bump to 3.3-beta1

(grafted from f22036f5f8
)

CMakeLists.txt

@@ -1,6 +1,6 @@
 project(Eigen)
 
-cmake_minimum_required(VERSION 2.8.4)
+cmake_minimum_required(VERSION 2.8.5)
 
 # guard against in-source builds
 
@@ -55,6 +55,7 @@ endif(EIGEN_HG_CHANGESET)
 
 
 include(CheckCXXCompilerFlag)
+include(GNUInstallDirs)
 
 set(CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake)
 
@@ -118,11 +119,7 @@ endmacro(ei_add_cxx_compiler_flag)
 
 if(NOT MSVC)
   # We assume that other compilers are partly compatible with GNUCC
-  
-#  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fexceptions")
-  set(CMAKE_CXX_FLAGS_DEBUG "-g3")
-  set(CMAKE_CXX_FLAGS_RELEASE "-g0 -O2")
-  
+
   # clang outputs some warnings for unknwon flags that are not caught by check_cxx_compiler_flag
   # adding -Werror turns such warnings into errors
   check_cxx_compiler_flag("-Werror" COMPILER_SUPPORT_WERROR)
@@ -341,24 +338,29 @@ option(EIGEN_TEST_CXX11 "Enable testing with C++11 and C++11 features (e.g. Tens
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
 
-# the user modifiable install path for header files
-set(EIGEN_INCLUDE_INSTALL_DIR ${EIGEN_INCLUDE_INSTALL_DIR} CACHE PATH "The directory where we install the header files (optional)")
-
-# set the internal install path for header files which depends on wether the user modifiable
-# EIGEN_INCLUDE_INSTALL_DIR has been set by the user or not.
+# Backward compatibility support for EIGEN_INCLUDE_INSTALL_DIR
 if(EIGEN_INCLUDE_INSTALL_DIR)
-  set(INCLUDE_INSTALL_DIR
-    ${EIGEN_INCLUDE_INSTALL_DIR}
-    CACHE INTERNAL
-    "The directory where we install the header files (internal)"
-  )
+  message(WARNING "EIGEN_INCLUDE_INSTALL_DIR is deprecated. Use INCLUDE_INSTALL_DIR instead.")
+endif()
+
+if(EIGEN_INCLUDE_INSTALL_DIR AND NOT INCLUDE_INSTALL_DIR)
+  set(INCLUDE_INSTALL_DIR ${EIGEN_INCLUDE_INSTALL_DIR}
+      CACHE PATH "The directory relative to CMAKE_PREFIX_PATH where Eigen header files are installed")
 else()
   set(INCLUDE_INSTALL_DIR
-    "${CMAKE_INSTALL_PREFIX}/include/eigen3"
-    CACHE INTERNAL
-    "The directory where we install the header files (internal)"
-  )
+      "${CMAKE_INSTALL_INCLUDEDIR}/eigen3"
+      CACHE PATH "The directory relative to CMAKE_PREFIX_PATH where Eigen header files are installed"
+      )
 endif()
+set(CMAKEPACKAGE_INSTALL_DIR
+    "${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)
@@ -377,21 +379,9 @@ install(FILES
   )
 
 if(EIGEN_BUILD_PKGCONFIG)
-    SET(path_separator ":")
-    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)
+    configure_file(eigen3.pc.in eigen3.pc @ONLY)
     install(FILES ${CMAKE_CURRENT_BINARY_DIR}/eigen3.pc
-        DESTINATION ${pkg_config_install_dir}/pkgconfig
+        DESTINATION ${PKGCONFIG_INSTALL_DIR}
         )
 endif(EIGEN_BUILD_PKGCONFIG)
 
@@ -454,12 +444,15 @@ 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 "              |     cmake . -DCMAKE_INSTALL_PREFIX=yourpath")
-  message(STATUS "              |   Eigen headers will then be installed to:")
-  message(STATUS "              |     ${INCLUDE_INSTALL_DIR}")
-  message(STATUS "              |   To install Eigen headers to a separate location, do:")
-  message(STATUS "              |     cmake . -DEIGEN_INCLUDE_INSTALL_DIR=yourpath")
+  message(STATUS "make install  | Install Eigen. Headers will be installed to:")
+  message(STATUS "              |     <CMAKE_INSTALL_PREFIX>/<INCLUDE_INSTALL_DIR>")
+  message(STATUS "              |   Using the following values:")
+  message(STATUS "              |     CMAKE_INSTALL_PREFIX: ${CMAKE_INSTALL_PREFIX}")
+  message(STATUS "              |     INCLUDE_INSTALL_DIR:  ${INCLUDE_INSTALL_DIR}")
+  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")
@@ -473,21 +466,13 @@ endif()
 
 message(STATUS "")
 
-set ( EIGEN_CONFIG_CMAKE_PATH
-      lib${LIB_SUFFIX}/cmake/eigen3
-      CACHE PATH "The directory where the CMake files are installed"
-    )
-if ( NOT IS_ABSOLUTE EIGEN_CONFIG_CMAKE_PATH )
-  set ( EIGEN_CONFIG_CMAKE_PATH ${CMAKE_INSTALL_PREFIX}/${EIGEN_CONFIG_CMAKE_PATH} )
-endif ()
 
-set ( EIGEN_USE_FILE ${EIGEN_CONFIG_CMAKE_PATH}/UseEigen3.cmake )
 set ( EIGEN_VERSION_STRING ${EIGEN_VERSION_NUMBER} )
 set ( EIGEN_VERSION_MAJOR  ${EIGEN_WORLD_VERSION} )
 set ( EIGEN_VERSION_MINOR  ${EIGEN_MAJOR_VERSION} )
 set ( EIGEN_VERSION_PATCH  ${EIGEN_MINOR_VERSION} )
 set ( EIGEN_DEFINITIONS "")
-set ( EIGEN_INCLUDE_DIR ${INCLUDE_INSTALL_DIR} )
+set ( EIGEN_INCLUDE_DIR "${CMAKE_INSTALL_PREFIX}/${INCLUDE_INSTALL_DIR}" )
 set ( EIGEN_INCLUDE_DIRS ${EIGEN_INCLUDE_DIR} )
 set ( EIGEN_ROOT_DIR ${CMAKE_INSTALL_PREFIX} )
 
@@ -498,7 +483,7 @@ configure_file ( ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Eigen3Config.cmake.in
 
 install ( FILES ${CMAKE_CURRENT_SOURCE_DIR}/cmake/UseEigen3.cmake
                 ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
-          DESTINATION ${EIGEN_CONFIG_CMAKE_PATH}
+          DESTINATION ${CMAKEPACKAGE_INSTALL_DIR}
         )
 
 # Add uninstall target

Eigen/Cholesky

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_CHOLESKY_MODULE_H
 #define EIGEN_CHOLESKY_MODULE_H
 

Eigen/CholmodSupport

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_CHOLMODSUPPORT_MODULE_H
 #define EIGEN_CHOLMODSUPPORT_MODULE_H
 

Eigen/Core

@@ -300,6 +300,7 @@ using std::ptrdiff_t;
 
 #include "src/Core/NumTraits.h"
 #include "src/Core/MathFunctions.h"
+#include "src/Core/SpecialFunctions.h"
 #include "src/Core/GenericPacketMath.h"
 
 #if defined EIGEN_VECTORIZE_AVX
@@ -382,8 +383,6 @@ using std::ptrdiff_t;
 #include "src/Core/DiagonalMatrix.h"
 #include "src/Core/Diagonal.h"
 #include "src/Core/DiagonalProduct.h"
-#include "src/Core/PermutationMatrix.h"
-#include "src/Core/Transpositions.h"
 #include "src/Core/Redux.h"
 #include "src/Core/Visitor.h"
 #include "src/Core/Fuzzy.h"
@@ -393,6 +392,9 @@ using std::ptrdiff_t;
 #include "src/Core/GeneralProduct.h"
 #include "src/Core/Solve.h"
 #include "src/Core/Inverse.h"
+#include "src/Core/SolverBase.h"
+#include "src/Core/PermutationMatrix.h"
+#include "src/Core/Transpositions.h"
 #include "src/Core/TriangularMatrix.h"
 #include "src/Core/SelfAdjointView.h"
 #include "src/Core/products/GeneralBlockPanelKernel.h"

Eigen/Eigenvalues

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_EIGENVALUES_MODULE_H
 #define EIGEN_EIGENVALUES_MODULE_H
 

Eigen/Geometry

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_GEOMETRY_MODULE_H
 #define EIGEN_GEOMETRY_MODULE_H
 

Eigen/Householder

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_HOUSEHOLDER_MODULE_H
 #define EIGEN_HOUSEHOLDER_MODULE_H
 

Eigen/IterativeLinearSolvers

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_ITERATIVELINEARSOLVERS_MODULE_H
 #define EIGEN_ITERATIVELINEARSOLVERS_MODULE_H
 
@@ -34,6 +41,7 @@
 #include "src/IterativeLinearSolvers/LeastSquareConjugateGradient.h"
 #include "src/IterativeLinearSolvers/BiCGSTAB.h"
 #include "src/IterativeLinearSolvers/IncompleteLUT.h"
+#include "src/IterativeLinearSolvers/IncompleteCholesky.h"
 
 #include "src/Core/util/ReenableStupidWarnings.h"
 

Eigen/Jacobi

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_JACOBI_MODULE_H
 #define EIGEN_JACOBI_MODULE_H
 

Eigen/LU

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_LU_MODULE_H
 #define EIGEN_LU_MODULE_H
 

Eigen/MetisSupport

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_METISSUPPORT_MODULE_H
 #define EIGEN_METISSUPPORT_MODULE_H
 

Eigen/OrderingMethods

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_ORDERINGMETHODS_MODULE_H
 #define EIGEN_ORDERINGMETHODS_MODULE_H
 

Eigen/PaStiXSupport

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_PASTIXSUPPORT_MODULE_H
 #define EIGEN_PASTIXSUPPORT_MODULE_H
 

Eigen/PardisoSupport

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_PARDISOSUPPORT_MODULE_H
 #define EIGEN_PARDISOSUPPORT_MODULE_H
 
@@ -7,8 +14,6 @@
 
 #include <mkl_pardiso.h>
 
-#include <unsupported/Eigen/SparseExtra>
-
 /** \ingroup Support_modules
   * \defgroup PardisoSupport_Module PardisoSupport module
   *

Eigen/QR

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_QR_MODULE_H
 #define EIGEN_QR_MODULE_H
 

Eigen/QtAlignedMalloc

@@ -1,3 +1,9 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
 #ifndef EIGEN_QTMALLOC_MODULE_H
 #define EIGEN_QTMALLOC_MODULE_H

Eigen/SPQRSupport

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_SPQRSUPPORT_MODULE_H
 #define EIGEN_SPQRSUPPORT_MODULE_H
 

Eigen/SVD

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_SVD_MODULE_H
 #define EIGEN_SVD_MODULE_H
 

Eigen/Sparse

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_SPARSE_MODULE_H
 #define EIGEN_SPARSE_MODULE_H
 

Eigen/SparseCore

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_SPARSECORE_MODULE_H
 #define EIGEN_SPARSECORE_MODULE_H
 
@@ -14,7 +21,7 @@
 /** 
   * \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/SparseQR

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_SPARSEQR_MODULE_H
 #define EIGEN_SPARSEQR_MODULE_H
 

Eigen/SuperLUSupport

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_SUPERLUSUPPORT_MODULE_H
 #define EIGEN_SUPERLUSUPPORT_MODULE_H
 
@@ -36,6 +43,8 @@ namespace Eigen { struct SluMatrix; }
   * - class SuperLU: a supernodal sequential LU factorization.
   * - class SuperILU: a supernodal sequential incomplete LU factorization (to be used as a preconditioner for iterative methods).
   *
+  * \warning This wrapper is only for the 4.x versions of SuperLU. The 3.x and 5.x versions are not supported.
+  *
   * \warning When including this module, you have to use SUPERLU_EMPTY instead of EMPTY which is no longer defined because it is too polluting.
   *
   * \code

Eigen/UmfPackSupport

@@ -1,3 +1,10 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
 #ifndef EIGEN_UMFPACKSUPPORT_MODULE_H
 #define EIGEN_UMFPACKSUPPORT_MODULE_H
 

Eigen/src/Cholesky/LDLT.h

@@ -99,14 +99,15 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * This calculates the decomposition for the input \a matrix.
       * \sa LDLT(Index size)
       */
-    explicit LDLT(const MatrixType& matrix)
+    template<typename InputType>
+    explicit LDLT(const EigenBase<InputType>& matrix)
       : m_matrix(matrix.rows(), matrix.cols()),
         m_transpositions(matrix.rows()),
         m_temporary(matrix.rows()),
         m_sign(internal::ZeroSign),
         m_isInitialized(false)
     {
-      compute(matrix);
+      compute(matrix.derived());
     }
 
     /** Clear any existing decomposition
@@ -188,7 +189,8 @@ template<typename _MatrixType, int _UpLo> class LDLT
     template<typename Derived>
     bool solveInPlace(MatrixBase<Derived> &bAndX) const;
 
-    LDLT& compute(const MatrixType& matrix);
+    template<typename InputType>
+    LDLT& compute(const EigenBase<InputType>& matrix);
 
     template <typename Derived>
     LDLT& rankUpdate(const MatrixBase<Derived>& w, const RealScalar& alpha=1);
@@ -427,14 +429,15 @@ template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
 /** Compute / recompute the LDLT decomposition A = L D L^* = U^* D U of \a matrix
   */
 template<typename MatrixType, int _UpLo>
-LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
+template<typename InputType>
+LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const EigenBase<InputType>& a)
 {
   check_template_parameters();
   
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
 
-  m_matrix = a;
+  m_matrix = a.derived();
 
   m_transpositions.resize(size);
   m_isInitialized = false;

Eigen/src/Cholesky/LLT.h

@@ -87,11 +87,12 @@ template<typename _MatrixType, int _UpLo> class LLT
     explicit LLT(Index size) : m_matrix(size, size),
                     m_isInitialized(false) {}
 
-    explicit LLT(const MatrixType& matrix)
+    template<typename InputType>
+    explicit LLT(const EigenBase<InputType>& matrix)
       : m_matrix(matrix.rows(), matrix.cols()),
         m_isInitialized(false)
     {
-      compute(matrix);
+      compute(matrix.derived());
     }
 
     /** \returns a view of the upper triangular matrix U */
@@ -131,7 +132,8 @@ template<typename _MatrixType, int _UpLo> class LLT
     template<typename Derived>
     void solveInPlace(MatrixBase<Derived> &bAndX) const;
 
-    LLT& compute(const MatrixType& matrix);
+    template<typename InputType>
+    LLT& compute(const EigenBase<InputType>& matrix);
 
     /** \returns the LLT decomposition matrix
       *
@@ -283,7 +285,7 @@ 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;
   }
@@ -381,14 +383,15 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
   * Output: \verbinclude TutorialLinAlgComputeTwice.out
   */
 template<typename MatrixType, int _UpLo>
-LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
+template<typename InputType>
+LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const EigenBase<InputType>& a)
 {
   check_template_parameters();
   
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
   m_matrix.resize(size, size);
-  m_matrix = a;
+  m_matrix = a.derived();
 
   m_isInitialized = true;
   bool ok = Traits::inplace_decomposition(m_matrix);

Eigen/src/CholmodSupport/CholmodSupport.h

@@ -78,7 +78,7 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_StorageIndex>& mat)
   {
     res.itype = CHOLMOD_INT;
   }
-  else if (internal::is_same<_StorageIndex,UF_long>::value)
+  else if (internal::is_same<_StorageIndex,SuiteSparse_long>::value)
   {
     res.itype = CHOLMOD_LONG;
   }
@@ -170,6 +170,10 @@ class CholmodBase : public SparseSolverBase<Derived>
     typedef typename MatrixType::RealScalar RealScalar;
     typedef MatrixType CholMatrixType;
     typedef typename MatrixType::StorageIndex StorageIndex;
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
 
   public:
 
@@ -350,6 +354,8 @@ class CholmodBase : public SparseSolverBase<Derived>
   * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
   *               or Upper. Default is Lower.
   *
+  * \implsparsesolverconcept
+  *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
   * \sa \ref TutorialSparseDirectSolvers, class CholmodSupernodalLLT, class SimplicialLLT
@@ -397,6 +403,8 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl
   * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
   *               or Upper. Default is Lower.
   *
+  * \implsparsesolverconcept
+  *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
   * \sa \ref TutorialSparseDirectSolvers, class CholmodSupernodalLLT, class SimplicialLDLT
@@ -442,6 +450,8 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp
   * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
   *               or Upper. Default is Lower.
   *
+  * \implsparsesolverconcept
+  *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
   * \sa \ref TutorialSparseDirectSolvers
@@ -489,6 +499,8 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper
   * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
   *               or Upper. Default is Lower.
   *
+  * \implsparsesolverconcept
+  *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
   * \sa \ref TutorialSparseDirectSolvers

Eigen/src/Core/ArrayBase.h

@@ -46,15 +46,14 @@ 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>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
     typedef DenseBase<Derived> Base;
+    using Base::operator*;
+    using Base::operator/;
     using Base::RowsAtCompileTime;
     using Base::ColsAtCompileTime;
     using Base::SizeAtCompileTime;

Eigen/src/Core/AssignEvaluator.h

@@ -54,6 +54,7 @@ private:
     InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime)
               : int(DstFlags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime)
               : int(Dst::MaxRowsAtCompileTime),
+    OuterStride = int(outer_stride_at_compile_time<Dst>::ret),
     MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
     PacketSize = unpacket_traits<PacketType>::size
   };
@@ -65,7 +66,9 @@ private:
     MightVectorize = StorageOrdersAgree
                   && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit)
                   && (functor_traits<AssignFunc>::PacketAccess),
-    MayInnerVectorize  = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0
+    MayInnerVectorize  = MightVectorize
+                       && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0
+                       && int(OuterStride)!=Dynamic && int(OuterStride)%int(PacketSize)==0
                        && int(JointAlignment)>=int(RequiredAlignment),
     MayLinearize = StorageOrdersAgree && (int(DstFlags) & int(SrcFlags) & LinearAccessBit),
     MayLinearVectorize = MightVectorize && MayLinearize && DstHasDirectAccess
@@ -95,10 +98,8 @@ private:
   enum {
     UnrollingLimit      = EIGEN_UNROLLING_LIMIT * (Vectorized ? int(PacketSize) : 1),
     MayUnrollCompletely = int(Dst::SizeAtCompileTime) != Dynamic
-                       && int(SrcEvaluator::CoeffReadCost) != Dynamic
                        && int(Dst::SizeAtCompileTime) * int(SrcEvaluator::CoeffReadCost) <= int(UnrollingLimit),
     MayUnrollInner      = int(InnerSize) != Dynamic
-                       && int(SrcEvaluator::CoeffReadCost) != Dynamic
                        && int(InnerSize) * int(SrcEvaluator::CoeffReadCost) <= int(UnrollingLimit)
   };
 
@@ -125,8 +126,8 @@ public:
     std::cerr << "DstXpr: " << typeid(typename DstEvaluator::XprType).name() << std::endl;
     std::cerr << "SrcXpr: " << typeid(typename SrcEvaluator::XprType).name() << std::endl;
     std::cerr.setf(std::ios::hex, std::ios::basefield);
-    EIGEN_DEBUG_VAR(DstFlags)
-    EIGEN_DEBUG_VAR(SrcFlags)
+    std::cerr << "DstFlags" << " = " << DstFlags << " (" << demangle_flags(DstFlags) << " )" << std::endl;
+    std::cerr << "SrcFlags" << " = " << SrcFlags << " (" << demangle_flags(SrcFlags) << " )" << std::endl;
     std::cerr.unsetf(std::ios::hex);
     EIGEN_DEBUG_VAR(DstAlignment)
     EIGEN_DEBUG_VAR(SrcAlignment)
@@ -141,11 +142,11 @@ public:
     EIGEN_DEBUG_VAR(MayInnerVectorize)
     EIGEN_DEBUG_VAR(MayLinearVectorize)
     EIGEN_DEBUG_VAR(MaySliceVectorize)
-    EIGEN_DEBUG_VAR(Traversal)
+    std::cerr << "Traversal" << " = " << Traversal << " (" << demangle_traversal(Traversal) << ")" << std::endl;
     EIGEN_DEBUG_VAR(UnrollingLimit)
     EIGEN_DEBUG_VAR(MayUnrollCompletely)
     EIGEN_DEBUG_VAR(MayUnrollInner)
-    EIGEN_DEBUG_VAR(Unrolling)
+    std::cerr << "Unrolling" << " = " << Unrolling << " (" << demangle_unrolling(Unrolling) << ")" << std::endl;
     std::cerr << std::endl;
   }
 #endif
@@ -288,7 +289,7 @@ struct dense_assignment_loop;
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, DefaultTraversal, NoUnrolling>
 {
-  EIGEN_DEVICE_FUNC static void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static void EIGEN_STRONG_INLINE run(Kernel &kernel)
   {
     for(Index outer = 0; outer < kernel.outerSize(); ++outer) {
       for(Index inner = 0; inner < kernel.innerSize(); ++inner) {
@@ -311,7 +312,6 @@ struct dense_assignment_loop<Kernel, DefaultTraversal, CompleteUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, DefaultTraversal, InnerUnrolling>
 {
-  typedef typename Kernel::StorageIndex StorageIndex;
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
@@ -392,7 +392,6 @@ struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, NoUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, CompleteUnrolling>
 {
-  typedef typename Kernel::StorageIndex StorageIndex;
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
@@ -414,7 +413,7 @@ template<typename Kernel>
 struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, NoUnrolling>
 {
   typedef typename Kernel::PacketType PacketType;
-  EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     const Index innerSize = kernel.innerSize();
     const Index outerSize = kernel.outerSize();
@@ -438,7 +437,6 @@ struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, CompleteUnrolling
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling>
 {
-  typedef typename Kernel::StorageIndex StorageIndex;
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
@@ -455,7 +453,7 @@ struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, LinearTraversal, NoUnrolling>
 {
-  EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     const Index size = kernel.size();
     for(Index i = 0; i < size; ++i)
@@ -545,7 +543,6 @@ public:
   typedef DstEvaluatorTypeT DstEvaluatorType;
   typedef SrcEvaluatorTypeT SrcEvaluatorType;
   typedef typename DstEvaluatorType::Scalar Scalar;
-  typedef typename DstEvaluatorType::StorageIndex StorageIndex;
   typedef copy_using_evaluator_traits<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor> AssignmentTraits;
   typedef typename AssignmentTraits::PacketType PacketType;
   
@@ -565,26 +562,23 @@ public:
   EIGEN_DEVICE_FUNC Index cols() const        { return m_dstExpr.cols(); }
   EIGEN_DEVICE_FUNC Index outerStride() const { return m_dstExpr.outerStride(); }
   
-  // TODO get rid of this one:
-  EIGEN_DEVICE_FUNC DstXprType& dstExpression() const { return m_dstExpr; }
-  
   EIGEN_DEVICE_FUNC DstEvaluatorType& dstEvaluator() { return m_dst; }
   EIGEN_DEVICE_FUNC const SrcEvaluatorType& srcEvaluator() const { return m_src; }
   
   /// Assign src(row,col) to dst(row,col) through the assignment functor.
-  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index row, Index col)
   {
     m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col));
   }
   
   /// \sa assignCoeff(Index,Index)
-  EIGEN_DEVICE_FUNC void assignCoeff(Index index)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index index)
   {
     m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index));
   }
   
   /// \sa assignCoeff(Index,Index)
-  EIGEN_DEVICE_FUNC void assignCoeffByOuterInner(Index outer, Index inner)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeffByOuterInner(Index outer, Index inner)
   {
     Index row = rowIndexByOuterInner(outer, inner); 
     Index col = colIndexByOuterInner(outer, inner); 
@@ -593,26 +587,26 @@ public:
   
   
   template<int StoreMode, int LoadMode, typename PacketType>
-  EIGEN_DEVICE_FUNC void assignPacket(Index row, Index col)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index row, Index col)
   {
     m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(row,col), m_src.template packet<LoadMode,PacketType>(row,col));
   }
   
   template<int StoreMode, int LoadMode, typename PacketType>
-  EIGEN_DEVICE_FUNC void assignPacket(Index index)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index index)
   {
     m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(index), m_src.template packet<LoadMode,PacketType>(index));
   }
   
   template<int StoreMode, int LoadMode, typename PacketType>
-  EIGEN_DEVICE_FUNC void assignPacketByOuterInner(Index outer, Index inner)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacketByOuterInner(Index outer, Index inner)
   {
     Index row = rowIndexByOuterInner(outer, inner); 
     Index col = colIndexByOuterInner(outer, inner);
     assignPacket<StoreMode,LoadMode,PacketType>(row, col);
   }
   
-  EIGEN_DEVICE_FUNC static Index rowIndexByOuterInner(Index outer, Index inner)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner)
   {
     typedef typename DstEvaluatorType::ExpressionTraits Traits;
     return int(Traits::RowsAtCompileTime) == 1 ? 0
@@ -621,7 +615,7 @@ public:
       : inner;
   }
 
-  EIGEN_DEVICE_FUNC static Index colIndexByOuterInner(Index outer, Index inner)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner)
   {
     typedef typename DstEvaluatorType::ExpressionTraits Traits;
     return int(Traits::ColsAtCompileTime) == 1 ? 0
@@ -719,14 +713,8 @@ EIGEN_DEVICE_FUNC void call_assignment(Dst& dst, const Src& src, const Func& fun
 }
 
 // by-pass AssumeAliasing
-// FIXME the const version should probably not be needed
 // When there is no aliasing, we require that 'dst' has been properly resized
 template<typename Dst, template <typename> class StorageBase, typename Src, typename Func>
-EIGEN_DEVICE_FUNC void call_assignment(const NoAlias<Dst,StorageBase>& dst, const Src& src, const Func& func)
-{
-  call_assignment_no_alias(dst.expression(), src, func);
-}
-template<typename Dst, template <typename> class StorageBase, typename Src, typename Func>
 EIGEN_DEVICE_FUNC void call_assignment(NoAlias<Dst,StorageBase>& dst, const Src& src, const Func& func)
 {
   call_assignment_no_alias(dst.expression(), src, func);
@@ -737,11 +725,9 @@ template<typename Dst, typename Src, typename Func>
 EIGEN_DEVICE_FUNC void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func)
 {
   enum {
-    NeedToTranspose = (  (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1)
-                        |   // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&".
-                                // revert to || as soon as not needed anymore.
-                         (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1))
-                     && int(Dst::SizeAtCompileTime) != 1
+    NeedToTranspose = (    (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1)
+                        || (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1)
+                      ) && int(Dst::SizeAtCompileTime) != 1
   };
 
   Index dstRows = NeedToTranspose ? src.cols() : src.rows();
@@ -756,11 +742,7 @@ EIGEN_DEVICE_FUNC void call_assignment_no_alias(Dst& dst, const Src& src, const
   // TODO check whether this is the right place to perform these checks:
   EIGEN_STATIC_ASSERT_LVALUE(Dst)
   EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(ActualDstTypeCleaned,Src)
-
-  // TODO this line is commented to allow matrix = permutation
-  // Actually, the "Scalar" type for a permutation matrix does not really make sense,
-  // perhaps it could be void, and EIGEN_CHECK_BINARY_COMPATIBILIY could allow micing void with anything...?
-//   EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar);
+  EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar);
   
   Assignment<ActualDstTypeCleaned,Src,Func>::run(actualDst, src, func);
 }

Eigen/src/Core/BooleanRedux.h

@@ -83,8 +83,6 @@ inline bool DenseBase<Derived>::all() const
   typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && Evaluator::CoeffReadCost != Dynamic
-          && NumTraits<Scalar>::AddCost != Dynamic
           && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
   Evaluator evaluator(derived());
@@ -109,8 +107,6 @@ inline bool DenseBase<Derived>::any() const
   typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && Evaluator::CoeffReadCost != Dynamic
-          && NumTraits<Scalar>::AddCost != Dynamic
           && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
   Evaluator evaluator(derived());
@@ -142,7 +138,11 @@ inline Eigen::Index DenseBase<Derived>::count() const
 template<typename Derived>
 inline bool DenseBase<Derived>::hasNaN() const
 {
+#if EIGEN_COMP_MSVC || (defined __FAST_MATH__)
+  return derived().array().isNaN().any();
+#else
   return !((derived().array()==derived().array()).all());
+#endif
 }
 
 /** \returns true if \c *this contains only finite numbers, i.e., no NaN and no +/-INF values.
@@ -152,7 +152,11 @@ inline bool DenseBase<Derived>::hasNaN() const
 template<typename Derived>
 inline bool DenseBase<Derived>::allFinite() const
 {
+#if EIGEN_COMP_MSVC || (defined __FAST_MATH__)
+  return derived().array().isFinite().all();
+#else
   return !((derived()-derived()).hasNaN());
+#endif
 }
     
 } // end namespace Eigen

Eigen/src/Core/CommaInitializer.h

@@ -106,7 +106,7 @@ struct CommaInitializer
   EIGEN_DEVICE_FUNC
   inline ~CommaInitializer()
 #if defined VERIFY_RAISES_ASSERT && (!defined EIGEN_NO_ASSERTION_CHECKING) && defined EIGEN_EXCEPTIONS
-  throw(Eigen::eigen_assert_exception)
+  EIGEN_EXCEPTION_SPEC(Eigen::eigen_assert_exception)
 #endif
   {
     eigen_assert((m_row+m_currentBlockRows) == m_xpr.rows()

Eigen/src/Core/CoreEvaluators.h

@@ -29,6 +29,7 @@ struct storage_kind_to_evaluator_kind {
 template<typename StorageKind> struct storage_kind_to_shape;
 
 template<> struct storage_kind_to_shape<Dense>                  { typedef DenseShape Shape;           };
+template<> struct storage_kind_to_shape<SolverStorage>          { typedef SolverShape Shape;           };
 template<> struct storage_kind_to_shape<PermutationStorage>     { typedef PermutationShape Shape;     };
 template<> struct storage_kind_to_shape<TranspositionsStorage>  { typedef TranspositionsShape Shape;  };
 
@@ -98,9 +99,6 @@ struct evaluator<const T>
 template<typename ExpressionType>
 struct evaluator_base : public noncopyable
 {
-  // FIXME is it really usefull?
-  typedef typename traits<ExpressionType>::StorageIndex StorageIndex;
-  
   // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices.
   typedef traits<ExpressionType> ExpressionTraits;
   
@@ -140,11 +138,15 @@ struct evaluator<PlainObjectBase<Derived> >
       m_outerStride(IsVectorAtCompileTime  ? 0 
                                            : int(IsRowMajor) ? ColsAtCompileTime 
                                            : RowsAtCompileTime)
-  {}
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
   
   EIGEN_DEVICE_FUNC explicit evaluator(const PlainObjectType& m)
     : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) 
-  { }
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 
   EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
   {
@@ -324,13 +326,15 @@ struct evaluator<CwiseNullaryOp<NullaryOp,PlainObjectType> >
           &  (  HereditaryBits
               | (functor_has_linear_access<NullaryOp>::ret  ? LinearAccessBit : 0)
               | (functor_traits<NullaryOp>::PacketAccess    ? PacketAccessBit : 0)))
-          | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit), // FIXME EvalBeforeNestingBit should be needed anymore
-    Alignment = 0 // FIXME alignment should not matter here, perhaps we could set it to AlignMax??
+          | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit),
+    Alignment = AlignedMax
   };
 
   EIGEN_DEVICE_FUNC explicit evaluator(const XprType& n)
     : m_functor(n.functor()) 
-  { }
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 
   typedef typename XprType::CoeffReturnType CoeffReturnType;
 
@@ -379,7 +383,10 @@ struct unary_evaluator<CwiseUnaryOp<UnaryOp, ArgType>, IndexBased >
   EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op)
     : m_functor(op.functor()), 
       m_argImpl(op.nestedExpression()) 
-  { }
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 
   typedef typename XprType::CoeffReturnType CoeffReturnType;
 
@@ -452,7 +459,10 @@ struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IndexBased, IndexBase
     : m_functor(xpr.functor()),
       m_lhsImpl(xpr.lhs()), 
       m_rhsImpl(xpr.rhs())  
-  { }
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 
   typedef typename XprType::CoeffReturnType CoeffReturnType;
 
@@ -505,7 +515,10 @@ struct unary_evaluator<CwiseUnaryView<UnaryOp, ArgType>, IndexBased>
   EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op)
     : m_unaryOp(op.functor()), 
       m_argImpl(op.nestedExpression()) 
-  { }
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 
   typedef typename XprType::Scalar Scalar;
   typedef typename XprType::CoeffReturnType CoeffReturnType;
@@ -562,6 +575,7 @@ struct mapbase_evaluator : evaluator_base<Derived>
   {
     EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(evaluator<Derived>::Flags&PacketAccessBit, internal::inner_stride_at_compile_time<Derived>::ret==1),
                         PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
   }
  
   EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
@@ -636,17 +650,9 @@ struct evaluator<Map<PlainObjectType, MapOptions, StrideType> >
     HasNoStride = HasNoInnerStride && HasNoOuterStride,
     IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic,
     
-    PacketAlignment = unpacket_traits<PacketScalar>::alignment,
-    
-    KeepsPacketAccess = bool(HasNoInnerStride)
-                        && ( bool(IsDynamicSize)
-                           || HasNoOuterStride
-                           || ( OuterStrideAtCompileTime!=Dynamic
-                           && ((static_cast<int>(sizeof(Scalar))*OuterStrideAtCompileTime) % PacketAlignment)==0 ) ),
-    Flags0 = evaluator<PlainObjectType>::Flags,
-    Flags1 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime))
-           ? int(Flags0) : int(Flags0 & ~LinearAccessBit),
-    Flags = KeepsPacketAccess ? int(Flags1) : (int(Flags1) & ~PacketAccessBit),
+    PacketAccessMask = bool(HasNoInnerStride) ? ~int(0) : ~int(PacketAccessBit),
+    LinearAccessMask = bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime) ? ~int(0) : ~int(LinearAccessBit),
+    Flags = int( evaluator<PlainObjectType>::Flags) & (LinearAccessMask&PacketAccessMask),
     
     Alignment = int(MapOptions)&int(AlignedMask)
   };
@@ -724,7 +730,10 @@ struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
     Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ArgType>::Alignment, Alignment0)
   };
   typedef block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel> block_evaluator_type;
-  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& block) : block_evaluator_type(block) {}
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& block) : block_evaluator_type(block)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 };
 
 // no direct-access => dispatch to a unary evaluator
@@ -825,14 +834,14 @@ struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /* HasDirectAc
   EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block)
     : mapbase_evaluator<XprType, typename XprType::PlainObject>(block) 
   {
-    // FIXME this should be an internal assertion
+    // TODO: for the 3.3 release, this should be turned to an internal assertion, but let's keep it as is for the beta lifetime
     eigen_assert(((size_t(block.data()) % EIGEN_PLAIN_ENUM_MAX(1,evaluator<XprType>::Alignment)) == 0) && "data is not aligned");
   }
 };
 
 
 // -------------------- Select --------------------
-// TODO shall we introduce a ternary_evaluator?
+// NOTE shall we introduce a ternary_evaluator?
 
 // TODO enable vectorization for Select
 template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
@@ -842,8 +851,8 @@ struct evaluator<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
   typedef Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> XprType;
   enum {
     CoeffReadCost = evaluator<ConditionMatrixType>::CoeffReadCost
-                  + EIGEN_SIZE_MAX(evaluator<ThenMatrixType>::CoeffReadCost,
-                                   evaluator<ElseMatrixType>::CoeffReadCost),
+                  + EIGEN_PLAIN_ENUM_MAX(evaluator<ThenMatrixType>::CoeffReadCost,
+                                         evaluator<ElseMatrixType>::CoeffReadCost),
 
     Flags = (unsigned int)evaluator<ThenMatrixType>::Flags & evaluator<ElseMatrixType>::Flags & HereditaryBits,
     
@@ -854,7 +863,9 @@ struct evaluator<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
     : m_conditionImpl(select.conditionMatrix()),
       m_thenImpl(select.thenMatrix()),
       m_elseImpl(select.elseMatrix())
-  { }
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
  
   typedef typename XprType::CoeffReturnType CoeffReturnType;
 
@@ -897,8 +908,8 @@ struct unary_evaluator<Replicate<ArgType, RowFactor, ColFactor> >
   
   enum {
     CoeffReadCost = evaluator<ArgTypeNestedCleaned>::CoeffReadCost,
-    
-    Flags = (evaluator<ArgTypeNestedCleaned>::Flags & HereditaryBits & ~RowMajorBit) | (traits<XprType>::Flags & RowMajorBit),
+    LinearAccessMask = XprType::IsVectorAtCompileTime ? LinearAccessBit : 0,
+    Flags = (evaluator<ArgTypeNestedCleaned>::Flags & (HereditaryBits|LinearAccessMask) & ~RowMajorBit) | (traits<XprType>::Flags & RowMajorBit),
     
     Alignment = evaluator<ArgTypeNestedCleaned>::Alignment
   };
@@ -957,7 +968,7 @@ struct unary_evaluator<Replicate<ArgType, RowFactor, ColFactor> >
   }
  
 protected:
-  const ArgTypeNested m_arg; // FIXME is it OK to store both the argument and its evaluator?? (we have the same situation in evaluator_product)
+  const ArgTypeNested m_arg;
   evaluator<ArgTypeNestedCleaned> m_argImpl;
   const variable_if_dynamic<Index, ArgType::RowsAtCompileTime> m_rows;
   const variable_if_dynamic<Index, ArgType::ColsAtCompileTime> m_cols;
@@ -965,48 +976,57 @@ protected:
 
 
 // -------------------- PartialReduxExpr --------------------
-//
-// This is a wrapper around the expression object. 
-// TODO: Find out how to write a proper evaluator without duplicating
-//       the row() and col() member functions.
 
 template< typename ArgType, typename MemberOp, int Direction>
 struct evaluator<PartialReduxExpr<ArgType, MemberOp, Direction> >
   : evaluator_base<PartialReduxExpr<ArgType, MemberOp, Direction> >
 {
   typedef PartialReduxExpr<ArgType, MemberOp, Direction> XprType;
-  typedef typename XprType::Scalar InputScalar;
+  typedef typename internal::nested_eval<ArgType,1>::type ArgTypeNested;
+  typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned;
+  typedef typename ArgType::Scalar InputScalar;
+  typedef typename XprType::Scalar Scalar;
   enum {
-    TraversalSize = Direction==int(Vertical) ? int(ArgType::RowsAtCompileTime) :  int(XprType::ColsAtCompileTime)
+    TraversalSize = Direction==int(Vertical) ? int(ArgType::RowsAtCompileTime) :  int(ArgType::ColsAtCompileTime)
   };
   typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
   enum {
-    CoeffReadCost = TraversalSize==Dynamic ? Dynamic
+    CoeffReadCost = TraversalSize==Dynamic ? HugeCost
                   : TraversalSize * evaluator<ArgType>::CoeffReadCost + int(CostOpType::value),
     
-    Flags = (traits<XprType>::Flags&RowMajorBit) | (evaluator<ArgType>::Flags&HereditaryBits),
+    Flags = (traits<XprType>::Flags&RowMajorBit) | (evaluator<ArgType>::Flags&(HereditaryBits&(~RowMajorBit))),
     
-    Alignment = 0 // FIXME this could be improved
+    Alignment = 0 // FIXME this will need to be improved once PartialReduxExpr is vectorized
   };
 
-  EIGEN_DEVICE_FUNC explicit evaluator(const XprType expr)
-    : m_expr(expr)
-  {}
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType xpr)
+    : m_arg(xpr.nestedExpression()), m_functor(xpr.functor())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(TraversalSize==Dynamic ? HugeCost : int(CostOpType::value));
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 
   typedef typename XprType::CoeffReturnType CoeffReturnType;
- 
-  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
-  { 
-    return m_expr.coeff(row, col);
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
+  {
+    if (Direction==Vertical)
+      return m_functor(m_arg.col(j));
+    else
+      return m_functor(m_arg.row(i));
   }
-  
-  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
-  { 
-    return m_expr.coeff(index);
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
+  {
+    if (Direction==Vertical)
+      return m_functor(m_arg.col(index));
+    else
+      return m_functor(m_arg.row(index));
   }
 
 protected:
-  const XprType m_expr;
+  const ArgTypeNested m_arg;
+  const MemberOp m_functor;
 };
 
 
@@ -1130,6 +1150,7 @@ struct unary_evaluator<Reverse<ArgType, Direction> >
     // FIXME enable DirectAccess with negative strides?
     Flags0 = evaluator<ArgType>::Flags,
     LinearAccess = ( (Direction==BothDirections) && (int(Flags0)&PacketAccessBit) )
+                  || ((ReverseRow && XprType::ColsAtCompileTime==1) || (ReverseCol && XprType::RowsAtCompileTime==1))
                  ? LinearAccessBit : 0,
 
     Flags = int(Flags0) & (HereditaryBits | PacketAccessBit | LinearAccess),
@@ -1139,8 +1160,8 @@ struct unary_evaluator<Reverse<ArgType, Direction> >
 
   EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& reverse)
     : m_argImpl(reverse.nestedExpression()),
-      m_rows(ReverseRow ? reverse.nestedExpression().rows() : 0),
-      m_cols(ReverseCol ? reverse.nestedExpression().cols() : 0)
+      m_rows(ReverseRow ? reverse.nestedExpression().rows() : 1),
+      m_cols(ReverseCol ? reverse.nestedExpression().cols() : 1)
   { }
  
   EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
@@ -1214,8 +1235,9 @@ protected:
   evaluator<ArgType> m_argImpl;
 
   // If we do not reverse rows, then we do not need to know the number of rows; same for columns
-  const variable_if_dynamic<Index, ReverseRow ? ArgType::RowsAtCompileTime : 0> m_rows;
-  const variable_if_dynamic<Index, ReverseCol ? ArgType::ColsAtCompileTime : 0> m_cols;
+  // Nonetheless, in this case it is important to set to 1 such that the coeff(index) method works fine for vectors.
+  const variable_if_dynamic<Index, ReverseRow ? ArgType::RowsAtCompileTime : 1> m_rows;
+  const variable_if_dynamic<Index, ReverseCol ? ArgType::ColsAtCompileTime : 1> m_cols;
 };
 
 
@@ -1331,20 +1353,16 @@ struct evaluator<EvalToTemp<ArgType> >
   typedef evaluator<PlainObject> Base;
   
   EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
-    : m_result(xpr.rows(), xpr.cols())
+    : m_result(xpr.arg())
   {
     ::new (static_cast<Base*>(this)) Base(m_result);
-    // TODO we should simply do m_result(xpr.arg());
-    call_dense_assignment_loop(m_result, xpr.arg());
   }
 
   // This constructor is used when nesting an EvalTo evaluator in another evaluator
   EIGEN_DEVICE_FUNC evaluator(const ArgType& arg)
-    : m_result(arg.rows(), arg.cols())
+    : m_result(arg)
   {
     ::new (static_cast<Base*>(this)) Base(m_result);
-    // TODO we should simply do m_result(xpr.arg());
-    call_dense_assignment_loop(m_result, arg);
   }
 
 protected:

Eigen/src/Core/DenseBase.h

@@ -40,18 +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,
-                                     typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>
+  : public internal::special_scalar_op_base<Derived, typename internal::traits<Derived>::Scalar,
+                                            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*;
-    using internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
-                typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>::operator/;
-
 
     /** Inner iterator type to iterate over the coefficients of a row or column.
       * \sa class InnerIterator
@@ -77,9 +73,10 @@ template<typename Derived> class DenseBase
     typedef Scalar value_type;
     
     typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef internal::special_scalar_op_base<Derived,Scalar,RealScalar, DenseCoeffsBase<Derived> > Base;
 
-    typedef internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
-                      typename NumTraits<typename internal::traits<Derived>::Scalar>::Real> Base;
+    using Base::operator*;
+    using Base::operator/;
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -215,10 +212,6 @@ template<typename Derived> class DenseBase
       * of stored coefficients. */
     EIGEN_DEVICE_FUNC
     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.
       *

Eigen/src/Core/DenseCoeffsBase.h

@@ -138,6 +138,8 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     EIGEN_STRONG_INLINE CoeffReturnType
     coeff(Index index) const
     {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
       eigen_internal_assert(index >= 0 && index < size());
       return internal::evaluator<Derived>(derived()).coeff(index);
     }
@@ -243,6 +245,8 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     template<int LoadMode>
     EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
     {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
       typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
       eigen_internal_assert(index >= 0 && index < size());
       return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(index);
@@ -370,6 +374,8 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     EIGEN_STRONG_INLINE Scalar&
     coeffRef(Index index)
     {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
       eigen_internal_assert(index >= 0 && index < size());
       return internal::evaluator<Derived>(derived()).coeffRef(index);
     }
@@ -617,7 +623,7 @@ static inline Index first_default_aligned(const DenseBase<Derived>& m)
 {
   typedef typename Derived::Scalar Scalar;
   typedef typename packet_traits<Scalar>::type DefaultPacketType;
-  return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(m);
+  return internal::first_aligned<int(unpacket_traits<DefaultPacketType>::alignment),Derived>(m);
 }
 
 template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret>

Eigen/src/Core/Dot.h

@@ -178,9 +178,11 @@ struct lpNorm_selector<Derived, Infinity>
 
 } // end namespace internal
 
-/** \returns the \f$ \ell^p \f$ norm of *this, that is, returns the p-th root of the sum of the p-th powers of the absolute values
-  *          of the coefficients of *this. If \a p is the special value \a Eigen::Infinity, this function returns the \f$ \ell^\infty \f$
-  *          norm, that is the maximum of the absolute values of the coefficients of *this.
+/** \returns the \b coefficient-wise \f$ \ell^p \f$ norm of \c *this, that is, returns the p-th root of the sum of the p-th powers of the absolute values
+  *          of the coefficients of \c *this. If \a p is the special value \a Eigen::Infinity, this function returns the \f$ \ell^\infty \f$
+  *          norm, that is the maximum of the absolute values of the coefficients of \c *this.
+  *
+  * \note For matrices, this function does not compute the <a href="https://en.wikipedia.org/wiki/Operator_norm">operator-norm</a>. That is, if \c *this is a matrix, then its coefficients are interpreted as a 1D vector. Nonetheless, you can easily compute the 1-norm and \f$\infty\f$-norm matrix operator norms using \link TutorialReductionsVisitorsBroadcastingReductionsNorm partial reductions \endlink.
   *
   * \sa norm()
   */

Eigen/src/Core/GeneralProduct.h

@@ -160,7 +160,7 @@ template<>              struct product_type_selector<Large,Large,Small>  { enum
 namespace internal {
 
 template<int Side, int StorageOrder, bool BlasCompatible>
-struct gemv_dense_sense_selector;
+struct gemv_dense_selector;
 
 } // end namespace internal
 
@@ -204,19 +204,19 @@ struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
 
 // The vector is on the left => transposition
 template<int StorageOrder, bool BlasCompatible>
-struct gemv_dense_sense_selector<OnTheLeft,StorageOrder,BlasCompatible>
+struct gemv_dense_selector<OnTheLeft,StorageOrder,BlasCompatible>
 {
   template<typename Lhs, typename Rhs, typename Dest>
   static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
     Transpose<Dest> destT(dest);
     enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
-    gemv_dense_sense_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
+    gemv_dense_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
       ::run(rhs.transpose(), lhs.transpose(), destT, alpha);
   }
 };
 
-template<> struct gemv_dense_sense_selector<OnTheRight,ColMajor,true>
+template<> struct gemv_dense_selector<OnTheRight,ColMajor,true>
 {
   template<typename Lhs, typename Rhs, typename Dest>
   static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
@@ -292,7 +292,7 @@ template<> struct gemv_dense_sense_selector<OnTheRight,ColMajor,true>
   }
 };
 
-template<> struct gemv_dense_sense_selector<OnTheRight,RowMajor,true>
+template<> struct gemv_dense_selector<OnTheRight,RowMajor,true>
 {
   template<typename Lhs, typename Rhs, typename Dest>
   static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
@@ -345,27 +345,28 @@ template<> struct gemv_dense_sense_selector<OnTheRight,RowMajor,true>
   }
 };
 
-template<> struct gemv_dense_sense_selector<OnTheRight,ColMajor,false>
+template<> struct gemv_dense_selector<OnTheRight,ColMajor,false>
 {
   template<typename Lhs, typename Rhs, typename Dest>
   static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    // TODO makes sure dest is sequentially stored in memory, otherwise use a temp
+    // TODO if rhs is large enough it might be beneficial to make sure that dest is sequentially stored in memory, otherwise use a temp
+    typename nested_eval<Rhs,1>::type actual_rhs(rhs);
     const Index size = rhs.rows();
     for(Index k=0; k<size; ++k)
-      dest += (alpha*rhs.coeff(k)) * lhs.col(k);
+      dest += (alpha*actual_rhs.coeff(k)) * lhs.col(k);
   }
 };
 
-template<> struct gemv_dense_sense_selector<OnTheRight,RowMajor,false>
+template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
 {
   template<typename Lhs, typename Rhs, typename Dest>
   static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp
+    typename nested_eval<Rhs,Lhs::RowsAtCompileTime>::type actual_rhs(rhs);
     const Index rows = dest.rows();
     for(Index i=0; i<rows; ++i)
-      dest.coeffRef(i) += alpha * (lhs.row(i).cwiseProduct(rhs.transpose())).sum();
+      dest.coeffRef(i) += alpha * (lhs.row(i).cwiseProduct(actual_rhs.transpose())).sum();
   }
 };
 

Eigen/src/Core/GenericPacketMath.h

@@ -43,7 +43,7 @@ struct default_packet_traits
 {
   enum {
     HasHalfPacket = 0,
-    
+
     HasAdd    = 1,
     HasSub    = 1,
     HasMul    = 1,
@@ -74,10 +74,15 @@ struct default_packet_traits
     HasSinh    = 0,
     HasCosh    = 0,
     HasTanh    = 0,
+    HasLGamma = 0,
+    HasErf = 0,
+    HasErfc = 0,
 
     HasRound  = 0,
     HasFloor  = 0,
-    HasCeil   = 0
+    HasCeil   = 0,
+
+    HasSign   = 0
   };
 };
 
@@ -430,6 +435,18 @@ Packet pfloor(const Packet& a) { using numext::floor; return floor(a); }
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); }
 
+/** \internal \returns the ln(|gamma(\a a)|) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plgamma(const Packet& a) { using numext::lgamma; return lgamma(a); }
+
+/** \internal \returns the erf(\a a) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet perf(const Packet& a) { using numext::erf; return erf(a); }
+
+/** \internal \returns the erfc(\a a) (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet perfc(const Packet& a) { using numext::erfc; return erfc(a); }
+
 /***************************************************************************
 * The following functions might not have to be overwritten for vectorized types
 ***************************************************************************/

Eigen/src/Core/GlobalFunctions.h

@@ -49,6 +49,9 @@ namespace Eigen
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sinh,scalar_sinh_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cosh,scalar_cosh_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tanh,scalar_tanh_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(lgamma,scalar_lgamma_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erf,scalar_erf_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erfc,scalar_erfc_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(exp,scalar_exp_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op)
@@ -64,6 +67,7 @@ namespace Eigen
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isnan,scalar_isnan_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isinf,scalar_isinf_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isfinite,scalar_isfinite_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sign,scalar_sign_op)
   
   template<typename Derived>
   inline const Eigen::CwiseUnaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar>, const Derived>

Eigen/src/Core/Inverse.h

@@ -12,8 +12,6 @@
 
 namespace Eigen { 
 
-// TODO move the general declaration in Core, and rename this file DenseInverseImpl.h, or something like this...
-
 template<typename XprType,typename StorageKind> class InverseImpl;
 
 namespace internal {
@@ -49,11 +47,13 @@ public:
   typedef typename XprType::PlainObject                       PlainObject;
   typedef typename internal::ref_selector<XprType>::type      XprTypeNested;
   typedef typename internal::remove_all<XprTypeNested>::type  XprTypeNestedCleaned;
+  typedef typename internal::ref_selector<Inverse>::type Nested;
+  typedef typename internal::remove_all<XprType>::type NestedExpression;
   
   explicit Inverse(const XprType &xpr)
     : m_xpr(xpr)
   {}
-  
+
   EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
   EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
 
@@ -63,25 +63,16 @@ protected:
   XprTypeNested m_xpr;
 };
 
-/** \internal
-  * Specialization of the Inverse expression for dense expressions.
-  * Direct access to the coefficients are discared.
-  * FIXME this intermediate class is probably not needed anymore.
-  */
-template<typename XprType>
-class InverseImpl<XprType,Dense>
-  : public MatrixBase<Inverse<XprType> >
+// Generic API dispatcher
+template<typename XprType, typename StorageKind>
+class InverseImpl
+  : public internal::generic_xpr_base<Inverse<XprType> >::type
 {
-  typedef Inverse<XprType> Derived;
-  
 public:
-  
-  typedef MatrixBase<Derived> Base;
-  EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
-  typedef typename internal::remove_all<XprType>::type NestedExpression;
-
+  typedef typename internal::generic_xpr_base<Inverse<XprType> >::type Base;
+  typedef typename XprType::Scalar Scalar;
 private:
-  
+
   Scalar coeff(Index row, Index col) const;
   Scalar coeff(Index i) const;
 };

Eigen/src/Core/MapBase.h

@@ -155,6 +155,10 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       checkSanity();
     }
 
+    #ifdef EIGEN_MAPBASE_PLUGIN
+    #include EIGEN_MAPBASE_PLUGIN
+    #endif
+
   protected:
 
     EIGEN_DEVICE_FUNC

Eigen/src/Core/MathFunctions.h

@@ -241,8 +241,8 @@ struct conj_retval
 * Implementation of abs2                                                 *
 ****************************************************************************/
 
-template<typename Scalar>
-struct abs2_impl
+template<typename Scalar,bool IsComplex>
+struct abs2_impl_default
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
   EIGEN_DEVICE_FUNC
@@ -252,16 +252,28 @@ struct abs2_impl
   }
 };
 
-template<typename RealScalar>
-struct abs2_impl<std::complex<RealScalar> >
+template<typename Scalar>
+struct abs2_impl_default<Scalar, true> // IsComplex
 {
+  typedef typename NumTraits<Scalar>::Real RealScalar;
   EIGEN_DEVICE_FUNC
-  static inline RealScalar run(const std::complex<RealScalar>& x)
+  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;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    return abs2_impl_default<Scalar,NumTraits<Scalar>::IsComplex>::run(x);
+  }
+};
+
 template<typename Scalar>
 struct abs2_retval
 {
@@ -314,8 +326,6 @@ struct hypot_impl
   typedef typename NumTraits<Scalar>::Real RealScalar;
   static inline RealScalar run(const Scalar& x, const Scalar& y)
   {
-    EIGEN_USING_STD_MATH(max);
-    EIGEN_USING_STD_MATH(min);
     EIGEN_USING_STD_MATH(abs);
     EIGEN_USING_STD_MATH(sqrt);
     RealScalar _x = abs(x);
@@ -607,8 +617,6 @@ struct random_default_impl<Scalar, false, true>
 {
   static inline Scalar run(const Scalar& x, const Scalar& y)
   { 
-    using std::max;
-    using std::min;
     typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
     if(y<x)
       return x;
@@ -667,6 +675,115 @@ inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random()
   return EIGEN_MATHFUNC_IMPL(random, Scalar)::run();
 }
 
+// Implementatin of is* functions
+
+// std::is* do not work with fast-math and gcc, std::is* are available on MSVC 2013 and newer, as well as in clang.
+#if (EIGEN_HAS_CXX11_MATH && !(EIGEN_COMP_GNUC_STRICT && __FINITE_MATH_ONLY__)) || (EIGEN_COMP_MSVC>=1800) || (EIGEN_COMP_CLANG)
+#define EIGEN_USE_STD_FPCLASSIFY 1
+#else
+#define EIGEN_USE_STD_FPCLASSIFY 0
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isnan_impl(const T&) { return false; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isinf_impl(const T&) { return false; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isfinite_impl(const T&) { return true; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isfinite_impl(const T& x)
+{
+  #if EIGEN_USE_STD_FPCLASSIFY
+    using std::isfinite;
+    return isfinite EIGEN_NOT_A_MACRO (x);
+  #else
+    return x<NumTraits<T>::highest() && x>NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isinf_impl(const T& x)
+{
+  #if EIGEN_USE_STD_FPCLASSIFY
+    using std::isinf;
+    return isinf EIGEN_NOT_A_MACRO (x);
+  #else
+    return x>NumTraits<T>::highest() || x<NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isnan_impl(const T& x)
+{
+  #if EIGEN_USE_STD_FPCLASSIFY
+    using std::isnan;
+    return isnan EIGEN_NOT_A_MACRO (x);
+  #else
+    return x != x;
+  #endif
+}
+
+#if (!EIGEN_USE_STD_FPCLASSIFY)
+
+#if EIGEN_COMP_MSVC
+
+template<typename T> EIGEN_DEVICE_FUNC bool isinf_msvc_helper(T x)
+{
+  return _fpclass(x)==_FPCLASS_NINF || _fpclass(x)==_FPCLASS_PINF;
+}
+
+//MSVC defines a _isnan builtin function, but for double only
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const long double& x) { return _isnan(x); }
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const double& x)      { return _isnan(x); }
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const float& x)       { return _isnan(x); }
+
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const long double& x) { return isinf_msvc_helper(x); }
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const double& x)      { return isinf_msvc_helper(x); }
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const float& x)       { return isinf_msvc_helper(x); }
+
+#elif (defined __FINITE_MATH_ONLY__ && __FINITE_MATH_ONLY__ && EIGEN_COMP_GNUC)
+
+#if EIGEN_GNUC_AT_LEAST(5,0)
+  #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((optimize("no-finite-math-only")))
+#else
+  // NOTE the inline qualifier and noinline attribute are both needed: the former is to avoid linking issue (duplicate symbol),
+  //      while the second prevent too aggressive optimizations in fast-math mode:
+  #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((noinline,optimize("no-finite-math-only")))
+#endif
+
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const long double& x) { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const double& x)      { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const float& x)       { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const double& x)      { return __builtin_isinf(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const float& x)       { return __builtin_isinf(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const long double& x) { return __builtin_isinf(x); }
+
+#undef EIGEN_TMP_NOOPT_ATTRIB
+
+#endif
+
+#endif
+
+// The following overload are defined at the end of this file
+template<typename T> bool isfinite_impl(const std::complex<T>& x);
+template<typename T> bool isnan_impl(const std::complex<T>& x);
+template<typename T> bool isinf_impl(const std::complex<T>& x);
+
 } // end namespace internal
 
 /****************************************************************************
@@ -810,59 +927,9 @@ inline EIGEN_MATHFUNC_RETVAL(pow, Scalar) pow(const Scalar& x, const Scalar& y)
   return EIGEN_MATHFUNC_IMPL(pow, Scalar)::run(x, y);
 }
 
-template<typename T>
-EIGEN_DEVICE_FUNC
-bool (isfinite)(const T& x)
-{
-  #if EIGEN_HAS_CXX11_MATH
-    using std::isfinite;
-    return isfinite EIGEN_NOT_A_MACRO (x);
-  #else
-    return x<NumTraits<T>::highest() && x>NumTraits<T>::lowest();
-  #endif
-}
-
-template<typename T>
-EIGEN_DEVICE_FUNC
-bool (isnan)(const T& x)
-{
-  #if EIGEN_HAS_CXX11_MATH
-    using std::isnan;
-    return isnan EIGEN_NOT_A_MACRO (x);
-  #else
-    return x != x;
-  #endif
-}
-
-template<typename T>
-EIGEN_DEVICE_FUNC
-bool (isinf)(const T& x)
-{
-  #if EIGEN_HAS_CXX11_MATH
-    using std::isinf;
-    return isinf EIGEN_NOT_A_MACRO (x);
-  #else
-    return x>NumTraits<T>::highest() || x<NumTraits<T>::lowest();
-  #endif
-}
-
-template<typename T>
-bool (isfinite)(const std::complex<T>& x)
-{
-  return (numext::isfinite)(numext::real(x)) && (numext::isfinite)(numext::imag(x));
-}
-
-template<typename T>
-bool (isnan)(const std::complex<T>& x)
-{
-  return (numext::isnan)(numext::real(x)) || (numext::isnan)(numext::imag(x));
-}
-
-template<typename T>
-bool (isinf)(const std::complex<T>& x)
-{
-  return ((numext::isinf)(numext::real(x)) || (numext::isinf)(numext::imag(x))) && (!(numext::isnan)(x));
-}
+template<typename T> EIGEN_DEVICE_FUNC bool (isnan)   (const T &x) { return internal::isnan_impl(x); }
+template<typename T> EIGEN_DEVICE_FUNC bool (isinf)   (const T &x) { return internal::isinf_impl(x); }
+template<typename T> EIGEN_DEVICE_FUNC bool (isfinite)(const T &x) { return internal::isfinite_impl(x); }
 
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
@@ -906,6 +973,24 @@ inline int log2(int x)
 
 namespace internal {
 
+template<typename T>
+bool isfinite_impl(const std::complex<T>& x)
+{
+  return (numext::isfinite)(numext::real(x)) && (numext::isfinite)(numext::imag(x));
+}
+
+template<typename T>
+bool isnan_impl(const std::complex<T>& x)
+{
+  return (numext::isnan)(numext::real(x)) || (numext::isnan)(numext::imag(x));
+}
+
+template<typename T>
+bool isinf_impl(const std::complex<T>& x)
+{
+  return ((numext::isinf)(numext::real(x)) || (numext::isinf)(numext::imag(x))) && (!(numext::isnan)(x));
+}
+
 /****************************************************************************
 * Implementation of fuzzy comparisons                                       *
 ****************************************************************************/
@@ -928,9 +1013,8 @@ struct scalar_fuzzy_default_impl<Scalar, false, false>
   EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
-    EIGEN_USING_STD_MATH(min);
     EIGEN_USING_STD_MATH(abs);
-    return abs(x - y) <= (min)(abs(x), abs(y)) * prec;
+    return abs(x - y) <= numext::mini(abs(x), abs(y)) * prec;
   }
   EIGEN_DEVICE_FUNC
   static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar& prec)
@@ -971,8 +1055,7 @@ struct scalar_fuzzy_default_impl<Scalar, true, false>
   }
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
-    EIGEN_USING_STD_MATH(min);
-    return numext::abs2(x - y) <= (min)(numext::abs2(x), numext::abs2(y)) * prec * prec;
+    return numext::abs2(x - y) <= numext::mini(numext::abs2(x), numext::abs2(y)) * prec * prec;
   }
 };
 

Eigen/src/Core/MatrixBase.h

@@ -328,23 +328,26 @@ template<typename Derived> class MatrixBase
 
 /////////// LU module ///////////
 
-    EIGEN_DEVICE_FUNC const FullPivLU<PlainObject> fullPivLu() const;
-    EIGEN_DEVICE_FUNC const PartialPivLU<PlainObject> partialPivLu() const;
-
-    const PartialPivLU<PlainObject> lu() const;
+    EIGEN_DEVICE_FUNC
+    inline const FullPivLU<PlainObject> fullPivLu() const;
+    EIGEN_DEVICE_FUNC
+    inline const PartialPivLU<PlainObject> partialPivLu() const;
 
     EIGEN_DEVICE_FUNC
-    const Inverse<Derived> inverse() const;
+    inline const PartialPivLU<PlainObject> lu() const;
+
+    EIGEN_DEVICE_FUNC
+    inline const Inverse<Derived> inverse() const;
     
     template<typename ResultType>
-    void computeInverseAndDetWithCheck(
+    inline void computeInverseAndDetWithCheck(
       ResultType& inverse,
       typename ResultType::Scalar& determinant,
       bool& invertible,
       const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
     ) const;
     template<typename ResultType>
-    void computeInverseWithCheck(
+    inline void computeInverseWithCheck(
       ResultType& inverse,
       bool& invertible,
       const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
@@ -353,22 +356,24 @@ template<typename Derived> class MatrixBase
 
 /////////// Cholesky module ///////////
 
-    const LLT<PlainObject>  llt() const;
-    const LDLT<PlainObject> ldlt() const;
+    inline const LLT<PlainObject>  llt() const;
+    inline const LDLT<PlainObject> ldlt() const;
 
 /////////// QR module ///////////
 
-    const HouseholderQR<PlainObject> householderQr() const;
-    const ColPivHouseholderQR<PlainObject> colPivHouseholderQr() const;
-    const FullPivHouseholderQR<PlainObject> fullPivHouseholderQr() const;
+    inline const HouseholderQR<PlainObject> householderQr() const;
+    inline const ColPivHouseholderQR<PlainObject> colPivHouseholderQr() const;
+    inline const FullPivHouseholderQR<PlainObject> fullPivHouseholderQr() const;
 
-    EigenvaluesReturnType eigenvalues() const;
-    RealScalar operatorNorm() const;
+/////////// Eigenvalues module ///////////
+
+    inline EigenvaluesReturnType eigenvalues() const;
+    inline RealScalar operatorNorm() const;
 
 /////////// SVD module ///////////
 
-    JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const;
-    BDCSVD<PlainObject>    bdcSvd(unsigned int computationOptions = 0) const;
+    inline JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const;
+    inline BDCSVD<PlainObject>    bdcSvd(unsigned int computationOptions = 0) const;
 
 /////////// Geometry module ///////////
 
@@ -381,24 +386,24 @@ template<typename Derived> class MatrixBase
     #endif // EIGEN_PARSED_BY_DOXYGEN
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    typename cross_product_return_type<OtherDerived>::type
+    inline typename cross_product_return_type<OtherDerived>::type
     cross(const MatrixBase<OtherDerived>& other) const;
     
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    PlainObject cross3(const MatrixBase<OtherDerived>& other) const;
+    inline PlainObject cross3(const MatrixBase<OtherDerived>& other) const;
     
     EIGEN_DEVICE_FUNC
-    PlainObject unitOrthogonal(void) const;
+    inline PlainObject unitOrthogonal(void) const;
     
-    Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const;
+    inline Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const;
     
-    ScalarMultipleReturnType operator*(const UniformScaling<Scalar>& s) const;
+    inline ScalarMultipleReturnType operator*(const UniformScaling<Scalar>& s) const;
     // put this as separate enum value to work around possible GCC 4.3 bug (?)
     enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1&&RowsAtCompileTime==1 ? ((internal::traits<Derived>::Flags&RowMajorBit)==RowMajorBit ? Horizontal : Vertical)
                                           : ColsAtCompileTime==1 ? Vertical : Horizontal };
     typedef Homogeneous<Derived, HomogeneousReturnTypeDirection> HomogeneousReturnType;
-    HomogeneousReturnType homogeneous() const;
+    inline HomogeneousReturnType homogeneous() const;
     
     enum {
       SizeMinusOne = SizeAtCompileTime==Dynamic ? Dynamic : SizeAtCompileTime-1
@@ -409,7 +414,7 @@ template<typename Derived> class MatrixBase
     typedef CwiseUnaryOp<internal::scalar_quotient1_op<typename internal::traits<Derived>::Scalar>,
                 const ConstStartMinusOne > HNormalizedReturnType;
 
-    const HNormalizedReturnType hnormalized() const;
+    inline const HNormalizedReturnType hnormalized() const;
 
 ////////// Householder module ///////////
 
@@ -433,6 +438,15 @@ 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/NumTraits.h

@@ -157,9 +157,9 @@ struct NumTraits<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
     IsInteger = NumTraits<Scalar>::IsInteger,
     IsSigned  = NumTraits<Scalar>::IsSigned,
     RequireInitialization = 1,
-    ReadCost = ArrayType::SizeAtCompileTime==Dynamic ? Dynamic : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::ReadCost,
-    AddCost  = ArrayType::SizeAtCompileTime==Dynamic ? Dynamic : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::AddCost,
-    MulCost  = ArrayType::SizeAtCompileTime==Dynamic ? Dynamic : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::MulCost
+    ReadCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::ReadCost,
+    AddCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::AddCost,
+    MulCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::MulCost
   };
   
   static inline RealScalar epsilon() { return NumTraits<RealScalar>::epsilon(); }

Eigen/src/Core/PermutationMatrix.h

@@ -2,7 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2009-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -13,9 +13,6 @@
 
 namespace Eigen { 
 
-// TODO: this does not seems to be needed at all:
-// template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKind> class PermutedImpl;
-
 /** \class PermutationBase
   * \ingroup Core_Module
   *
@@ -67,8 +64,10 @@ class PermutationBase : public EigenBase<Derived>
             DenseMatrixType;
     typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,StorageIndex>
             PlainPermutationType;
+    typedef PlainPermutationType PlainObject;
     using Base::derived;
-    typedef Transpose<PermutationBase> TransposeReturnType;
+    typedef Inverse<Derived> InverseReturnType;
+    typedef void Scalar;
     #endif
 
     /** Copies the other permutation into *this */
@@ -195,14 +194,14 @@ class PermutationBase : public EigenBase<Derived>
       *
       * \note \note_try_to_help_rvo
       */
-    inline TransposeReturnType inverse() const
-    { return TransposeReturnType(derived()); }
+    inline InverseReturnType inverse() const
+    { return InverseReturnType(derived()); }
     /** \returns the tranpose permutation matrix.
       *
       * \note \note_try_to_help_rvo
       */
-    inline TransposeReturnType transpose() const
-    { return TransposeReturnType(derived()); }
+    inline InverseReturnType transpose() const
+    { return InverseReturnType(derived()); }
 
     /**** multiplication helpers to hopefully get RVO ****/
 
@@ -237,7 +236,7 @@ class PermutationBase : public EigenBase<Derived>
       * \note \note_try_to_help_rvo
       */
     template<typename Other>
-    inline PlainPermutationType operator*(const Transpose<PermutationBase<Other> >& other) const
+    inline PlainPermutationType operator*(const InverseImpl<Other,PermutationStorage>& other) const
     { return PlainPermutationType(internal::PermPermProduct, *this, other.eval()); }
 
     /** \returns the product of an inverse permutation with another permutation.
@@ -245,7 +244,7 @@ class PermutationBase : public EigenBase<Derived>
       * \note \note_try_to_help_rvo
       */
     template<typename Other> friend
-    inline PlainPermutationType operator*(const Transpose<PermutationBase<Other> >& other, const PermutationBase& perm)
+    inline PlainPermutationType operator*(const InverseImpl<Other, PermutationStorage>& 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.
@@ -303,6 +302,7 @@ struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _Storag
   typedef PermutationStorage StorageKind;
   typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
   typedef _StorageIndex StorageIndex;
+  typedef void Scalar;
 };
 }
 
@@ -396,13 +396,13 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Other>
-    PermutationMatrix(const Transpose<PermutationBase<Other> >& other)
-      : m_indices(other.nestedExpression().size())
+    PermutationMatrix(const InverseImpl<Other,PermutationStorage>& other)
+      : m_indices(other.derived().nestedExpression().size())
     {
       eigen_internal_assert(m_indices.size() <= NumTraits<StorageIndex>::highest());
       StorageIndex end = StorageIndex(m_indices.size());
       for (StorageIndex i=0; i<end;++i)
-        m_indices.coeffRef(other.nestedExpression().indices().coeff(i)) = i;
+        m_indices.coeffRef(other.derived().nestedExpression().indices().coeff(i)) = i;
     }
     template<typename Lhs,typename Rhs>
     PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs)
@@ -426,6 +426,7 @@ struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _St
   typedef PermutationStorage StorageKind;
   typedef Map<const Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
   typedef _StorageIndex StorageIndex;
+  typedef void Scalar;
 };
 }
 
@@ -499,7 +500,7 @@ template<typename _IndicesType>
 struct traits<PermutationWrapper<_IndicesType> >
 {
   typedef PermutationStorage StorageKind;
-  typedef typename _IndicesType::Scalar Scalar;
+  typedef void Scalar;
   typedef typename _IndicesType::Scalar StorageIndex;
   typedef _IndicesType IndicesType;
   enum {
@@ -561,84 +562,61 @@ operator*(const PermutationBase<PermutationDerived> &permutation,
             (permutation.derived(), matrix.derived());
 }
 
-namespace internal {
 
-/* Template partial specialization for transposed/inverse permutations */
-
-template<typename Derived>
-struct traits<Transpose<PermutationBase<Derived> > >
- : traits<Derived>
-{};
-
-} // end namespace internal
-
-// TODO: the specificties should be handled by the evaluator,
-// at the very least we should only specialize TransposeImpl
-template<typename Derived>
-class Transpose<PermutationBase<Derived> >
-  : public EigenBase<Transpose<PermutationBase<Derived> > >
+template<typename PermutationType>
+class InverseImpl<PermutationType, PermutationStorage>
+  : public EigenBase<Inverse<PermutationType> >
 {
-    typedef Derived PermutationType;
-    typedef typename PermutationType::IndicesType IndicesType;
     typedef typename PermutationType::PlainPermutationType PlainPermutationType;
+    typedef internal::traits<PermutationType> PermTraits;
+  protected:
+    InverseImpl() {}
   public:
+    typedef Inverse<PermutationType> InverseType;
+    using EigenBase<Inverse<PermutationType> >::derived;
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
-    typedef internal::traits<PermutationType> Traits;
-    typedef typename Derived::DenseMatrixType DenseMatrixType;
+    typedef typename PermutationType::DenseMatrixType DenseMatrixType;
     enum {
-      Flags = Traits::Flags,
-      RowsAtCompileTime = Traits::RowsAtCompileTime,
-      ColsAtCompileTime = Traits::ColsAtCompileTime,
-      MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
-      MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
+      RowsAtCompileTime = PermTraits::RowsAtCompileTime,
+      ColsAtCompileTime = PermTraits::ColsAtCompileTime,
+      MaxRowsAtCompileTime = PermTraits::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = PermTraits::MaxColsAtCompileTime
     };
-    typedef typename Traits::Scalar Scalar;
-    typedef typename Traits::StorageIndex StorageIndex;
     #endif
 
-    Transpose(const PermutationType& p) : m_permutation(p) {}
-
-    inline Index rows() const { return m_permutation.rows(); }
-    inline Index cols() const { return m_permutation.cols(); }
-
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename DenseDerived>
     void evalTo(MatrixBase<DenseDerived>& other) const
     {
       other.setZero();
-      for (Index i=0; i<rows();++i)
-        other.coeffRef(i, m_permutation.indices().coeff(i)) = typename DenseDerived::Scalar(1);
+      for (Index i=0; i<derived().rows();++i)
+        other.coeffRef(i, derived().nestedExpression().indices().coeff(i)) = typename DenseDerived::Scalar(1);
     }
     #endif
 
     /** \return the equivalent permutation matrix */
-    PlainPermutationType eval() const { return *this; }
+    PlainPermutationType eval() const { return derived(); }
 
-    DenseMatrixType toDenseMatrix() const { return *this; }
+    DenseMatrixType toDenseMatrix() const { return derived(); }
 
     /** \returns the matrix with the inverse permutation applied to the columns.
       */
     template<typename OtherDerived> friend
-    const Product<OtherDerived, Transpose, AliasFreeProduct>
-    operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trPerm)
+    const Product<OtherDerived, InverseType, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix, const InverseType& trPerm)
     {
-      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trPerm.derived());
+      return Product<OtherDerived, InverseType, AliasFreeProduct>(matrix.derived(), trPerm.derived());
     }
 
     /** \returns the matrix with the inverse permutation applied to the rows.
       */
     template<typename OtherDerived>
-    const Product<Transpose, OtherDerived, AliasFreeProduct>
+    const Product<InverseType, OtherDerived, AliasFreeProduct>
     operator*(const MatrixBase<OtherDerived>& matrix) const
     {
-      return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived());
+      return Product<InverseType, OtherDerived, AliasFreeProduct>(derived(), matrix.derived());
     }
-
-    const PermutationType& nestedExpression() const { return m_permutation; }
-
-  protected:
-    const PermutationType& m_permutation;
 };
 
 template<typename Derived>

Eigen/src/Core/PlainObjectBase.h

@@ -263,7 +263,6 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
         m_storage.resize(size, rows, cols);
         if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
       #else
-        internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols);
         m_storage.resize(rows*cols, rows, cols);
       #endif
     }
@@ -450,6 +449,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       return Base::operator=(func);
     }
 
+    // Prevent user from trying to instantiate PlainObjectBase objects
+    // by making all its constructor protected. See bug 1074.
+  protected:
+
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE PlainObjectBase() : m_storage()
     {
@@ -496,17 +499,6 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 //       EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
     }
 
-    /** \copydoc MatrixBase::operator=(const EigenBase<OtherDerived>&)
-      */
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC 
-    EIGEN_STRONG_INLINE Derived& operator=(const EigenBase<OtherDerived> &other)
-    {
-      _resize_to_match(other);
-      Base::operator=(other.derived());
-      return this->derived();
-    }
-
     /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
@@ -520,7 +512,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 
     /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
       : m_storage()
     {
@@ -539,6 +531,19 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       other.evalTo(this->derived());
     }
 
+  public:
+
+    /** \copydoc MatrixBase::operator=(const EigenBase<OtherDerived>&)
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE Derived& operator=(const EigenBase<OtherDerived> &other)
+    {
+      _resize_to_match(other);
+      Base::operator=(other.derived());
+      return this->derived();
+    }
+
     /** \name Map
       * These are convenience functions returning Map objects. The Map() static functions return unaligned Map objects,
       * while the AlignedMap() functions return aligned Map objects and thus should be called only with 16-byte-aligned

Eigen/src/Core/Product.h

@@ -217,29 +217,6 @@ class ProductImpl<Lhs,Rhs,Option,Dense>
   
 };
 
-/***************************************************************************
-* Implementation of matrix base methods
-***************************************************************************/
-
-
-/** \internal used to test the evaluator only
-  */
-template<typename Lhs,typename Rhs>
-const Product<Lhs,Rhs>
-prod(const Lhs& lhs, const Rhs& rhs)
-{
-  return Product<Lhs,Rhs>(lhs,rhs);
-}
-
-/** \internal used to test the evaluator only
-  */
-template<typename Lhs,typename Rhs>
-const Product<Lhs,Rhs,LazyProduct>
-lazyprod(const Lhs& lhs, const Rhs& rhs)
-{
-  return Product<Lhs,Rhs,LazyProduct>(lhs,rhs);
-}
-
 } // end namespace Eigen
 
 #endif // EIGEN_PRODUCT_H

Eigen/src/Core/ProductEvaluators.h

@@ -38,6 +38,12 @@ struct evaluator<Product<Lhs, Rhs, Options> >
 // Catch scalar * ( A * B ) and transform it to (A*scalar) * B
 // TODO we should apply that rule only if that's really helpful
 template<typename Lhs, typename Rhs, typename Scalar>
+struct evaluator_traits<CwiseUnaryOp<internal::scalar_multiple_op<Scalar>,  const Product<Lhs, Rhs, DefaultProduct>  > >
+ : evaluator_traits_base<CwiseUnaryOp<internal::scalar_multiple_op<Scalar>,  const Product<Lhs, Rhs, DefaultProduct>  > >
+{
+  enum { AssumeAliasing = 1 };
+};
+template<typename Lhs, typename Rhs, typename Scalar>
 struct evaluator<CwiseUnaryOp<internal::scalar_multiple_op<Scalar>,  const Product<Lhs, Rhs, DefaultProduct>  > > 
  : public evaluator<Product<CwiseUnaryOp<internal::scalar_multiple_op<Scalar>,const Lhs>, Rhs, DefaultProduct> >
 {
@@ -91,8 +97,7 @@ struct evaluator_traits<Product<Lhs, Rhs, AliasFreeProduct> >
 // This is the default evaluator implementation for products:
 // It creates a temporary and call generic_product_impl
 template<typename Lhs, typename Rhs, int Options, int ProductTag, typename LhsShape, typename RhsShape>
-struct product_evaluator<Product<Lhs, Rhs, Options>, ProductTag, LhsShape, RhsShape, typename traits<Lhs>::Scalar, typename traits<Rhs>::Scalar,
-  EnableIf<(Options==DefaultProduct || Options==AliasFreeProduct)> >
+struct product_evaluator<Product<Lhs, Rhs, Options>, ProductTag, LhsShape, RhsShape>
   : public evaluator<typename Product<Lhs, Rhs, Options>::PlainObject>
 {
   typedef Product<Lhs, Rhs, Options> XprType;
@@ -177,11 +182,41 @@ struct Assignment<DstXprType, CwiseUnaryOp<internal::scalar_multiple_op<ScalarBi
                        const Product<Lhs,Rhs,DefaultProduct> > SrcXprType;
   static void run(DstXprType &dst, const SrcXprType &src, const AssignFunc& func)
   {
-    // TODO use operator* instead of prod() once we have made enough progress
-    call_assignment(dst.noalias(), prod(src.functor().m_other * src.nestedExpression().lhs(), src.nestedExpression().rhs()), func);
+    call_assignment_no_alias(dst, (src.functor().m_other * src.nestedExpression().lhs())*src.nestedExpression().rhs(), func);
   }
 };
 
+//----------------------------------------
+// Catch "Dense ?= xpr + Product<>" expression to save one temporary
+// FIXME we could probably enable these rules for any product, i.e., not only Dense and DefaultProduct
+
+template<typename DstXprType, typename OtherXpr, typename ProductType, typename Scalar, typename Func1, typename Func2>
+struct assignment_from_xpr_plus_product
+{
+  typedef CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const OtherXpr, const ProductType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const Func1& func)
+  {
+    call_assignment_no_alias(dst, src.lhs(), func);
+    call_assignment_no_alias(dst, src.rhs(), Func2());
+  }
+};
+
+template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const OtherXpr,
+                                           const Product<Lhs,Rhs,DefaultProduct> >, internal::assign_op<Scalar>, Dense2Dense>
+  : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, Scalar, internal::assign_op<Scalar>, internal::add_assign_op<Scalar> >
+{};
+template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const OtherXpr,
+                                           const Product<Lhs,Rhs,DefaultProduct> >, internal::add_assign_op<Scalar>, Dense2Dense>
+  : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, Scalar, internal::add_assign_op<Scalar>, internal::add_assign_op<Scalar> >
+{};
+template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const OtherXpr,
+                                           const Product<Lhs,Rhs,DefaultProduct> >, internal::sub_assign_op<Scalar>, Dense2Dense>
+  : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, Scalar, internal::sub_assign_op<Scalar>, internal::sub_assign_op<Scalar> >
+{};
+//----------------------------------------
 
 template<typename Lhs, typename Rhs>
 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,InnerProduct>
@@ -213,12 +248,12 @@ template<typename Dst, typename Lhs, typename Rhs, typename Func>
 EIGEN_DONT_INLINE void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&)
 {
   evaluator<Rhs> rhsEval(rhs);
-  // FIXME make sure lhs is sequentially stored
+  typename nested_eval<Lhs,Rhs::SizeAtCompileTime>::type actual_lhs(lhs);
+  // FIXME if cols is large enough, then it might be useful to make sure that lhs is sequentially stored
   // FIXME not very good if rhs is real and lhs complex while alpha is real too
-  // FIXME we should probably build an evaluator for dst
   const Index cols = dst.cols();
   for (Index j=0; j<cols; ++j)
-    func(dst.col(j), rhsEval.coeff(0,j) * lhs);
+    func(dst.col(j), rhsEval.coeff(0,j) * actual_lhs);
 }
 
 // Row major result
@@ -226,18 +261,18 @@ template<typename Dst, typename Lhs, typename Rhs, typename Func>
 EIGEN_DONT_INLINE void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&)
 {
   evaluator<Lhs> lhsEval(lhs);
-  // FIXME make sure rhs is sequentially stored
+  typename nested_eval<Rhs,Lhs::SizeAtCompileTime>::type actual_rhs(rhs);
+  // FIXME if rows is large enough, then it might be useful to make sure that rhs is sequentially stored
   // FIXME not very good if lhs is real and rhs complex while alpha is real too
-  // FIXME we should probably build an evaluator for dst
   const Index rows = dst.rows();
   for (Index i=0; i<rows; ++i)
-    func(dst.row(i), lhsEval.coeff(i,0) * rhs);
+    func(dst.row(i), lhsEval.coeff(i,0) * actual_rhs);
 }
 
 template<typename Lhs, typename Rhs>
 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,OuterProduct>
 {
-  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 {};
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   
   // TODO it would be nice to be able to exploit our *_assign_op functors for that purpose
@@ -255,25 +290,25 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,OuterProduct>
   template<typename Dst>
   static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    internal::outer_product_selector_run(dst, lhs, rhs, set(), IsRowMajor<Dst>());
+    internal::outer_product_selector_run(dst, lhs, rhs, set(), is_row_major<Dst>());
   }
   
   template<typename Dst>
   static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    internal::outer_product_selector_run(dst, lhs, rhs, add(), IsRowMajor<Dst>());
+    internal::outer_product_selector_run(dst, lhs, rhs, add(), is_row_major<Dst>());
   }
   
   template<typename Dst>
   static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    internal::outer_product_selector_run(dst, lhs, rhs, sub(), IsRowMajor<Dst>());
+    internal::outer_product_selector_run(dst, lhs, rhs, sub(), is_row_major<Dst>());
   }
   
   template<typename Dst>
   static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   {
-    internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), IsRowMajor<Dst>());
+    internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), is_row_major<Dst>());
   }
   
 };
@@ -314,7 +349,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct>
   template<typename Dest>
   static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   {
-    internal::gemv_dense_sense_selector<Side,
+    internal::gemv_dense_selector<Side,
                             (int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
                             bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)
                            >::run(lhs, rhs, dst, alpha);
@@ -329,28 +364,28 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
   template<typename Dst>
   static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    // TODO: use the following instead of calling call_assignment, same for the other methods
-    // dst = lazyprod(lhs,rhs);
-    call_assignment(dst, lazyprod(lhs,rhs), internal::assign_op<Scalar>());
+    // Same as: dst.noalias() = lhs.lazyProduct(rhs);
+    // but easier on the compiler side
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::assign_op<Scalar>());
   }
   
   template<typename Dst>
   static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    // dst += lazyprod(lhs,rhs);
-    call_assignment(dst, lazyprod(lhs,rhs), internal::add_assign_op<Scalar>());
+    // dst.noalias() += lhs.lazyProduct(rhs);
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::add_assign_op<Scalar>());
   }
   
   template<typename Dst>
   static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    // dst -= lazyprod(lhs,rhs);
-    call_assignment(dst, lazyprod(lhs,rhs), internal::sub_assign_op<Scalar>());
+    // dst.noalias() -= lhs.lazyProduct(rhs);
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<Scalar>());
   }
   
 //   template<typename Dst>
 //   static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
-//   { dst += alpha * lazyprod(lhs,rhs); }
+//   { dst.noalias() += alpha * lhs.lazyProduct(rhs); }
 };
 
 // This specialization enforces the use of a coefficient-based evaluation strategy
@@ -371,7 +406,7 @@ template<int StorageOrder, int UnrollingIndex, typename Lhs, typename Rhs, typen
 struct etor_product_packet_impl;
 
 template<typename Lhs, typename Rhs, int ProductTag>
-struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar > 
+struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, DenseShape>
     : evaluator_base<Product<Lhs, Rhs, LazyProduct> >
 {
   typedef Product<Lhs, Rhs, LazyProduct> XprType;
@@ -387,7 +422,11 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
       m_rhsImpl(m_rhs),     //       Moreover, they are only useful for the packet path, so we could completely disable them when not needed,
                             //       or perhaps declare them on the fly on the packet method... We have experiment to check what's best.
       m_innerDim(xpr.lhs().cols())
-  { }
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::AddCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
 
   // Everything below here is taken from CoeffBasedProduct.h
 
@@ -408,15 +447,15 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
     MaxColsAtCompileTime = RhsNestedCleaned::MaxColsAtCompileTime,
       
     PacketSize = packet_traits<Scalar>::size,
-    
+
     LhsCoeffReadCost = LhsEtorType::CoeffReadCost,
     RhsCoeffReadCost = RhsEtorType::CoeffReadCost,
     CoeffReadCost = InnerSize==0 ? NumTraits<Scalar>::ReadCost
-                  : (InnerSize == Dynamic || LhsCoeffReadCost==Dynamic || RhsCoeffReadCost==Dynamic || NumTraits<Scalar>::AddCost==Dynamic || NumTraits<Scalar>::MulCost==Dynamic) ? Dynamic
+                  : InnerSize == Dynamic ? HugeCost
                   : InnerSize * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost)
                     + (InnerSize - 1) * NumTraits<Scalar>::AddCost,
 
-    Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT,
+    Unroll = CoeffReadCost <= EIGEN_UNROLLING_LIMIT,
     
     LhsFlags = LhsEtorType::Flags,
     RhsFlags = RhsEtorType::Flags,
@@ -424,19 +463,16 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
     LhsAlignment = LhsEtorType::Alignment,
     RhsAlignment = RhsEtorType::Alignment,
     
-    LhsIsAligned = int(LhsAlignment) >= int(unpacket_traits<PacketScalar>::alignment),
-    RhsIsAligned = int(RhsAlignment) >= int(unpacket_traits<PacketScalar>::alignment),
-    
     LhsRowMajor = LhsFlags & RowMajorBit,
     RhsRowMajor = RhsFlags & RowMajorBit,
       
     SameType = is_same<typename LhsNestedCleaned::Scalar,typename RhsNestedCleaned::Scalar>::value,
 
     CanVectorizeRhs = RhsRowMajor && (RhsFlags & PacketAccessBit)
-                    && (ColsAtCompileTime == Dynamic || ( (ColsAtCompileTime % PacketSize) == 0 && RhsIsAligned ) ),
+                    && (ColsAtCompileTime == Dynamic || ((ColsAtCompileTime % PacketSize) == 0) ),
 
     CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit)
-                    && (RowsAtCompileTime == Dynamic || ( (RowsAtCompileTime % PacketSize) == 0 && LhsIsAligned ) ),
+                    && (RowsAtCompileTime == Dynamic || ((RowsAtCompileTime % PacketSize) == 0) ),
 
     EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
                     : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
@@ -445,12 +481,16 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
     Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit)
           | (EvalToRowMajor ? RowMajorBit : 0)
           // TODO enable vectorization for mixed types
-          | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0),
+          | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0)
+          | (XprType::IsVectorAtCompileTime ? LinearAccessBit : 0),
           
-    Alignment = CanVectorizeLhs ? LhsAlignment
-              : CanVectorizeRhs ? RhsAlignment
+    LhsOuterStrideBytes = int(LhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename LhsNestedCleaned::Scalar)),
+    RhsOuterStrideBytes = int(RhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename RhsNestedCleaned::Scalar)),
+
+    Alignment = CanVectorizeLhs ? (LhsOuterStrideBytes<0 || (int(LhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,LhsAlignment))!=0 ? 0 : LhsAlignment)
+              : CanVectorizeRhs ? (RhsOuterStrideBytes<0 || (int(RhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,RhsAlignment))!=0 ? 0 : RhsAlignment)
               : 0,
-          
+
     /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside
     * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner
     * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect
@@ -460,13 +500,11 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
                         && LhsRowMajor
                         && (!RhsRowMajor)
                         && (LhsFlags & RhsFlags & ActualPacketAccessBit)
-                        && (LhsIsAligned && RhsIsAligned)
                         && (InnerSize % packet_traits<Scalar>::size == 0)
   };
   
-  EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index row, Index col) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index row, Index col) const
   {
-    // TODO check performance regression wrt to Eigen 3.2 which has special handling of this function
     return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum();
   }
 
@@ -478,7 +516,6 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
   {
     const Index row = RowsAtCompileTime == 1 ? 0 : index;
     const Index col = RowsAtCompileTime == 1 ? index : 0;
-    // TODO check performance regression wrt to Eigen 3.2 which has special handling of this function
     return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum();
   }
 
@@ -486,14 +523,21 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
   const PacketType packet(Index row, Index col) const
   {
     PacketType res;
-    typedef etor_product_packet_impl<Flags&RowMajorBit ? RowMajor : ColMajor,
-                                     Unroll ? InnerSize : Dynamic,
+    typedef etor_product_packet_impl<bool(int(Flags)&RowMajorBit) ? RowMajor : ColMajor,
+                                     Unroll ? int(InnerSize) : Dynamic,
                                      LhsEtorType, RhsEtorType, PacketType, LoadMode> PacketImpl;
-
     PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res);
     return res;
   }
 
+  template<int LoadMode, typename PacketType>
+  const PacketType packet(Index index) const
+  {
+    const Index row = RowsAtCompileTime == 1 ? 0 : index;
+    const Index col = RowsAtCompileTime == 1 ? index : 0;
+    return packet<LoadMode,PacketType>(row,col);
+  }
+
 protected:
   const LhsNested m_lhs;
   const RhsNested m_rhs;
@@ -506,12 +550,12 @@ protected:
 };
 
 template<typename Lhs, typename Rhs>
-struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, LazyCoeffBasedProductMode, DenseShape, DenseShape, typename traits<Lhs>::Scalar, typename traits<Rhs>::Scalar > 
-  : product_evaluator<Product<Lhs, Rhs, LazyProduct>, CoeffBasedProductMode, DenseShape, DenseShape, typename traits<Lhs>::Scalar, typename traits<Rhs>::Scalar >
+struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, LazyCoeffBasedProductMode, DenseShape, DenseShape>
+  : product_evaluator<Product<Lhs, Rhs, LazyProduct>, CoeffBasedProductMode, DenseShape, DenseShape>
 {
   typedef Product<Lhs, Rhs, DefaultProduct> XprType;
   typedef Product<Lhs, Rhs, LazyProduct> BaseProduct;
-  typedef product_evaluator<BaseProduct, CoeffBasedProductMode, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar > Base;
+  typedef product_evaluator<BaseProduct, CoeffBasedProductMode, DenseShape, DenseShape> Base;
   enum {
     Flags = Base::Flags | EvalBeforeNestingBit
   };
@@ -703,6 +747,8 @@ public:
   diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag)
     : m_diagImpl(diag), m_matImpl(mat)
   {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
   }
   
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const
@@ -735,7 +781,7 @@ protected:
 
 // diagonal * dense
 template<typename Lhs, typename Rhs, int ProductKind, int ProductTag>
-struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> 
+struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalShape, DenseShape>
   : diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft>
 {
   typedef diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft> Base;
@@ -781,7 +827,7 @@ struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalSha
 
 // dense * diagonal
 template<typename Lhs, typename Rhs, int ProductKind, int ProductTag>
-struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DenseShape, DiagonalShape, typename Lhs::Scalar, typename Rhs::Scalar> 
+struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DenseShape, DiagonalShape>
   : diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight>
 {
   typedef diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight> Base;
@@ -911,20 +957,20 @@ struct generic_product_impl<Lhs, Rhs, MatrixShape, PermutationShape, ProductTag>
 };
 
 template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
-struct generic_product_impl<Transpose<Lhs>, Rhs, PermutationShape, MatrixShape, ProductTag>
+struct generic_product_impl<Inverse<Lhs>, Rhs, PermutationShape, MatrixShape, ProductTag>
 {
   template<typename Dest>
-  static void evalTo(Dest& dst, const Transpose<Lhs>& lhs, const Rhs& rhs)
+  static void evalTo(Dest& dst, const Inverse<Lhs>& lhs, const Rhs& rhs)
   {
     permutation_matrix_product<Rhs, OnTheLeft, true, MatrixShape>::run(dst, lhs.nestedExpression(), rhs);
   }
 };
 
 template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
-struct generic_product_impl<Lhs, Transpose<Rhs>, MatrixShape, PermutationShape, ProductTag>
+struct generic_product_impl<Lhs, Inverse<Rhs>, MatrixShape, PermutationShape, ProductTag>
 {
   template<typename Dest>
-  static void evalTo(Dest& dst, const Lhs& lhs, const Transpose<Rhs>& rhs)
+  static void evalTo(Dest& dst, const Lhs& lhs, const Inverse<Rhs>& rhs)
   {
     permutation_matrix_product<Lhs, OnTheRight, true, MatrixShape>::run(dst, rhs.nestedExpression(), lhs);
   }

Eigen/src/Core/Redux.h

@@ -50,20 +50,14 @@ public:
 
 public:
   enum {
-    Cost = (  Derived::SizeAtCompileTime == Dynamic
-           || Derived::CoeffReadCost == Dynamic
-           || (Derived::SizeAtCompileTime!=1 && functor_traits<Func>::Cost == Dynamic)
-           ) ? Dynamic
-           : Derived::SizeAtCompileTime * Derived::CoeffReadCost
-               + (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
+    Cost = Derived::SizeAtCompileTime == Dynamic ? HugeCost
+         : Derived::SizeAtCompileTime * Derived::CoeffReadCost + (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
     UnrollingLimit = EIGEN_UNROLLING_LIMIT * (int(Traversal) == int(DefaultTraversal) ? 1 : int(PacketSize))
   };
 
 public:
   enum {
-    Unrolling = Cost != Dynamic && Cost <= UnrollingLimit
-              ? CompleteUnrolling
-              : NoUnrolling
+    Unrolling = Cost <= UnrollingLimit ? CompleteUnrolling : NoUnrolling
   };
   
 #ifdef EIGEN_DEBUG_ASSIGN
@@ -269,8 +263,9 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
   }
 };
 
-template<typename Func, typename Derived>
-struct redux_impl<Func, Derived, SliceVectorizedTraversal, NoUnrolling>
+// NOTE: for SliceVectorizedTraversal we simply bypass unrolling
+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 PacketType;
@@ -414,17 +409,7 @@ typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::redux(const Func& func) const
 {
   eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix");
-  
-  // FIXME, eval_nest should be handled by redux_evaluator, however:
-  //  - it is currently difficult to provide the right Flags since they are still handled by the expressions
-  //  - handling it here might reduce the number of template instantiations
-//   typedef typename internal::nested_eval<Derived,1>::type ThisNested;
-//   typedef typename internal::remove_all<ThisNested>::type ThisNestedCleaned;
-//   typedef typename internal::redux_evaluator<ThisNestedCleaned> ThisEvaluator;
-//   
-//   ThisNested thisNested(derived());
-//   ThisEvaluator thisEval(thisNested);
-  
+
   typedef typename internal::redux_evaluator<Derived> ThisEvaluator;
   ThisEvaluator thisEval(derived());
   

Eigen/src/Core/Solve.h

@@ -34,12 +34,11 @@ template<typename Decomposition, typename RhsType,typename StorageKind> struct s
 template<typename Decomposition, typename RhsType>
 struct solve_traits<Decomposition,RhsType,Dense>
 {
-  typedef typename Decomposition::MatrixType MatrixType;
   typedef Matrix<typename RhsType::Scalar,
-                 MatrixType::ColsAtCompileTime,
+                 Decomposition::ColsAtCompileTime,
                  RhsType::ColsAtCompileTime,
                  RhsType::PlainObject::Options,
-                 MatrixType::MaxColsAtCompileTime,
+                 Decomposition::MaxColsAtCompileTime,
                  RhsType::MaxColsAtCompileTime> PlainObject;  
 };
 
@@ -52,7 +51,7 @@ struct traits<Solve<Decomposition, RhsType> >
   typedef traits<PlainObject> BaseTraits;
   enum {
     Flags = BaseTraits::Flags & RowMajorBit,
-    CoeffReadCost = Dynamic
+    CoeffReadCost = HugeCost
   };
 };
 
@@ -118,6 +117,8 @@ struct evaluator<Solve<Decomposition,RhsType> >
   typedef Solve<Decomposition,RhsType> SolveType;
   typedef typename SolveType::PlainObject PlainObject;
   typedef evaluator<PlainObject> Base;
+
+  enum { Flags = Base::Flags | EvalBeforeNestingBit };
   
   EIGEN_DEVICE_FUNC explicit evaluator(const SolveType& solve)
     : m_result(solve.rows(), solve.cols())
@@ -143,6 +144,28 @@ struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar
   }
 };
 
+// Specialization for "dst = dec.transpose().solve(rhs)"
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<Transpose<const DecType>,RhsType>, internal::assign_op<Scalar>, Dense2Dense, Scalar>
+{
+  typedef Solve<Transpose<const DecType>,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar> &)
+  {
+    src.dec().nestedExpression().template _solve_impl_transposed<false>(src.rhs(), dst);
+  }
+};
+
+// Specialization for "dst = dec.adjoint().solve(rhs)"
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType>, internal::assign_op<Scalar>, Dense2Dense, Scalar>
+{
+  typedef Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar> &)
+  {
+    src.dec().nestedExpression().nestedExpression().template _solve_impl_transposed<true>(src.rhs(), dst);
+  }
+};
+
 } // end namepsace internal
 
 } // end namespace Eigen

Eigen/src/Core/SolveTriangular.h

@@ -107,32 +107,32 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
 * meta-unrolling implementation
 ***************************************************************************/
 
-template<typename Lhs, typename Rhs, int Mode, int Index, int Size,
-         bool Stop = Index==Size>
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size,
+         bool Stop = LoopIndex==Size>
 struct triangular_solver_unroller;
 
-template<typename Lhs, typename Rhs, int Mode, int Index, int Size>
-struct triangular_solver_unroller<Lhs,Rhs,Mode,Index,Size,false> {
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,false> {
   enum {
     IsLower = ((Mode&Lower)==Lower),
-    I = IsLower ? Index : Size - Index - 1,
-    S = IsLower ? 0     : I+1
+    DiagIndex  = IsLower ? LoopIndex : Size - LoopIndex - 1,
+    StartIndex = IsLower ? 0         : DiagIndex+1
   };
   static void run(const Lhs& lhs, Rhs& rhs)
   {
-    if (Index>0)
-      rhs.coeffRef(I) -= lhs.row(I).template segment<Index>(S).transpose()
-                         .cwiseProduct(rhs.template segment<Index>(S)).sum();
+    if (LoopIndex>0)
+      rhs.coeffRef(DiagIndex) -= lhs.row(DiagIndex).template segment<LoopIndex>(StartIndex).transpose()
+                                .cwiseProduct(rhs.template segment<LoopIndex>(StartIndex)).sum();
 
     if(!(Mode & UnitDiag))
-      rhs.coeffRef(I) /= lhs.coeff(I,I);
+      rhs.coeffRef(DiagIndex) /= lhs.coeff(DiagIndex,DiagIndex);
 
-    triangular_solver_unroller<Lhs,Rhs,Mode,Index+1,Size>::run(lhs,rhs);
+    triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex+1,Size>::run(lhs,rhs);
   }
 };
 
-template<typename Lhs, typename Rhs, int Mode, int Index, int Size>
-struct triangular_solver_unroller<Lhs,Rhs,Mode,Index,Size,true> {
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,true> {
   static void run(const Lhs&, Rhs&) {}
 };
 
@@ -161,13 +161,6 @@ struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
 * TriangularView methods
 ***************************************************************************/
 
-/** "in-place" version of TriangularView::solve() where the result is written in \a other
-  *
-  * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
-  * This function will const_cast it, so constness isn't honored here.
-  *
-  * See TriangularView:solve() for the details.
-  */
 template<typename MatrixType, unsigned int Mode>
 template<int Side, typename OtherDerived>
 void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
@@ -188,27 +181,6 @@ void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<Ot
     other = otherCopy;
 }
 
-/** \returns the product of the inverse of \c *this with \a other, \a *this being triangular.
-  *
-  * This function computes the inverse-matrix matrix product inverse(\c *this) * \a other if
-  * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
-  * \a Side==OnTheRight.
-  *
-  * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
-  * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
-  * is an upper (resp. lower) triangular matrix.
-  *
-  * Example: \include Triangular_solve.cpp
-  * Output: \verbinclude Triangular_solve.out
-  *
-  * This function returns an expression of the inverse-multiply and can works in-place if it is assigned
-  * to the same matrix or vector \a other.
-  *
-  * For users coming from BLAS, this function (and more specifically solveInPlace()) offer
-  * all the operations supported by the \c *TRSV and \c *TRSM BLAS routines.
-  *
-  * \sa TriangularView::solveInPlace()
-  */
 template<typename Derived, unsigned int Mode>
 template<int Side, typename Other>
 const internal::triangular_solve_retval<Side,TriangularView<Derived,Mode>,Other>

Eigen/src/Core/SolverBase.h

@@ -0,0 +1,130 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVERBASE_H
+#define EIGEN_SOLVERBASE_H
+
+namespace Eigen {
+
+namespace internal {
+
+
+
+} // end namespace internal
+
+/** \class SolverBase
+  * \brief A base class for matrix decomposition and solvers
+  *
+  * \tparam Derived the actual type of the decomposition/solver.
+  *
+  * Any matrix decomposition inheriting this base class provide the following API:
+  *
+  * \code
+  * MatrixType A, b, x;
+  * DecompositionType dec(A);
+  * x = dec.solve(b);             // solve A   * x = b
+  * x = dec.transpose().solve(b); // solve A^T * x = b
+  * x = dec.adjoint().solve(b);   // solve A'  * x = b
+  * \endcode
+  *
+  * \warning Currently, any other usage of transpose() and adjoint() are not supported and will produce compilation errors.
+  *
+  * \sa class PartialPivLU, class FullPivLU
+  */
+template<typename Derived>
+class SolverBase : public EigenBase<Derived>
+{
+  public:
+
+    typedef EigenBase<Derived> Base;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef Scalar CoeffReturnType;
+
+    enum {
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+                                                          internal::traits<Derived>::ColsAtCompileTime>::ret),
+      MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime,
+      MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime,
+                                                             internal::traits<Derived>::MaxColsAtCompileTime>::ret),
+      IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1
+                           || internal::traits<Derived>::MaxColsAtCompileTime == 1
+    };
+
+    /** Default constructor */
+    SolverBase()
+    {}
+
+    ~SolverBase()
+    {}
+
+    using Base::derived;
+
+    /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A.
+      */
+    template<typename Rhs>
+    inline const Solve<Derived, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<Derived, Rhs>(derived(), b.derived());
+    }
+
+    /** \internal the return type of transpose() */
+    typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    /** \returns an expression of the transposed of the factored matrix.
+      *
+      * A typical usage is to solve for the transposed problem A^T x = b:
+      * \code x = dec.transpose().solve(b); \endcode
+      *
+      * \sa adjoint(), solve()
+      */
+    inline ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(derived());
+    }
+
+    /** \internal the return type of adjoint() */
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                        CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>,
+                        ConstTransposeReturnType
+                     >::type AdjointReturnType;
+    /** \returns an expression of the adjoint of the factored matrix
+      *
+      * A typical usage is to solve for the adjoint problem A' x = b:
+      * \code x = dec.adjoint().solve(b); \endcode
+      *
+      * For real scalar types, this function is equivalent to transpose().
+      *
+      * \sa transpose(), solve()
+      */
+    inline AdjointReturnType adjoint() const
+    {
+      return AdjointReturnType(derived().transpose());
+    }
+
+  protected:
+};
+
+namespace internal {
+
+template<typename Derived>
+struct generic_xpr_base<Derived, MatrixXpr, SolverStorage>
+{
+  typedef SolverBase<Derived> type;
+
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVERBASE_H

Eigen/src/Core/SpecialFunctions.h

@@ -0,0 +1,160 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPECIAL_FUNCTIONS_H
+#define EIGEN_SPECIAL_FUNCTIONS_H
+
+namespace Eigen {
+namespace internal {
+
+/****************************************************************************
+ * Implementation of lgamma                                                 *
+ ****************************************************************************/
+
+template<typename Scalar>
+struct lgamma_impl
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar&)
+  {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template<typename Scalar>
+struct lgamma_retval
+{
+  typedef Scalar type;
+};
+
+#ifdef EIGEN_HAS_C99_MATH
+template<>
+struct lgamma_impl<float>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(const float& x) { return ::lgammaf(x); }
+};
+
+template<>
+struct lgamma_impl<double>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(const double& x) { return ::lgamma(x); }
+};
+#endif
+
+/****************************************************************************
+ * Implementation of erf                                                    *
+ ****************************************************************************/
+
+template<typename Scalar>
+struct erf_impl
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar&)
+  {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template<typename Scalar>
+struct erf_retval
+{
+  typedef Scalar type;
+};
+
+#ifdef EIGEN_HAS_C99_MATH
+template<>
+struct erf_impl<float>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(const float& x) { return ::erff(x); }
+};
+
+template<>
+struct erf_impl<double>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(const double& x) { return ::erf(x); }
+};
+#endif  // EIGEN_HAS_C99_MATH
+
+/***************************************************************************
+* Implementation of erfc                                                   *
+****************************************************************************/
+
+template<typename Scalar>
+struct erfc_impl
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Scalar&)
+  {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar, Scalar>::value == false),
+                        THIS_TYPE_IS_NOT_SUPPORTED);
+    return Scalar(0);
+  }
+};
+
+template<typename Scalar>
+struct erfc_retval
+{
+  typedef Scalar type;
+};
+
+#ifdef EIGEN_HAS_C99_MATH
+template<>
+struct erfc_impl<float>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE float run(const float x) { return ::erfcf(x); }
+};
+
+template<>
+struct erfc_impl<double>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE double run(const double x) { return ::erfc(x); }
+};
+#endif  // EIGEN_HAS_C99_MATH
+
+}  // end namespace internal
+
+
+namespace numext {
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(lgamma, Scalar) lgamma(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(lgamma, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(erf, Scalar) erf(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(erf, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(erfc, Scalar) erfc(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(erfc, Scalar)::run(x);
+}
+
+}  // end namespace numext
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_SPECIAL_FUNCTIONS_H

Eigen/src/Core/Transpose.h

@@ -39,7 +39,7 @@ struct traits<Transpose<MatrixType> > : public traits<MatrixType>
     MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime,
     MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
     FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
-    Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit),
+    Flags0 = traits<MatrixTypeNestedPlain>::Flags & ~(LvalueBit | NestByRefBit),
     Flags1 = Flags0 | FlagsLvalueBit,
     Flags = Flags1 ^ RowMajorBit,
     InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret,

Eigen/src/Core/TriangularMatrix.h

@@ -222,18 +222,23 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     TriangularView& operator=(const TriangularView &other)
     { return Base::operator=(other); }
 
+    /** \copydoc EigenBase::rows() */
     EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows(); }
+    /** \copydoc EigenBase::cols() */
     EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_matrix.cols(); }
 
+    /** \returns a const reference to the nested expression */
     EIGEN_DEVICE_FUNC
     const NestedExpression& nestedExpression() const { return m_matrix; }
+
+    /** \returns a reference to the nested expression */
     EIGEN_DEVICE_FUNC
     NestedExpression& nestedExpression() { return *const_cast<NestedExpression*>(&m_matrix); }
     
-    /** \sa MatrixBase::conjugate() const */
     typedef TriangularView<const MatrixConjugateReturnType,Mode> ConjugateReturnType;
+    /** \sa MatrixBase::conjugate() const */
     EIGEN_DEVICE_FUNC
     inline const ConjugateReturnType conjugate() const
     { return ConjugateReturnType(m_matrix.conjugate()); }
@@ -279,19 +284,28 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     using Base::solve;
   #endif
 
-    EIGEN_DEVICE_FUNC
-    const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const
-    {
-      EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
-      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
-    }
+    /** \returns a selfadjoint view of the referenced triangular part which must be either \c #Upper or \c #Lower.
+      *
+      * This is a shortcut for \code this->nestedExpression().selfadjointView<(*this)::Mode>() \endcode
+      * \sa MatrixBase::selfadjointView() */
     EIGEN_DEVICE_FUNC
     SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView()
     {
-      EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
+      EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR);
       return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
     }
 
+    /** This is the const version of selfadjointView() */
+    EIGEN_DEVICE_FUNC
+    const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const
+    {
+      EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR);
+      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
+    }
+
+
+    /** \returns the determinant of the triangular matrix
+      * \sa MatrixBase::determinant() */
     EIGEN_DEVICE_FUNC
     Scalar determinant() const
     {
@@ -341,12 +355,16 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
       Flags = internal::traits<TriangularViewType>::Flags
     };
 
+    /** \returns the outer-stride of the underlying dense matrix
+      * \sa DenseCoeffsBase::outerStride() */
     EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
+    /** \returns the inner-stride of the underlying dense matrix
+      * \sa DenseCoeffsBase::innerStride() */
     EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
 
-    /** \sa MatrixBase::operator+=() */    
+    /** \sa MatrixBase::operator+=() */
     template<typename Other>
     EIGEN_DEVICE_FUNC
     TriangularViewType&  operator+=(const DenseBase<Other>& other) {
@@ -364,7 +382,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
     /** \sa MatrixBase::operator*=() */
     EIGEN_DEVICE_FUNC
     TriangularViewType&  operator*=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() * other; }
-    /** \sa MatrixBase::operator/=() */
+    /** \sa DenseBase::operator/=() */
     EIGEN_DEVICE_FUNC
     TriangularViewType&  operator/=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() / other; }
 
@@ -408,21 +426,26 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
     EIGEN_DEVICE_FUNC
     TriangularViewType& operator=(const TriangularBase<OtherDerived>& other);
 
+    /** Shortcut for\code *this = other.other.triangularView<(*this)::Mode>() \endcode */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     TriangularViewType& operator=(const MatrixBase<OtherDerived>& other);
 
+#ifndef EIGEN_PARSED_BY_DOXYGEN
     EIGEN_DEVICE_FUNC
     TriangularViewType& operator=(const TriangularViewImpl& other)
     { return *this = other.derived().nestedExpression(); }
 
+    /** \deprecated */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     void lazyAssign(const TriangularBase<OtherDerived>& other);
 
+    /** \deprecated */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    void lazyAssign(const MatrixBase<OtherDerived>& other);  
+    void lazyAssign(const MatrixBase<OtherDerived>& other);
+#endif
 
     /** Efficient triangular matrix times vector/matrix product */
     template<typename OtherDerived>
@@ -442,11 +465,39 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
       return Product<OtherDerived,TriangularViewType>(lhs.derived(),rhs.derived());
     }
 
+    /** \returns the product of the inverse of \c *this with \a other, \a *this being triangular.
+      *
+      * This function computes the inverse-matrix matrix product inverse(\c *this) * \a other if
+      * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
+      * \a Side==OnTheRight.
+      *
+      * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
+      * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
+      * is an upper (resp. lower) triangular matrix.
+      *
+      * Example: \include Triangular_solve.cpp
+      * Output: \verbinclude Triangular_solve.out
+      *
+      * This function returns an expression of the inverse-multiply and can works in-place if it is assigned
+      * to the same matrix or vector \a other.
+      *
+      * For users coming from BLAS, this function (and more specifically solveInPlace()) offer
+      * all the operations supported by the \c *TRSV and \c *TRSM BLAS routines.
+      *
+      * \sa TriangularView::solveInPlace()
+      */
     template<int Side, typename Other>
     EIGEN_DEVICE_FUNC
     inline const internal::triangular_solve_retval<Side,TriangularViewType, Other>
     solve(const MatrixBase<Other>& other) const;
 
+    /** "in-place" version of TriangularView::solve() where the result is written in \a other
+      *
+      * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
+      * This function will const_cast it, so constness isn't honored here.
+      *
+      * See TriangularView:solve() for the details.
+      */
     template<int Side, typename OtherDerived>
     EIGEN_DEVICE_FUNC
     void solveInPlace(const MatrixBase<OtherDerived>& other) const;
@@ -456,18 +507,26 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
     void solveInPlace(const MatrixBase<OtherDerived>& other) const
     { return solveInPlace<OnTheLeft>(other); }
 
+    /** Swaps the coefficients of the common triangular parts of two matrices */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+    void swap(TriangularBase<OtherDerived> &other)
+#else
     void swap(TriangularBase<OtherDerived> const & other)
+#endif
     {
+      EIGEN_STATIC_ASSERT_LVALUE(OtherDerived);
       call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
     }
 
-    // TODO: this overload is ambiguous and it should be deprecated (Gael)
+    /** \deprecated
+      * Shortcut for \code (*this).swap(other.triangularView<(*this)::Mode>()) \endcode */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     void swap(MatrixBase<OtherDerived> const & other)
     {
+      EIGEN_STATIC_ASSERT_LVALUE(OtherDerived);
       call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
     }
 
@@ -503,7 +562,7 @@ template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
 void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const MatrixBase<OtherDerived>& other)
 {
-  internal::call_assignment(derived().noalias(), other.template triangularView<Mode>());
+  internal::call_assignment_no_alias(derived(), other.template triangularView<Mode>());
 }
 
 
@@ -523,7 +582,7 @@ template<typename OtherDerived>
 void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBase<OtherDerived>& other)
 {
   eigen_assert(Mode == int(OtherDerived::Mode));
-  internal::call_assignment(derived().noalias(), other.derived());
+  internal::call_assignment_no_alias(derived(), other.derived());
 }
 
 /***************************************************************************
@@ -745,7 +804,7 @@ EIGEN_DEVICE_FUNC void call_triangular_assignment_loop(const DstXprType& dst, co
   
   enum {
       unroll = DstXprType::SizeAtCompileTime != Dynamic
-            && SrcEvaluatorType::CoeffReadCost != Dynamic
+            && SrcEvaluatorType::CoeffReadCost < HugeCost
             && DstXprType::SizeAtCompileTime * SrcEvaluatorType::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT
     };
   

Eigen/src/Core/VectorwiseOp.h

@@ -11,7 +11,7 @@
 #ifndef EIGEN_PARTIAL_REDUX_H
 #define EIGEN_PARTIAL_REDUX_H
 
-namespace Eigen { 
+namespace Eigen {
 
 /** \class PartialReduxExpr
   * \ingroup Core_Module
@@ -41,8 +41,6 @@ struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> >
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
   typedef typename MatrixType::Scalar InputScalar;
-  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
-  typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::RowsAtCompileTime,
     ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime,
@@ -62,8 +60,6 @@ class PartialReduxExpr : public internal::dense_xpr_base< PartialReduxExpr<Matri
 
     typedef typename internal::dense_xpr_base<PartialReduxExpr>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(PartialReduxExpr)
-    typedef typename internal::traits<PartialReduxExpr>::MatrixTypeNested MatrixTypeNested;
-    typedef typename internal::traits<PartialReduxExpr>::_MatrixTypeNested _MatrixTypeNested;
 
     EIGEN_DEVICE_FUNC
     explicit PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
@@ -74,24 +70,14 @@ class PartialReduxExpr : public internal::dense_xpr_base< PartialReduxExpr<Matri
     EIGEN_DEVICE_FUNC
     Index cols() const { return (Direction==Horizontal ? 1 : m_matrix.cols()); }
 
-    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
-    {
-      if (Direction==Vertical)
-        return m_functor(m_matrix.col(j));
-      else
-        return m_functor(m_matrix.row(i));
-    }
+    EIGEN_DEVICE_FUNC
+    typename MatrixType::Nested nestedExpression() const { return m_matrix; }
 
-    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      if (Direction==Vertical)
-        return m_functor(m_matrix.col(index));
-      else
-        return m_functor(m_matrix.row(index));
-    }
+    EIGEN_DEVICE_FUNC
+    const MemberOp& functor() const { return m_functor; }
 
   protected:
-    MatrixTypeNested m_matrix;
+    typename MatrixType::Nested m_matrix;
     const MemberOp m_functor;
 };
 
@@ -124,6 +110,16 @@ EIGEN_MEMBER_FUNCTOR(any, (Size-1)*NumTraits<Scalar>::AddCost);
 EIGEN_MEMBER_FUNCTOR(count, (Size-1)*NumTraits<Scalar>::AddCost);
 EIGEN_MEMBER_FUNCTOR(prod, (Size-1)*NumTraits<Scalar>::MulCost);
 
+template <int p, typename ResultType>
+struct member_lpnorm {
+  typedef ResultType result_type;
+  template<typename Scalar, int Size> struct Cost
+  { enum { value = (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost }; };
+  EIGEN_DEVICE_FUNC member_lpnorm() {}
+  template<typename XprType>
+  EIGEN_DEVICE_FUNC inline ResultType operator()(const XprType& mat) const
+  { return mat.template lpNorm<p>(); }
+};
 
 template <typename BinaryOp, typename Scalar>
 struct member_redux {
@@ -230,7 +226,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
                        isVertical   ? 1 : m_matrix.rows(),
                        isHorizontal ? 1 : m_matrix.cols());
     }
-    
+
     template<typename OtherDerived> struct OppositeExtendedType {
       typedef Replicate<OtherDerived,
                         isHorizontal ? 1 : ExpressionType::RowsAtCompileTime,
@@ -290,9 +286,13 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     typedef typename ReturnType<internal::member_prod>::Type ProdReturnType;
     typedef Reverse<ExpressionType, Direction> ReverseReturnType;
 
+    template<int p> struct LpNormReturnType {
+      typedef PartialReduxExpr<ExpressionType, internal::member_lpnorm<p,RealScalar>,Direction> Type;
+    };
+
     /** \returns a row (or column) vector expression of the smallest coefficient
       * of each column (or row) of the referenced expression.
-      * 
+      *
       * \warning the result is undefined if \c *this contains NaN.
       *
       * Example: \include PartialRedux_minCoeff.cpp
@@ -305,7 +305,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** \returns a row (or column) vector expression of the largest coefficient
       * of each column (or row) of the referenced expression.
-      * 
+      *
       * \warning the result is undefined if \c *this contains NaN.
       *
       * Example: \include PartialRedux_maxCoeff.cpp
@@ -340,10 +340,23 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     const NormReturnType norm() const
     { return NormReturnType(_expression()); }
 
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * Example: \include PartialRedux_norm.cpp
+      * Output: \verbinclude PartialRedux_norm.out
+      *
+      * \sa DenseBase::norm() */
+    template<int p>
+    EIGEN_DEVICE_FUNC
+    const typename LpNormReturnType<p>::Type lpNorm() const
+    { return typename LpNormReturnType<p>::Type(_expression()); }
+
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression, using
-      * Blue's algorithm. 
+      * Blue's algorithm.
       * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::blueNorm() */
@@ -408,7 +421,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * \sa DenseBase::any() */
     EIGEN_DEVICE_FUNC
     const AnyReturnType any() const
-    { return Any(_expression()); }
+    { return AnyReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * the number of \c true coefficients of each respective column (or row).
@@ -579,7 +592,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
       return m_matrix / extendedTo(other.derived());
     }
-    
+
     /** \returns an expression where each column of row of the referenced matrix are normalized.
       * The referenced matrix is \b not modified.
       * \sa MatrixBase::normalized(), normalize()
@@ -589,8 +602,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
                   const ExpressionTypeNestedCleaned,
                   const typename OppositeExtendedType<typename ReturnType<internal::member_norm,RealScalar>::Type>::Type>
     normalized() const { return m_matrix.cwiseQuotient(extendedToOpposite(this->norm())); }
-    
-    
+
+
     /** Normalize in-place each row or columns of the referenced matrix.
       * \sa MatrixBase::normalize(), normalized()
       */

Eigen/src/Core/Visitor.h

@@ -109,14 +109,11 @@ void DenseBase<Derived>::visit(Visitor& visitor) const
   typedef typename internal::visitor_evaluator<Derived> ThisEvaluator;
   ThisEvaluator thisEval(derived());
   
-  enum { unroll =   SizeAtCompileTime != Dynamic
-                &&  ThisEvaluator::CoeffReadCost != Dynamic
-                &&  (SizeAtCompileTime == 1 || internal::functor_traits<Visitor>::Cost != Dynamic)
-                &&  SizeAtCompileTime * ThisEvaluator::CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost
-                <= EIGEN_UNROLLING_LIMIT };
-  return internal::visitor_impl<Visitor, ThisEvaluator,
-      unroll ? int(SizeAtCompileTime) : Dynamic
-    >::run(thisEval, visitor);
+  enum {
+    unroll =  SizeAtCompileTime != Dynamic
+           && SizeAtCompileTime * ThisEvaluator::CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost <= EIGEN_UNROLLING_LIMIT
+  };
+  return internal::visitor_impl<Visitor, ThisEvaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(thisEval, visitor);
 }
 
 namespace internal {

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

@@ -38,10 +38,10 @@ psin<Packet8f>(const Packet8f& _x) {
   _EIGEN_DECLARE_CONST_Packet8f(two, 2.0f);
   _EIGEN_DECLARE_CONST_Packet8f(one_over_four, 0.25f);
   _EIGEN_DECLARE_CONST_Packet8f(one_over_pi, 3.183098861837907e-01f);
-  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_first, -3.140625000000000e+00);
-  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_second, -9.670257568359375e-04);
-  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_third, -6.278329571784980e-07);
-  _EIGEN_DECLARE_CONST_Packet8f(four_over_pi, 1.273239544735163e+00);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_first, -3.140625000000000e+00f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_second, -9.670257568359375e-04f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_third, -6.278329571784980e-07f);
+  _EIGEN_DECLARE_CONST_Packet8f(four_over_pi, 1.273239544735163e+00f);
 
   // Map x from [-Pi/4,3*Pi/4] to z in [-1,3] and subtract the shifted period.
   Packet8f z = pmul(x, p8f_one_over_pi);
@@ -55,14 +55,14 @@ psin<Packet8f>(const Packet8f& _x) {
   // is odd.
   Packet8i shift_ints = _mm256_cvtps_epi32(shift);
   Packet8i shift_isodd =
-      (__m256i)_mm256_and_ps((__m256)shift_ints, (__m256)p8i_one);
+      _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(shift_ints), _mm256_castsi256_ps(p8i_one)));
 #ifdef EIGEN_VECTORIZE_AVX2
   Packet8i sign_flip_mask = _mm256_slli_epi32(shift_isodd, 31);
 #else
   __m128i lo =
-      _mm_slli_epi32(_mm256_extractf128_si256((__m256i)shift_isodd, 0), 31);
+      _mm_slli_epi32(_mm256_extractf128_si256(shift_isodd, 0), 31);
   __m128i hi =
-      _mm_slli_epi32(_mm256_extractf128_si256((__m256i)shift_isodd, 1), 31);
+      _mm_slli_epi32(_mm256_extractf128_si256(shift_isodd, 1), 31);
   Packet8i sign_flip_mask = _mm256_setr_m128(lo, hi);
 #endif
 
@@ -72,9 +72,9 @@ psin<Packet8f>(const Packet8f& _x) {
 
   // Evaluate the polynomial for the interval [1,3] in z.
   _EIGEN_DECLARE_CONST_Packet8f(coeff_right_0, 9.999999724233232e-01f);
-  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_2, -3.084242535619928e-01);
-  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_4, 1.584991525700324e-02);
-  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_6, -3.188805084631342e-04);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_2, -3.084242535619928e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_4, 1.584991525700324e-02f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_6, -3.188805084631342e-04f);
   Packet8f z_minus_two = psub(z, p8f_two);
   Packet8f z_minus_two2 = pmul(z_minus_two, z_minus_two);
   Packet8f right = pmadd(p8f_coeff_right_6, z_minus_two2, p8f_coeff_right_4);
@@ -82,10 +82,10 @@ psin<Packet8f>(const Packet8f& _x) {
   right = pmadd(right, z_minus_two2, p8f_coeff_right_0);
 
   // Evaluate the polynomial for the interval [-1,1] in z.
-  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_1, 7.853981525427295e-01);
-  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_3, -8.074536727092352e-02);
-  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_5, 2.489871967827018e-03);
-  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_7, -3.587725841214251e-05);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_1, 7.853981525427295e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_3, -8.074536727092352e-02f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_5, 2.489871967827018e-03f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_7, -3.587725841214251e-05f);
   Packet8f z2 = pmul(z, z);
   Packet8f left = pmadd(p8f_coeff_left_7, z2, p8f_coeff_left_5);
   left = pmadd(left, z2, p8f_coeff_left_3);
@@ -98,7 +98,7 @@ psin<Packet8f>(const Packet8f& _x) {
   Packet8f res = _mm256_or_ps(left, right);
 
   // Flip the sign on the odd intervals and return the result.
-  res = _mm256_xor_ps(res, (__m256)sign_flip_mask);
+  res = _mm256_xor_ps(res, _mm256_castsi256_ps(sign_flip_mask));
   return res;
 }
 
@@ -145,10 +145,10 @@ plog<Packet8f>(const Packet8f& _x) {
 // Extract the shifted exponents (No bitwise shifting in regular AVX, so
 // convert to SSE and do it there).
 #ifdef EIGEN_VECTORIZE_AVX2
-  Packet8f emm0 = _mm256_cvtepi32_ps(_mm256_srli_epi32((__m256i)x, 23));
+  Packet8f emm0 = _mm256_cvtepi32_ps(_mm256_srli_epi32(_mm256_castps_si256(x), 23));
 #else
-  __m128i lo = _mm_srli_epi32(_mm256_extractf128_si256((__m256i)x, 0), 23);
-  __m128i hi = _mm_srli_epi32(_mm256_extractf128_si256((__m256i)x, 1), 23);
+  __m128i lo = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(x), 0), 23);
+  __m128i hi = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(x), 1), 23);
   Packet8f emm0 = _mm256_cvtepi32_ps(_mm256_setr_m128(lo, hi));
 #endif
   Packet8f e = _mm256_sub_ps(emm0, p8f_126f);
@@ -348,7 +348,7 @@ pexp<Packet4d>(const Packet4d& _x) {
 
   // Construct the result 2^n * exp(g) = e * x. The max is used to catch
   // non-finite values in the input.
-  return pmax(pmul(x, Packet4d(e)), _x);
+  return pmax(pmul(x, _mm256_castsi256_pd(e)), _x);
 }
 
 // Functions for sqrt.
@@ -393,7 +393,7 @@ Packet4d psqrt<Packet4d>(const Packet4d& x) {
 
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet8f prsqrt<Packet8f>(const Packet8f& _x) {
- _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000);
   _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(nan, 0x7fc00000);
   _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
   _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);

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

@@ -43,7 +43,7 @@ template<> struct is_arithmetic<__m256d> { enum { value = true }; };
   const Packet4d p4d_##NAME = pset1<Packet4d>(X)
 
 #define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \
-  const Packet8f p8f_##NAME = (__m256)pset1<Packet8i>(X)
+  const Packet8f p8f_##NAME = _mm256_castsi256_ps(pset1<Packet8i>(X))
 
 #define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \
   const Packet8i p8i_##NAME = pset1<Packet8i>(X)
@@ -66,7 +66,10 @@ template<> struct packet_traits<float>  : default_packet_traits
     HasExp  = 1,
     HasSqrt = 1,
     HasRsqrt = 1,
-    HasBlend = 1
+    HasBlend = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
   };
 };
 template<> struct packet_traits<double> : default_packet_traits
@@ -83,7 +86,10 @@ template<> struct packet_traits<double> : default_packet_traits
     HasExp  = 1,
     HasSqrt = 1,
     HasRsqrt = 1,
-    HasBlend = 1
+    HasBlend = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
   };
 };
 
@@ -176,6 +182,15 @@ template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const
 template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_max_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_max_pd(a,b); }
 
+template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); }
+template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pceil<Packet8f>(const Packet8f& a) { return _mm256_ceil_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { return _mm256_ceil_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pfloor<Packet8f>(const Packet8f& a) { return _mm256_floor_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet4d pfloor<Packet4d>(const Packet4d& a) { return _mm256_floor_pd(a); }
+
 template<> EIGEN_STRONG_INLINE Packet8f pand<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_and_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet4d pand<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_and_pd(a,b); }
 

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

@@ -66,6 +66,43 @@ double2 prsqrt<double2>(const double2& a)
   return make_double2(rsqrt(a.x), rsqrt(a.y));
 }
 
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plgamma<float4>(const float4& a)
+{
+  return make_float4(lgammaf(a.x), lgammaf(a.y), lgammaf(a.z), lgammaf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plgamma<double2>(const double2& a)
+{
+  return make_double2(lgamma(a.x), lgamma(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 perf<float4>(const float4& a)
+{
+  return make_float4(erf(a.x), erf(a.y), erf(a.z), erf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 perf<double2>(const double2& a)
+{
+  return make_double2(erf(a.x), erf(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 perfc<float4>(const float4& a)
+{
+  return make_float4(erfc(a.x), erfc(a.y), erfc(a.z), erfc(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 perfc<double2>(const double2& a)
+{
+  return make_double2(erfc(a.x), erfc(a.y));
+}
+
+
 #endif
 
 } // end namespace internal

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

@@ -39,6 +39,9 @@ template<> struct packet_traits<float> : default_packet_traits
     HasExp  = 1,
     HasSqrt = 1,
     HasRsqrt = 1,
+    HasLGamma = 1,
+    HasErf = 1,
+    HasErfc = 1,
 
     HasBlend = 0,
   };
@@ -59,6 +62,9 @@ template<> struct packet_traits<double> : default_packet_traits
     HasExp  = 1,
     HasSqrt = 1,
     HasRsqrt = 1,
+    HasLGamma = 1,
+    HasErf = 1,
+    HasErfc = 1,
 
     HasBlend = 0,
   };
@@ -177,7 +183,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<double>(double* to
   to[1] = from.y;
 }
 
-#ifdef __CUDA_ARCH__
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
 template<>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const float* from) {
   return __ldg((const float4*)from);
@@ -244,6 +250,13 @@ template<> EIGEN_DEVICE_FUNC inline double predux_min<double2>(const double2& a)
   return fmin(a.x, a.y);
 }
 
+template<> EIGEN_DEVICE_FUNC inline float  predux_mul<float4>(const float4& a) {
+  return a.x * a.y * a.z * a.w;
+}
+template<> EIGEN_DEVICE_FUNC inline double predux_mul<double2>(const double2& a) {
+  return a.x * a.y;
+}
+
 template<> EIGEN_DEVICE_FUNC inline float4  pabs<float4>(const float4& a) {
   return make_float4(fabsf(a.x), fabsf(a.y), fabsf(a.z), fabsf(a.w));
 }

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

@@ -73,7 +73,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
 
   // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
   v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0));
-  // Get the real values of a | a1_im | a1_im | a2_im | a2_im |
+  // Get the imag values of a | a1_im | a1_im | a2_im | a2_im |
   v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1));
   // Multiply the real a with b
   v1 = vmulq_f32(v1, b.v);
@@ -325,8 +325,8 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con
 
   // Get the real values of a 
   v1 = vdupq_lane_f64(vget_low_f64(a.v), 0);
-  // Get the real values of a 
-  v2 = vdupq_lane_f64(vget_high_f64(a.v), 1);
+  // Get the imag values of a
+  v2 = vdupq_lane_f64(vget_high_f64(a.v), 0);
   // Multiply the real a with b
   v1 = vmulq_f64(v1, b.v);
   // Multiply the imag a with b

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

@@ -67,7 +67,6 @@ template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
 
 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
-  // TODO optimize it for SSE3 and 4
   #ifdef EIGEN_VECTORIZE_SSE3
   return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
                                  _mm_mul_ps(_mm_movehdup_ps(a.v),
@@ -310,9 +309,8 @@ template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
 
 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
-  // TODO optimize it for SSE3 and 4
   #ifdef EIGEN_VECTORIZE_SSE3
-  return Packet1cd(_mm_addsub_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
+  return Packet1cd(_mm_addsub_pd(_mm_mul_pd(_mm_movedup_pd(a.v), b.v),
                                  _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
                                             vec2d_swizzle1(b.v, 1, 0))));
   #else

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

@@ -110,6 +110,13 @@ template<> struct packet_traits<float>  : default_packet_traits
     HasSqrt = 1,
     HasRsqrt = 1,
     HasBlend = 1
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+    ,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+#endif
   };
 };
 template<> struct packet_traits<double> : default_packet_traits
@@ -127,6 +134,13 @@ template<> struct packet_traits<double> : default_packet_traits
     HasSqrt = 1,
     HasRsqrt = 1,
     HasBlend = 1
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+    ,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+#endif
   };
 };
 #endif
@@ -135,7 +149,6 @@ template<> struct packet_traits<int>    : default_packet_traits
   typedef Packet4i type;
   typedef Packet4i half;
   enum {
-    // FIXME check the Has*
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size=4,
@@ -223,10 +236,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const
 
 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
-{ eigen_assert(false && "packet integer division are not supported by SSE");
-  return pset1<Packet4i>(0);
-}
 
 // for some weird raisons, it has to be overloaded for packet of integers
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
@@ -261,6 +270,17 @@ template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const
 #endif
 }
 
+#ifdef EIGEN_VECTORIZE_SSE4_1
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); }
+#endif
+
 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); }
 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); }
@@ -287,8 +307,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { E
     #if (EIGEN_COMP_MSVC==1600)
     // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps
     // (i.e., it does not generate an unaligned load!!
-    // TODO On most architectures this version should also be faster than a single _mm_loadu_ps
-    // so we could also enable it for MSVC08 but first we have to make this later does not generate crap when doing so...
     __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from));
     res = _mm_loadh_pi(res, (const __m64*)(from+2));
     return res;
@@ -299,24 +317,16 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { E
   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 __m128i*>(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 EIGEN_COMP_GNUC && (EIGEN_ARCH_i386 || (EIGEN_ARCH_x86_64 && EIGEN_GNUC_AT_LEAST(4, 8)))
   // bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd
   #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
-  #define EIGEN_AVOID_CUSTOM_UNALIGNED_STORES 1
 #elif EIGEN_COMP_CLANG
   // bug 201: Segfaults in __mm_loadh_pd with clang 2.8
   #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
-  #define EIGEN_AVOID_CUSTOM_UNALIGNED_STORES 0
 #else
   #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0
-  #define EIGEN_AVOID_CUSTOM_UNALIGNED_STORES 0
 #endif
 
 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
@@ -374,17 +384,9 @@ template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& f
 template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
 template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
 
-template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) {
-  EIGEN_DEBUG_UNALIGNED_STORE
-#if EIGEN_AVOID_CUSTOM_UNALIGNED_STORES
-  _mm_storeu_pd(to, from);
-#else
-  _mm_storel_pd((to), from);
-  _mm_storeh_pd((to+1), from);
-#endif
-}
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), Packet2d(_mm_castps_pd(from))); }
-template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), Packet2d(_mm_castsi128_pd(from))); }
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); }
 
 template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
 {
@@ -547,7 +549,6 @@ EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
 }
 
 #ifdef EIGEN_VECTORIZE_SSE3
-// TODO implement SSE2 versions as well as integer versions
 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
 {
   return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
@@ -556,11 +557,6 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
 {
   return _mm_hadd_pd(vecs[0], vecs[1]);
 }
-// SSSE3 version:
-// EIGEN_STRONG_INLINE Packet4i preduxp(const Packet4i* vecs)
-// {
-//   return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
-// }
 
 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
 {
@@ -569,23 +565,16 @@ template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
 }
 
 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return pfirst<Packet2d>(_mm_hadd_pd(a, a)); }
-
-// SSSE3 version:
-// EIGEN_STRONG_INLINE float predux(const Packet4i& a)
-// {
-//   Packet4i tmp0 = _mm_hadd_epi32(a,a);
-//   return pfirst(_mm_hadd_epi32(tmp0, tmp0));
-// }
 #else
 // SSE2 versions
 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
 {
   Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
-  return pfirst(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+  return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
 }
 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
 {
-  return pfirst(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
+  return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
 }
 
 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
@@ -608,6 +597,18 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
 }
 #endif  // SSE3
 
+
+#ifdef EIGEN_VECTORIZE_SSSE3
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
+}
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i tmp0 = _mm_hadd_epi32(a,a);
+  return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0));
+}
+#else
 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
 {
   Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
@@ -627,7 +628,7 @@ template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
   tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
   return _mm_add_epi32(tmp0, tmp2);
 }
-
+#endif
 // Other reduction functions:
 
 // mul

Eigen/src/Core/functors/BinaryFunctors.h

@@ -26,10 +26,10 @@ template<typename Scalar> struct scalar_sum_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
   { return internal::padd(a,b); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
   { return internal::predux(a); }
 };
 template<typename Scalar>
@@ -65,10 +65,10 @@ template<typename LhsScalar,typename RhsScalar> struct scalar_product_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
   { return internal::pmul(a,b); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
   { return internal::predux_mul(a); }
 };
 template<typename LhsScalar,typename RhsScalar>
@@ -97,7 +97,7 @@ template<typename LhsScalar,typename RhsScalar> struct scalar_conj_product_op {
   { return conj_helper<LhsScalar,RhsScalar,Conj,false>().pmul(a,b); }
   
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
   { return conj_helper<Packet,Packet,Conj,false>().pmul(a,b); }
 };
 template<typename LhsScalar,typename RhsScalar>
@@ -117,10 +117,10 @@ template<typename Scalar> struct scalar_min_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return numext::mini(a, b); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
   { return internal::pmin(a,b); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
   { return internal::predux_min(a); }
 };
 template<typename Scalar>
@@ -140,10 +140,10 @@ template<typename Scalar> struct scalar_max_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return numext::maxi(a, b); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
   { return internal::pmax(a,b); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
   { return internal::predux_max(a); }
 };
 template<typename Scalar>
@@ -175,22 +175,37 @@ struct result_of<scalar_cmp_op<Scalar, Cmp>(Scalar,Scalar)> {
 
 
 template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_EQ> {
+  typedef bool result_type;
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
   EIGEN_DEVICE_FUNC 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> {
+  typedef bool result_type;
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
   EIGEN_DEVICE_FUNC 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> {
+  typedef bool result_type;
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a<=b;}
 };
+template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_GT> {
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a>b;}
+};
+template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_GE> {
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC 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> {
+  typedef bool result_type;
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
   EIGEN_DEVICE_FUNC 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> {
+  typedef bool result_type;
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a!=b;}
 };
@@ -252,7 +267,7 @@ template<typename Scalar> struct scalar_difference_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a - b; }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
   { return internal::psub(a,b); }
 };
 template<typename Scalar>
@@ -277,7 +292,7 @@ template<typename LhsScalar,typename RhsScalar> struct scalar_quotient_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
   { return internal::pdiv(a,b); }
 };
 template<typename LhsScalar,typename RhsScalar>
@@ -349,7 +364,7 @@ struct scalar_multiple_op {
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a * m_other; }
   template <typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
   { return internal::pmul(a, pset1<Packet>(m_other)); }
   typename add_const_on_value_type<typename NumTraits<Scalar>::Nested>::type m_other;
 };
@@ -384,7 +399,7 @@ struct scalar_quotient1_op {
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {}
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; }
   template <typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
   { return internal::pdiv(a, pset1<Packet>(m_other)); }
   typename add_const_on_value_type<typename NumTraits<Scalar>::Nested>::type m_other;
 };
@@ -426,7 +441,7 @@ struct scalar_add_op {
   EIGEN_DEVICE_FUNC inline scalar_add_op(const Scalar& other) : m_other(other) { }
   EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a + m_other; }
   template <typename Packet>
-  inline const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
   { return internal::padd(a, pset1<Packet>(m_other)); }
   const Scalar m_other;
 };
@@ -440,11 +455,11 @@ struct functor_traits<scalar_add_op<Scalar> >
   */
 template<typename Scalar>
 struct scalar_sub_op {
-  inline scalar_sub_op(const scalar_sub_op& other) : m_other(other.m_other) { }
-  inline scalar_sub_op(const Scalar& other) : m_other(other) { }
-  inline Scalar operator() (const Scalar& a) const { return a - m_other; }
+  EIGEN_DEVICE_FUNC inline scalar_sub_op(const scalar_sub_op& other) : m_other(other.m_other) { }
+  EIGEN_DEVICE_FUNC inline scalar_sub_op(const Scalar& other) : m_other(other) { }
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a - m_other; }
   template <typename Packet>
-  inline const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
   { return internal::psub(a, pset1<Packet>(m_other)); }
   const Scalar m_other;
 };
@@ -458,11 +473,11 @@ struct functor_traits<scalar_sub_op<Scalar> >
   */
 template<typename Scalar>
 struct scalar_rsub_op {
-  inline scalar_rsub_op(const scalar_rsub_op& other) : m_other(other.m_other) { }
-  inline scalar_rsub_op(const Scalar& other) : m_other(other) { }
-  inline Scalar operator() (const Scalar& a) const { return m_other - a; }
+  EIGEN_DEVICE_FUNC inline scalar_rsub_op(const scalar_rsub_op& other) : m_other(other.m_other) { }
+  EIGEN_DEVICE_FUNC inline scalar_rsub_op(const Scalar& other) : m_other(other) { }
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return m_other - a; }
   template <typename Packet>
-  inline const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
   { return internal::psub(pset1<Packet>(m_other), a); }
   const Scalar m_other;
 };
@@ -477,8 +492,8 @@ struct functor_traits<scalar_rsub_op<Scalar> >
 template<typename Scalar>
 struct scalar_pow_op {
   // FIXME default copy constructors seems bugged with std::complex<>
-  inline scalar_pow_op(const scalar_pow_op& other) : m_exponent(other.m_exponent) { }
-  inline scalar_pow_op(const Scalar& exponent) : m_exponent(exponent) {}
+  EIGEN_DEVICE_FUNC inline scalar_pow_op(const scalar_pow_op& other) : m_exponent(other.m_exponent) { }
+  EIGEN_DEVICE_FUNC inline scalar_pow_op(const Scalar& exponent) : m_exponent(exponent) {}
   EIGEN_DEVICE_FUNC
   inline Scalar operator() (const Scalar& a) const { return numext::pow(a, m_exponent); }
   const Scalar m_exponent;
@@ -493,10 +508,10 @@ struct functor_traits<scalar_pow_op<Scalar> >
   */
 template<typename Scalar>
 struct scalar_inverse_mult_op {
-  scalar_inverse_mult_op(const Scalar& other) : m_other(other) {}
+  EIGEN_DEVICE_FUNC scalar_inverse_mult_op(const Scalar& other) : m_other(other) {}
   EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return m_other / a; }
   template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
   { return internal::pdiv(pset1<Packet>(m_other),a); }
   Scalar m_other;
 };

Eigen/src/Core/functors/NullaryFunctors.h

@@ -21,12 +21,11 @@ struct scalar_constant_op {
   template<typename Index>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; }
   template<typename Index, typename PacketType>
-  EIGEN_STRONG_INLINE const PacketType packetOp(Index, Index = 0) const { return internal::pset1<PacketType>(m_other); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packetOp(Index, Index = 0) const { return internal::pset1<PacketType>(m_other); }
   const Scalar m_other;
 };
 template<typename Scalar>
 struct functor_traits<scalar_constant_op<Scalar> >
-// FIXME replace this packet test by a safe one
 { enum { Cost = 1, PacketAccess = packet_traits<Scalar>::Vectorizable, IsRepeatable = true }; };
 
 template<typename Scalar> struct scalar_identity_op {
@@ -64,7 +63,7 @@ struct linspaced_op_impl<Scalar,Packet,false>
   }
 
   template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); }
 
   const Scalar m_low;
   const Scalar m_step;
@@ -86,7 +85,7 @@ struct linspaced_op_impl<Scalar,Packet,true>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; }
 
   template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index i) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index i) const
   { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1<Packet>(Scalar(i)),m_interPacket))); }
 
   const Scalar m_low;
@@ -121,12 +120,12 @@ template <typename Scalar, typename PacketType, bool RandomAccess> struct linspa
   }
 
   template<typename Index, typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index i) const { return impl.packetOp(i); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index i) const { return impl.packetOp(i); }
 
   // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since
   // there row==0 and col is used for the actual iteration.
   template<typename Index, typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index row, Index col) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index row, Index col) const
   {
     eigen_assert(col==0 || row==0);
     return impl.packetOp(col + row);
@@ -135,14 +134,12 @@ template <typename Scalar, typename PacketType, bool RandomAccess> struct linspa
   // This proxy object handles the actual required temporaries, the different
   // implementations (random vs. sequential access) as well as the
   // correct piping to size 2/4 packet operations.
-  // TODO find a way to make the packet type configurable
   const linspaced_op_impl<Scalar,PacketType,RandomAccess> impl;
 };
 
 // all functors allow linear access, except scalar_identity_op. So we fix here a quick meta
 // to indicate whether a functor allows linear access, just always answering 'yes' except for
 // scalar_identity_op.
-// FIXME move this to functor_traits adding a functor_default
 template<typename Functor> struct functor_has_linear_access { enum { ret = 1 }; };
 template<typename Scalar> struct functor_has_linear_access<scalar_identity_op<Scalar> > { enum { ret = 0 }; };
 

Eigen/src/Core/functors/UnaryFunctors.h

@@ -23,7 +23,7 @@ template<typename Scalar> struct scalar_opposite_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
   { return internal::pnegate(a); }
 };
 template<typename Scalar>
@@ -43,7 +43,7 @@ template<typename Scalar> struct scalar_abs_op {
   typedef typename NumTraits<Scalar>::Real result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using std::abs; return abs(a); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
   { return internal::pabs(a); }
 };
 template<typename Scalar>
@@ -94,7 +94,7 @@ template<typename Scalar> struct scalar_abs2_op {
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
   { return internal::pmul(a,a); }
 };
 template<typename Scalar>
@@ -111,7 +111,7 @@ template<typename Scalar> struct scalar_conjugate_op {
   EIGEN_DEVICE_FUNC
   EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_conjugate_op<Scalar> >
@@ -132,7 +132,7 @@ template<typename Scalar> struct scalar_arg_op {
   typedef typename NumTraits<Scalar>::Real result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using numext::arg; return arg(a); }
   template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
   { return internal::parg(a); }
 };
 template<typename Scalar>
@@ -232,7 +232,7 @@ template<typename Scalar> struct scalar_exp_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::exp; return exp(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_exp_op<Scalar> >
@@ -248,7 +248,7 @@ template<typename Scalar> struct scalar_log_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::log; return log(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::plog(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_log_op<Scalar> >
@@ -264,7 +264,7 @@ template<typename Scalar> struct scalar_log10_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_log10_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::log10; return log10(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::plog10(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog10(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_log10_op<Scalar> >
@@ -278,7 +278,7 @@ template<typename Scalar> struct scalar_sqrt_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return sqrt(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_sqrt_op<Scalar> >
@@ -296,7 +296,7 @@ template<typename Scalar> struct scalar_rsqrt_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_rsqrt_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return Scalar(1)/sqrt(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::prsqrt(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::prsqrt(a); }
 };
 
 template<typename Scalar>
@@ -315,7 +315,7 @@ template<typename Scalar> struct scalar_cos_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op)
   EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::pcos(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcos(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_cos_op<Scalar> >
@@ -334,7 +334,7 @@ template<typename Scalar> struct scalar_sin_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::psin(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psin(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_sin_op<Scalar> >
@@ -354,7 +354,7 @@ template<typename Scalar> struct scalar_tan_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::ptan(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::ptan(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_tan_op<Scalar> >
@@ -373,7 +373,7 @@ template<typename Scalar> struct scalar_acos_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::pacos(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pacos(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_acos_op<Scalar> >
@@ -392,7 +392,7 @@ template<typename Scalar> struct scalar_asin_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op)
   EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::pasin(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pasin(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_asin_op<Scalar> >
@@ -403,15 +403,86 @@ struct functor_traits<scalar_asin_op<Scalar> >
   };
 };
 
+
+/** \internal
+ * \brief Template functor to compute the natural log of the absolute
+ * value of Gamma of a scalar
+ * \sa class CwiseUnaryOp, Cwise::lgamma()
+ */
+template<typename Scalar> struct scalar_lgamma_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_lgamma_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const {
+    using numext::lgamma; return lgamma(a);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  inline Packet packetOp(const Packet& a) const { return internal::plgamma(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_lgamma_op<Scalar> >
+{
+  enum {
+    // Guesstimate
+    Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasLGamma
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Gauss error function of a
+ * scalar
+ * \sa class CwiseUnaryOp, Cwise::erf()
+ */
+template<typename Scalar> struct scalar_erf_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_erf_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const {
+    using numext::erf; return erf(a);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  inline Packet packetOp(const Packet& a) const { return internal::perf(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_erf_op<Scalar> >
+{
+  enum {
+    // Guesstimate
+    Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasErf
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the Complementary Error Function
+ * of a scalar
+ * \sa class CwiseUnaryOp, Cwise::erfc()
+ */
+template<typename Scalar> struct scalar_erfc_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_erfc_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const {
+    using numext::erfc; return erfc(a);
+  }
+  typedef typename packet_traits<Scalar>::type Packet;
+  inline Packet packetOp(const Packet& a) const { return internal::perfc(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_erfc_op<Scalar> >
+{
+  enum {
+    // Guesstimate
+    Cost = 10 * NumTraits<Scalar>::MulCost + 5 * NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasErfc
+  };
+};
+
+
 /** \internal
   * \brief Template functor to compute the atan of a scalar
   * \sa class CwiseUnaryOp, ArrayBase::atan()
   */
 template<typename Scalar> struct scalar_atan_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_atan_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::atan; return atan(a); }
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::atan; return atan(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::patan(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::patan(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_atan_op<Scalar> >
@@ -422,15 +493,16 @@ struct functor_traits<scalar_atan_op<Scalar> >
   };
 };
 
+
 /** \internal
   * \brief Template functor to compute the tanh of a scalar
   * \sa class CwiseUnaryOp, ArrayBase::tanh()
   */
 template<typename Scalar> struct scalar_tanh_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_tanh_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::tanh; return tanh(a); }
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::tanh; return tanh(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::ptanh(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::ptanh(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_tanh_op<Scalar> >
@@ -447,9 +519,9 @@ struct functor_traits<scalar_tanh_op<Scalar> >
   */
 template<typename Scalar> struct scalar_sinh_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_sinh_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::sinh; return sinh(a); }
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sinh; return sinh(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::psinh(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psinh(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_sinh_op<Scalar> >
@@ -466,9 +538,9 @@ struct functor_traits<scalar_sinh_op<Scalar> >
   */
 template<typename Scalar> struct scalar_cosh_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cosh_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::cosh; return cosh(a); }
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::cosh; return cosh(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::pcosh(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcosh(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_cosh_op<Scalar> >
@@ -488,7 +560,7 @@ struct scalar_inverse_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op)
   EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; }
   template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
   { return internal::pdiv(pset1<Packet>(Scalar(1)),a); }
 };
 template<typename Scalar>
@@ -504,7 +576,7 @@ struct scalar_square_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op)
   EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a; }
   template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
   { return internal::pmul(a,a); }
 };
 template<typename Scalar>
@@ -520,7 +592,7 @@ struct scalar_cube_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op)
   EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a*a; }
   template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
   { return internal::pmul(a,pmul(a,a)); }
 };
 template<typename Scalar>
@@ -535,7 +607,7 @@ template<typename Scalar> struct scalar_round_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_round_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::round(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::pround(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pround(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_round_op<Scalar> >
@@ -554,7 +626,7 @@ template<typename Scalar> struct scalar_floor_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_floor_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::floor(a); }
   template <typename Packet>
-  inline Packet packetOp(const Packet& a) const { return internal::pfloor(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pfloor(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_floor_op<Scalar> >
@@ -573,7 +645,7 @@ template<typename Scalar> struct scalar_ceil_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_ceil_op)
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::ceil(a); }
   typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pceil(a); }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pceil(a); }
 };
 template<typename Scalar>
 struct functor_traits<scalar_ceil_op<Scalar> >
@@ -655,6 +727,49 @@ struct functor_traits<scalar_boolean_not_op<Scalar> > {
   };
 };
 
+/** \internal
+  * \brief Template functor to compute the signum of a scalar
+  * \sa class CwiseUnaryOp, Cwise::sign()
+  */
+template<typename Scalar,bool iscpx=(NumTraits<Scalar>::IsComplex!=0) > struct scalar_sign_op;
+template<typename Scalar> 
+struct scalar_sign_op<Scalar,false> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const 
+  {
+      return Scalar( (a>Scalar(0)) - (a<Scalar(0)) );
+  }
+  //TODO
+  //template <typename Packet>
+  //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); }
+};
+template<typename Scalar> 
+struct scalar_sign_op<Scalar,true> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const 
+  {
+    using std::abs;
+    typedef typename NumTraits<Scalar>::Real real_type;
+    real_type aa = abs(a);
+    if (aa==0)
+      return Scalar(0); 
+    aa = 1./aa; 
+    return Scalar(real(a)*aa, imag(a)*aa );
+  }
+  //TODO
+  //template <typename Packet>
+  //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_sign_op<Scalar> >
+{ enum {
+    Cost = 
+        NumTraits<Scalar>::IsComplex
+        ? ( 8*NumTraits<Scalar>::MulCost  ) // roughly
+        : ( 3*NumTraits<Scalar>::AddCost),
+    PacketAccess = packet_traits<Scalar>::HasSign
+  };
+};
 
 } // end namespace internal
 

Eigen/src/Core/products/GeneralBlockPanelKernel.h

@@ -36,37 +36,40 @@ const std::ptrdiff_t defaultL3CacheSize = 512*1024;
 #endif
 
 /** \internal */
-inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff_t* l2, std::ptrdiff_t* l3)
-{
-  static bool m_cache_sizes_initialized = false;
-  static std::ptrdiff_t m_l1CacheSize = 0;
-  static std::ptrdiff_t m_l2CacheSize = 0;
-  static std::ptrdiff_t m_l3CacheSize = 0;
-
-  if(!m_cache_sizes_initialized)
-  {
+struct CacheSizes { 
+  CacheSizes(): m_l1(-1),m_l2(-1),m_l3(-1) {
     int l1CacheSize, l2CacheSize, l3CacheSize;
     queryCacheSizes(l1CacheSize, l2CacheSize, l3CacheSize);
-    m_l1CacheSize = manage_caching_sizes_helper(l1CacheSize, defaultL1CacheSize);
-    m_l2CacheSize = manage_caching_sizes_helper(l2CacheSize, defaultL2CacheSize);
-    m_l3CacheSize = manage_caching_sizes_helper(l3CacheSize, defaultL3CacheSize);
-    m_cache_sizes_initialized = true;
+    m_l1 = manage_caching_sizes_helper(l1CacheSize, defaultL1CacheSize);
+    m_l2 = manage_caching_sizes_helper(l2CacheSize, defaultL2CacheSize);
+    m_l3 = manage_caching_sizes_helper(l3CacheSize, defaultL3CacheSize);
   }
 
+  std::ptrdiff_t m_l1;
+  std::ptrdiff_t m_l2;
+  std::ptrdiff_t m_l3;
+};
+
+
+/** \internal */
+inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff_t* l2, std::ptrdiff_t* l3)
+{
+  static CacheSizes m_cacheSizes;
+
   if(action==SetAction)
   {
     // set the cpu cache size and cache all block sizes from a global cache size in byte
     eigen_internal_assert(l1!=0 && l2!=0);
-    m_l1CacheSize = *l1;
-    m_l2CacheSize = *l2;
-    m_l3CacheSize = *l3;
+    m_cacheSizes.m_l1 = *l1;
+    m_cacheSizes.m_l2 = *l2;
+    m_cacheSizes.m_l3 = *l3;
   }
   else if(action==GetAction)
   {
     eigen_internal_assert(l1!=0 && l2!=0);
-    *l1 = m_l1CacheSize;
-    *l2 = m_l2CacheSize;
-    *l3 = m_l3CacheSize;
+    *l1 = m_cacheSizes.m_l1;
+    *l2 = m_cacheSizes.m_l2;
+    *l3 = m_cacheSizes.m_l3;
   }
   else
   {
@@ -200,8 +203,6 @@ void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index n
     const Index actual_l2 = 1572864; // == 1.5 MB
     #endif
     
-    
-    
     // Here, nc is chosen such that a block of kc x nc of the rhs fit within half of L2.
     // The second half is implicitly reserved to access the result and lhs coefficients.
     // When k<max_kc, then nc can arbitrarily growth. In practice, it seems to be fruitful

Eigen/src/Core/products/GeneralMatrixMatrix.h

@@ -149,7 +149,7 @@ static void run(Index rows, Index cols, Index depth,
       {
       for(Index i=0; i<threads; ++i)
         #pragma omp atomic
-        --(info[i].users);
+        info[i].users -= 1;
       }
     }
   }

Eigen/src/Core/products/Parallelizer.h

@@ -102,21 +102,17 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
   // - we are not already in a parallel code
   // - the sizes are large enough
 
-  // 1- are we already in a parallel session?
-  // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
-  if((!Condition) || (omp_get_num_threads()>1))
-    return func(0,rows, 0,cols);
-
-  Index size = transpose ? rows : cols;
-
-  // 2- compute the maximal number of threads from the size of the product:
+  // compute the maximal number of threads from the size of the product:
   // FIXME this has to be fine tuned
-  Index max_threads = std::max<Index>(1,size / 32);
+  Index size = transpose ? rows : cols;
+  Index pb_max_threads = std::max<Index>(1,size / 32);
+  // compute the number of threads we are going to use
+  Index threads = std::min<Index>(nbThreads(), pb_max_threads);
 
-  // 3 - compute the number of threads we are going to use
-  Index threads = std::min<Index>(nbThreads(), max_threads);
-
-  if(threads==1)
+  // if multi-threading is explicitely disabled, not useful, or if we already are in a parallel session,
+  // then abort multi-threading
+  // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
+  if((!Condition) || (threads==1) || (omp_get_num_threads()>1))
     return func(0,rows, 0,cols);
 
   Eigen::initParallel();

Eigen/src/Core/products/SelfadjointMatrixVector.h

@@ -30,7 +30,7 @@ struct selfadjoint_matrix_vector_product
 static EIGEN_DONT_INLINE void run(
   Index size,
   const Scalar*  lhs, Index lhsStride,
-  const Scalar* _rhs, Index rhsIncr,
+  const Scalar*  rhs,
   Scalar* res,
   Scalar alpha);
 };
@@ -39,11 +39,12 @@ template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool Conju
 EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run(
   Index size,
   const Scalar*  lhs, Index lhsStride,
-  const Scalar* _rhs, Index rhsIncr,
+  const Scalar*  rhs,
   Scalar* res,
   Scalar alpha)
 {
   typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
   const Index PacketSize = sizeof(Packet)/sizeof(Scalar);
 
   enum {
@@ -54,23 +55,13 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
 
   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> cj0;
   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> cj1;
-  conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex, ConjugateRhs> cjd;
+  conj_helper<RealScalar,Scalar,false, ConjugateRhs> cjd;
 
   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> pcj0;
   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> pcj1;
 
   Scalar cjAlpha = ConjugateRhs ? numext::conj(alpha) : alpha;
 
-  // FIXME this copy is now handled outside product_selfadjoint_vector, so it could probably be removed.
-  // if the rhs is not sequentially stored in memory we copy it to a temporary buffer,
-  // this is because we need to extract packets
-  ei_declare_aligned_stack_constructed_variable(Scalar,rhs,size,rhsIncr==1 ? const_cast<Scalar*>(_rhs) : 0);  
-  if (rhsIncr!=1)
-  {
-    const Scalar* it = _rhs;
-    for (Index i=0; i<size; ++i, it+=rhsIncr)
-      rhs[i] = *it;
-  }
 
   Index bound = (std::max)(Index(0),size-8) & 0xfffffffe;
   if (FirstTriangular)
@@ -97,7 +88,6 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
     size_t alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti);
     size_t alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
 
-    // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
     res[j]   += cjd.pmul(numext::real(A0[j]), t0);
     res[j+1] += cjd.pmul(numext::real(A1[j+1]), t1);
     if(FirstTriangular)
@@ -151,7 +141,6 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
 
     Scalar t1 = cjAlpha * rhs[j];
     Scalar t2(0);
-    // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
     res[j] += cjd.pmul(numext::real(A0[j]), t1);
     for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++)
     {
@@ -238,7 +227,7 @@ struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true>
       (
         lhs.rows(),                             // size
         &lhs.coeffRef(0,0),  lhs.outerStride(), // lhs info
-        actualRhsPtr, 1,                        // rhs info
+        actualRhsPtr,                           // rhs info
         actualDestPtr,                          // result info
         actualAlpha                             // scale factor
       );

Eigen/src/Core/products/SelfadjointMatrixVector_MKL.h

@@ -52,16 +52,16 @@ template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool Con
 struct selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Specialized> { \
 static void run( \
   Index size, const Scalar*  lhs, Index lhsStride, \
-  const Scalar* _rhs, Index rhsIncr, Scalar* res, Scalar alpha) { \
+  const Scalar* _rhs, Scalar* res, Scalar alpha) { \
     enum {\
       IsColMajor = StorageOrder==ColMajor \
     }; \
     if (IsColMajor == ConjugateLhs) {\
       selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn>::run( \
-        size, lhs, lhsStride, _rhs, rhsIncr, res, alpha);  \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
     } else {\
       selfadjoint_matrix_vector_product_symv<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs>::run( \
-        size, lhs, lhsStride, _rhs, rhsIncr, res, alpha);  \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
     }\
   } \
 }; \
@@ -79,13 +79,13 @@ typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> SYMVVector;\
 \
 static void run( \
 Index size, const EIGTYPE*  lhs, Index lhsStride, \
-const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* res, EIGTYPE alpha) \
+const EIGTYPE* _rhs, EIGTYPE* res, EIGTYPE alpha) \
 { \
   enum {\
     IsRowMajor = StorageOrder==RowMajor ? 1 : 0, \
     IsLower = UpLo == Lower ? 1 : 0 \
   }; \
-  MKL_INT n=size, lda=lhsStride, incx=rhsIncr, incy=1; \
+  MKL_INT n=size, lda=lhsStride, incx=1, incy=1; \
   MKLTYPE alpha_, beta_; \
   const EIGTYPE *x_ptr, myone(1); \
   char uplo=(IsRowMajor) ? (IsLower ? 'U' : 'L') : (IsLower ? 'L' : 'U'); \
@@ -93,10 +93,9 @@ const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* res, EIGTYPE alpha) \
   assign_scalar_eig2mkl(beta_, myone); \
   SYMVVector x_tmp; \
   if (ConjugateRhs) { \
-    Map<const SYMVVector, 0, InnerStride<> > map_x(_rhs,size,1,InnerStride<>(incx)); \
+    Map<const SYMVVector, 0 > map_x(_rhs,size,1); \
     x_tmp=map_x.conjugate(); \
     x_ptr=x_tmp.data(); \
-    incx=1; \
   } else x_ptr=_rhs; \
   MKLFUNC(&uplo, &n, &alpha_, (const MKLTYPE*)lhs, &lda, (const MKLTYPE*)x_ptr, &incx, &beta_, (MKLTYPE*)res, &incy); \
 }\

Eigen/src/Core/products/TriangularSolverMatrix.h

@@ -304,9 +304,12 @@ 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));
-              for (Index i=0; i<actual_mc; ++i)
-                r[i] *= b;
+              if((Mode & UnitDiag)==0)
+              {
+                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/Constants.h

@@ -30,6 +30,14 @@ const int DynamicIndex = 0xffffff;
   */
 const int Infinity = -1;
 
+/** This value means that the cost to evaluate an expression coefficient is either very expensive or
+  * cannot be known at compile time.
+  *
+  * This value has to be positive to (1) simplify cost computation, and (2) allow to distinguish between a very expensive and very very expensive expressions.
+  * It thus must also be large enough to make sure unrolling won't happen and that sub expressions will be evaluated, but not too large to avoid overflow.
+  */
+const int HugeCost = 10000;
+
 /** \defgroup flags Flags
   * \ingroup Core_Module
   *
@@ -189,8 +197,8 @@ const unsigned int HereditaryBits = RowMajorBit
   */
 
 /** \ingroup enums
-  * Enum containing possible values for the \p Mode parameter of 
-  * MatrixBase::selfadjointView() and MatrixBase::triangularView(). */
+  * Enum containing possible values for the \c Mode or \c UpLo parameter of
+  * MatrixBase::selfadjointView() and MatrixBase::triangularView(), and selfadjoint solvers. */
 enum {
   /** View matrix as a lower triangular matrix. */
   Lower=0x1,                      
@@ -484,6 +492,9 @@ struct Dense {};
 /** The type used to identify a general sparse storage. */
 struct Sparse {};
 
+/** The type used to identify a general solver (foctored) storage. */
+struct SolverStorage {};
+
 /** The type used to identify a permutation storage. */
 struct PermutationStorage {};
 
@@ -498,6 +509,7 @@ struct ArrayXpr {};
 
 // An evaluator must define its shape. By default, it can be one of the following:
 struct DenseShape             { static std::string debugName() { return "DenseShape"; } };
+struct SolverShape            { static std::string debugName() { return "SolverShape"; } };
 struct HomogeneousShape       { static std::string debugName() { return "HomogeneousShape"; } };
 struct DiagonalShape          { static std::string debugName() { return "DiagonalShape"; } };
 struct BandShape              { static std::string debugName() { return "BandShape"; } };
@@ -523,7 +535,9 @@ enum ComparisonName {
   cmp_LT = 1,
   cmp_LE = 2,
   cmp_UNORD = 3,
-  cmp_NEQ = 4
+  cmp_NEQ = 4,
+  cmp_GT = 5,
+  cmp_GE = 6
 };
 } // end namespace internal
 

Eigen/src/Core/util/DisableStupidWarnings.h

@@ -10,6 +10,7 @@
   // 4244 - 'argument' : conversion from 'type1' to 'type2', possible loss of data
   // 4273 - QtAlignedMalloc, inconsistent DLL linkage
   // 4324 - structure was padded due to declspec(align())
+  // 4503 - decorated name length exceeded, name was truncated
   // 4512 - assignment operator could not be generated
   // 4522 - 'class' : multiple assignment operators specified
   // 4700 - uninitialized local variable 'xyz' used
@@ -17,17 +18,19 @@
   #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
     #pragma warning( push )
   #endif
-  #pragma warning( disable : 4100 4101 4127 4181 4211 4244 4273 4324 4512 4522 4700 4717 )
+  #pragma warning( disable : 4100 4101 4127 4181 4211 4244 4273 4324 4503 4512 4522 4700 4717 )
 #elif defined __INTEL_COMPILER
   // 2196 - routine is both "inline" and "noinline" ("noinline" assumed)
   //        ICC 12 generates this warning even without any inline keyword, when defining class methods 'inline' i.e. inside of class body
   //        typedef that may be a reference type.
   // 279  - controlling expression is constant
   //        ICC 12 generates this warning on assert(constant_expression_depending_on_template_params) and frankly this is a legitimate use case.
+  // 1684 - conversion from pointer to same-sized integral type (potential portability problem)
+  // 2259 - non-pointer conversion from "Eigen::Index={ptrdiff_t={long}}" to "int" may lose significant bits
   #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
     #pragma warning push
   #endif
-  #pragma warning disable 2196 279
+  #pragma warning disable 2196 279 1684 2259
 #elif defined __clang__
   // -Wconstant-logical-operand - warning: use of logical && with constant operand; switch to bitwise & or remove constant
   //     this is really a stupid warning as it warns on compile-time expressions involving enums

Eigen/src/Core/util/ForwardDeclarations.h

@@ -132,6 +132,7 @@ template<typename MatrixType> struct CommaInitializer;
 template<typename Derived> class ReturnByValue;
 template<typename ExpressionType> class ArrayWrapper;
 template<typename ExpressionType> class MatrixWrapper;
+template<typename Derived> class SolverBase;
 template<typename XprType> class InnerIterator;
 
 namespace internal {
@@ -160,8 +161,7 @@ template< typename T,
           typename LhsShape = typename evaluator_traits<typename T::Lhs>::Shape,
           typename RhsShape = typename evaluator_traits<typename T::Rhs>::Shape,
           typename LhsScalar = typename traits<typename T::Lhs>::Scalar,
-          typename RhsScalar = typename traits<typename T::Rhs>::Scalar,
-          typename = EnableIf<true> // extra template parameter for SFINAE-based specialization
+          typename RhsScalar = typename traits<typename T::Rhs>::Scalar
         > struct product_evaluator;
 }
 
@@ -209,6 +209,7 @@ template<typename Scalar> struct scalar_random_op;
 template<typename Scalar> struct scalar_add_op;
 template<typename Scalar> struct scalar_constant_op;
 template<typename Scalar> struct scalar_identity_op;
+template<typename Scalar,bool iscpx> struct scalar_sign_op;
 
 template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_product_op;
 template<typename LhsScalar,typename RhsScalar> struct scalar_multiple2_op;
@@ -266,7 +267,6 @@ template<typename Scalar> class Rotation2D;
 template<typename Scalar> class AngleAxis;
 template<typename Scalar,int Dim> class Translation;
 template<typename Scalar,int Dim> class AlignedBox;
-
 template<typename Scalar, int Options = AutoAlign> class Quaternion;
 template<typename Scalar,int Dim,int Mode,int _Options=AutoAlign> class Transform;
 template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class ParametrizedLine;
@@ -274,6 +274,9 @@ template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class Hyperp
 template<typename Scalar> class UniformScaling;
 template<typename MatrixType,int Direction> class Homogeneous;
 
+// Sparse module:
+template<typename Derived> class SparseMatrixBase;
+
 // MatrixFunctions module
 template<typename Derived> struct MatrixExponentialReturnValue;
 template<typename Derived> class MatrixFunctionReturnValue;

Eigen/src/Core/util/Macros.h

@@ -13,7 +13,7 @@
 
 #define EIGEN_WORLD_VERSION 3
 #define EIGEN_MAJOR_VERSION 2
-#define EIGEN_MINOR_VERSION 91
+#define EIGEN_MINOR_VERSION 92
 
 #define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \
                                       (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \
@@ -341,6 +341,13 @@
   #define EIGEN_HAVE_RVALUE_REFERENCES
 #endif
 
+// Does the compiler support C99?
+#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901))       \
+  || (defined(__GNUC__) && defined(_GLIBCXX_USE_C99)) \
+  || (defined(_LIBCPP_VERSION) && !defined(_MSC_VER))
+#define EIGEN_HAS_C99_MATH 1
+#endif
+
 // Does the compiler support result_of?
 #if (__has_feature(cxx_lambdas) || (defined(__cplusplus) && __cplusplus >= 201103L))
 #define EIGEN_HAS_STD_RESULT_OF 1
@@ -353,16 +360,19 @@
 
 // Does the compiler support const expressions?
 #ifdef __CUDACC__
-  // Const expressions are not supported regardless of what host compiler is used 
+// Const expressions are supported provided that c++11 is enabled and we're using nvcc 7.5 or above
+#if defined(__CUDACC_VER__) &&  __CUDACC_VER__ >= 70500 && __cplusplus > 199711L
+  #define EIGEN_HAS_CONSTEXPR 1
+#endif
 #elif (defined(__cplusplus) && __cplusplus >= 201402L) || \
-    EIGEN_GNUC_AT_LEAST(4,9)
+    EIGEN_GNUC_AT_LEAST(4,8)
 #define EIGEN_HAS_CONSTEXPR 1
 #endif
 
 // Does the compiler support C++11 math?
 // Let's be conservative and enable the default C++11 implementation only if we are sure it exists
 #ifndef EIGEN_HAS_CXX11_MATH
-  #if (__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC || EIGEN_COMP_ICC)  \
+  #if (__cplusplus > 201103L) || (__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC || EIGEN_COMP_ICC)  \
       && (EIGEN_ARCH_i386_OR_x86_64) && (EIGEN_OS_GNULINUX || EIGEN_OS_WIN_STRICT || EIGEN_OS_MAC)
     #define EIGEN_HAS_CXX11_MATH 1
   #else
@@ -372,17 +382,30 @@
 
 // Does the compiler support proper C++11 containers?
 #ifndef EIGEN_HAS_CXX11_CONTAINERS
-  #if ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG)) || EIGEN_COMP_MSVC >= 1900
+  #if    (__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \
+      || EIGEN_COMP_MSVC >= 1900
     #define EIGEN_HAS_CXX11_CONTAINERS 1
   #else
     #define EIGEN_HAS_CXX11_CONTAINERS 0
   #endif
 #endif
 
+// Does the compiler support C++11 noexcept?
+#ifndef EIGEN_HAS_CXX11_NOEXCEPT
+  #if    (__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \
+      || EIGEN_COMP_MSVC >= 1900
+    #define EIGEN_HAS_CXX11_NOEXCEPT 1
+  #else
+    #define EIGEN_HAS_CXX11_NOEXCEPT 0
+  #endif
+#endif
+
 /** Allows to disable some optimizations which might affect the accuracy of the result.
   * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them.
   * They currently include:
-  *   - single precision ArrayBase::sin() and ArrayBase::cos() when SSE vectorization is enabled.
+  *   - single precision ArrayBase::sin() and ArrayBase::cos() for SSE and AVX vectorization.
   */
 #ifndef EIGEN_FAST_MATH
 #define EIGEN_FAST_MATH 1
@@ -609,10 +632,14 @@ namespace Eigen {
   // 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
   // certain common platform (compiler+architecture combinations) to avoid these problems.
-  // Only static alignment is really problematic (relies on nonstandard compiler extensions that don't
-  // work everywhere, for example don't work on GCC/ARM), try to keep heap alignment even
-  // when we have to disable static alignment.
-  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64)
+  // Only static alignment is really problematic (relies on nonstandard compiler extensions),
+  // try to keep heap alignment even when we have to disable static alignment.
+  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64)
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+  #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6)
+  // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support.
+  // Not sure which version fixed it, hopefully it doesn't affect 4.7, which is still somewhat in use.
+  // 4.8 and newer seem definitely unaffected.
   #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
   #else
   #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
@@ -747,8 +774,6 @@ namespace Eigen {
 * documentation in a single line.
 **/
 
-// TODO The EIGEN_DENSE_PUBLIC_INTERFACE should not exists anymore
-  
 #define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \
   typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex<float>. */ \
   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex<T>, T were corresponding to RealScalar. */ \
@@ -761,17 +786,17 @@ namespace Eigen {
         Flags = Eigen::internal::traits<Derived>::Flags, \
         SizeAtCompileTime = Base::SizeAtCompileTime, \
         MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \
-        IsVectorAtCompileTime = Base::IsVectorAtCompileTime };
-
-
-#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
-  EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \
-  typedef typename Base::PacketScalar PacketScalar; \
-  enum { MaxRowsAtCompileTime = Eigen::internal::traits<Derived>::MaxRowsAtCompileTime, \
-        MaxColsAtCompileTime = Eigen::internal::traits<Derived>::MaxColsAtCompileTime}; \
+        IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \
   using Base::derived; \
   using Base::const_cast_derived;
 
+
+// FIXME Maybe the EIGEN_DENSE_PUBLIC_INTERFACE could be removed as importing PacketScalar is rarely needed
+#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
+  EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \
+  typedef typename Base::PacketScalar PacketScalar;
+
+
 #define EIGEN_PLAIN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b)
 #define EIGEN_PLAIN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b)
 
@@ -837,4 +862,12 @@ namespace Eigen {
 #  define EIGEN_CATCH(X) else
 #endif
 
+#if EIGEN_HAS_CXX11_NOEXCEPT
+#   define EIGEN_NO_THROW noexcept(true)
+#   define EIGEN_EXCEPTION_SPEC(X) noexcept(false)
+#else
+#   define EIGEN_NO_THROW throw()
+#   define EIGEN_EXCEPTION_SPEC(X) throw(X)
+#endif
+
 #endif // EIGEN_MACROS_H

Eigen/src/Core/util/Memory.h

@@ -732,7 +732,7 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
 
 #if EIGEN_MAX_ALIGN_BYTES!=0
   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
-      void* operator new(size_t size, const std::nothrow_t&) throw() { \
+      void* operator new(size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \
         EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
         EIGEN_CATCH (...) { return 0; } \
       }
@@ -743,20 +743,20 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
       void *operator new[](size_t size) { \
         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
       } \
-      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); } \
+      void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_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. */ \
       static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
       static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \
-      void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
-      void operator delete[](void * memory, void *ptr) throw() { return ::operator delete[](memory,ptr); } \
+      void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \
+      void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \
       /* nothrow-new (returns zero instead of std::bad_alloc) */ \
       EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
-      void operator delete(void *ptr, const std::nothrow_t&) throw() { \
+      void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \
         Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
       } \
       typedef void eigen_aligned_operator_new_marker_type;

Eigen/src/Core/util/Meta.h

@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -11,6 +11,11 @@
 #ifndef EIGEN_META_H
 #define EIGEN_META_H
 
+#if defined(__CUDA_ARCH__)
+#include <cfloat>
+#include <math_constants.h>
+#endif
+
 namespace Eigen {
 
 namespace internal {
@@ -68,6 +73,18 @@ template<> struct is_arithmetic<unsigned int>  { enum { value = true }; };
 template<> struct is_arithmetic<signed long>   { enum { value = true }; };
 template<> struct is_arithmetic<unsigned long> { enum { value = true }; };
 
+template<typename T> struct is_integral        { enum { value = false }; };
+template<> struct is_integral<bool>            { enum { value = true }; };
+template<> struct is_integral<char>            { enum { value = true }; };
+template<> struct is_integral<signed char>     { enum { value = true }; };
+template<> struct is_integral<unsigned char>   { enum { value = true }; };
+template<> struct is_integral<signed short>    { enum { value = true }; };
+template<> struct is_integral<unsigned short>  { enum { value = true }; };
+template<> struct is_integral<signed int>      { enum { value = true }; };
+template<> struct is_integral<unsigned int>    { enum { value = true }; };
+template<> struct is_integral<signed long>     { enum { value = true }; };
+template<> struct is_integral<unsigned long>   { enum { value = true }; };
+
 template <typename T> struct add_const { typedef const T type; };
 template <typename T> struct add_const<T&> { typedef T& type; };
 
@@ -138,16 +155,16 @@ template<> struct numeric_limits<float>
   EIGEN_DEVICE_FUNC
   static float (max)() { return CUDART_MAX_NORMAL_F; }
   EIGEN_DEVICE_FUNC
-  static float (min)() { return __FLT_EPSILON__; }
+  static float (min)() { return FLT_MIN; }
 };
 template<> struct numeric_limits<double>
 {
   EIGEN_DEVICE_FUNC
   static double epsilon() { return __DBL_EPSILON__; }
   EIGEN_DEVICE_FUNC
-  static double (max)() { return CUDART_INF; }
+  static double (max)() { return DBL_MAX; }
   EIGEN_DEVICE_FUNC
-  static double (min)() { return __DBL_EPSILON__; }
+  static double (min)() { return DBL_MIN; }
 };
 template<> struct numeric_limits<int>
 {
@@ -158,6 +175,15 @@ template<> struct numeric_limits<int>
   EIGEN_DEVICE_FUNC
   static int (min)() { return INT_MIN; }
 };
+template<> struct numeric_limits<unsigned int>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned int epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned int (max)() { return UINT_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned int (min)() { return 0; }
+};
 template<> struct numeric_limits<long>
 {
   EIGEN_DEVICE_FUNC
@@ -167,6 +193,15 @@ template<> struct numeric_limits<long>
   EIGEN_DEVICE_FUNC
   static long (min)() { return LONG_MIN; }
 };
+template<> struct numeric_limits<unsigned long>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long (max)() { return ULONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long (min)() { return 0; }
+};
 template<> struct numeric_limits<long long>
 {
   EIGEN_DEVICE_FUNC
@@ -176,6 +211,15 @@ template<> struct numeric_limits<long long>
   EIGEN_DEVICE_FUNC
   static long long (min)() { return LLONG_MIN; }
 };
+template<> struct numeric_limits<unsigned long long>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned long long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long long (max)() { return ULLONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long long (min)() { return 0; }
+};
 
 }
 
@@ -193,7 +237,6 @@ protected:
   EIGEN_DEVICE_FUNC ~noncopyable() {}
 };
 
-
 /** \internal
   * Convenient struct to get the result type of a unary or binary functor.
   *

Eigen/src/Core/util/StaticAssert.h

@@ -93,7 +93,11 @@
         THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH,
         OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG,
         IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY,
-        STORAGE_LAYOUT_DOES_NOT_MATCH
+        STORAGE_LAYOUT_DOES_NOT_MATCH,
+        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE,
+        THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS,
+        MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY,
+        THIS_TYPE_IS_NOT_SUPPORTED
       };
     };
 
@@ -200,5 +204,9 @@
                                             >::value), \
                           YOU_CANNOT_MIX_ARRAYS_AND_MATRICES)
 
+// Check that a cost value is positive, and that is stay within a reasonable range
+// TODO this check could be enabled for internal debugging only
+#define EIGEN_INTERNAL_CHECK_COST_VALUE(C) \
+      EIGEN_STATIC_ASSERT((C)>=0 && (C)<=HugeCost*HugeCost, EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE);
 
 #endif // EIGEN_STATIC_ASSERT_H

Eigen/src/Core/util/XprHelper.h

@@ -233,33 +233,33 @@ template<typename XprType> struct size_of_xpr_at_compile_time
  */
 
 template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_matrix_type;
-template<typename T, typename BaseClassType> struct plain_matrix_type_dense;
+template<typename T, typename BaseClassType, int Flags> struct plain_matrix_type_dense;
 template<typename T> struct plain_matrix_type<T,Dense>
 {
-  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind>::type type;
+  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, traits<T>::Flags>::type type;
 };
 template<typename T> struct plain_matrix_type<T,DiagonalShape>
 {
   typedef typename T::PlainObject type;
 };
 
-template<typename T> struct plain_matrix_type_dense<T,MatrixXpr>
+template<typename T, int Flags> struct plain_matrix_type_dense<T,MatrixXpr,Flags>
 {
   typedef Matrix<typename traits<T>::Scalar,
                 traits<T>::RowsAtCompileTime,
                 traits<T>::ColsAtCompileTime,
-                AutoAlign | (traits<T>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor),
                 traits<T>::MaxRowsAtCompileTime,
                 traits<T>::MaxColsAtCompileTime
           > type;
 };
 
-template<typename T> struct plain_matrix_type_dense<T,ArrayXpr>
+template<typename T, int Flags> struct plain_matrix_type_dense<T,ArrayXpr,Flags>
 {
   typedef Array<typename traits<T>::Scalar,
                 traits<T>::RowsAtCompileTime,
                 traits<T>::ColsAtCompileTime,
-                AutoAlign | (traits<T>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor),
                 traits<T>::MaxRowsAtCompileTime,
                 traits<T>::MaxColsAtCompileTime
           > type;
@@ -303,6 +303,15 @@ struct eval<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>, Dense>
 };
 
 
+/* similar to plain_matrix_type, but using the evaluator's Flags */
+template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_object_eval;
+
+template<typename T>
+struct plain_object_eval<T,Dense>
+{
+  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, evaluator<T>::Flags>::type type;
+};
+
 
 /* plain_matrix_type_column_major : same as plain_matrix_type but guaranteed to be column-major
  */
@@ -385,29 +394,23 @@ struct transfer_constness
   * \param n the number of coefficient accesses in the nested expression for each coefficient access in the bigger expression.
   * \param PlainObject the type of the temporary if needed.
   */
-template<typename T, int n, typename PlainObject = typename eval<T>::type> struct nested_eval
+template<typename T, int n, typename PlainObject = typename plain_object_eval<T>::type> struct nested_eval
 {
   enum {
-    // For the purpose of this test, to keep it reasonably simple, we arbitrarily choose a value of Dynamic values.
-    // the choice of 10000 makes it larger than any practical fixed value and even most dynamic values.
-    // in extreme cases where these assumptions would be wrong, we would still at worst suffer performance issues
-    // (poor choice of temporaries).
-    // It's important that this value can still be squared without integer overflowing.
-    DynamicAsInteger = 10000,
     ScalarReadCost = NumTraits<typename traits<T>::Scalar>::ReadCost,
-    ScalarReadCostAsInteger = ScalarReadCost == Dynamic ? int(DynamicAsInteger) : int(ScalarReadCost),
-    CoeffReadCost = evaluator<T>::CoeffReadCost,  // TODO What if an evaluator evaluate itself into a tempory?
-                                                  // Then CoeffReadCost will be small but we still have to evaluate if n>1... 
-                                                  // The solution might be to ask the evaluator if it creates a temp. Perhaps we could even ask the number of temps?
-    CoeffReadCostAsInteger = CoeffReadCost == Dynamic ? int(DynamicAsInteger) : int(CoeffReadCost),
-    NAsInteger = n == Dynamic ? int(DynamicAsInteger) : n,
-    CostEvalAsInteger   = (NAsInteger+1) * ScalarReadCostAsInteger + CoeffReadCostAsInteger,
-    CostNoEvalAsInteger = NAsInteger * CoeffReadCostAsInteger
+    CoeffReadCost = evaluator<T>::CoeffReadCost,  // NOTE What if an evaluator evaluate itself into a tempory?
+                                                  //      Then CoeffReadCost will be small (e.g., 1) but we still have to evaluate, especially if n>1.
+                                                  //      This situation is already taken care by the EvalBeforeNestingBit flag, which is turned ON
+                                                  //      for all evaluator creating a temporary. This flag is then propagated by the parent evaluators.
+                                                  //      Another solution could be to count the number of temps?
+    NAsInteger = n == Dynamic ? HugeCost : n,
+    CostEval   = (NAsInteger+1) * ScalarReadCost + CoeffReadCost,
+    CostNoEval = NAsInteger * CoeffReadCost
   };
 
   typedef typename conditional<
         ( (int(evaluator<T>::Flags) & EvalBeforeNestingBit) ||
-          (int(CostEvalAsInteger) < int(CostNoEvalAsInteger)) ),
+          (int(CostEval) < int(CostNoEval)) ),
         PlainObject,
         typename ref_selector<T>::type
   >::type type;
@@ -449,9 +452,9 @@ struct generic_xpr_base<Derived, XprKind, Dense>
 
 /** \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
@@ -460,8 +463,8 @@ struct special_scalar_op_base : public DenseCoeffsBase<Derived>
   void operator/(dummy) const;
 };
 
-template<typename Derived,typename Scalar,typename OtherScalar>
-struct special_scalar_op_base<Derived,Scalar,OtherScalar,true>  : public DenseCoeffsBase<Derived>
+template<typename Derived,typename Scalar,typename OtherScalar, typename BaseType>
+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
@@ -654,6 +657,43 @@ bool is_same_dense(const T1 &, const T2 &, typename enable_if<!(has_direct_acces
   return false;
 }
 
+template<typename T, typename U> struct is_same_or_void { enum { value = is_same<T,U>::value }; };
+template<typename T> struct is_same_or_void<void,T>     { enum { value = 1 }; };
+template<typename T> struct is_same_or_void<T,void>     { enum { value = 1 }; };
+template<>           struct is_same_or_void<void,void>  { enum { value = 1 }; };
+
+#ifdef EIGEN_DEBUG_ASSIGN
+std::string demangle_traversal(int t)
+{
+  if(t==DefaultTraversal) return "DefaultTraversal";
+  if(t==LinearTraversal) return "LinearTraversal";
+  if(t==InnerVectorizedTraversal) return "InnerVectorizedTraversal";
+  if(t==LinearVectorizedTraversal) return "LinearVectorizedTraversal";
+  if(t==SliceVectorizedTraversal) return "SliceVectorizedTraversal";
+  return "?";
+}
+std::string demangle_unrolling(int t)
+{
+  if(t==NoUnrolling) return "NoUnrolling";
+  if(t==InnerUnrolling) return "InnerUnrolling";
+  if(t==CompleteUnrolling) return "CompleteUnrolling";
+  return "?";
+}
+std::string demangle_flags(int f)
+{
+  std::string res;
+  if(f&RowMajorBit)                 res += " | RowMajor";
+  if(f&PacketAccessBit)             res += " | Packet";
+  if(f&LinearAccessBit)             res += " | Linear";
+  if(f&LvalueBit)                   res += " | Lvalue";
+  if(f&DirectAccessBit)             res += " | Direct";
+  if(f&NestByRefBit)                res += " | NestByRef";
+  if(f&NoPreferredStorageOrderBit)  res += " | NoPreferredStorageOrderBit";
+  
+  return res;
+}
+#endif
+
 } // end namespace internal
 
 // we require Lhs and Rhs to have the same scalar type. Currently there is no example of a binary functor
@@ -666,7 +706,7 @@ bool is_same_dense(const T1 &, const T2 &, typename enable_if<!(has_direct_acces
 #define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \
   EIGEN_STATIC_ASSERT((internal::functor_is_product_like<BINOP>::ret \
                         ? int(internal::scalar_product_traits<LHS, RHS>::Defined) \
-                        : int(internal::is_same<LHS, RHS>::value)), \
+                        : int(internal::is_same_or_void<LHS, RHS>::value)), \
     YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
     
 } // end namespace Eigen

Eigen/src/Eigenvalues/ComplexEigenSolver.h

@@ -122,7 +122,8 @@ template<typename _MatrixType> class ComplexEigenSolver
       *
       * This constructor calls compute() to compute the eigendecomposition.
       */
-    explicit ComplexEigenSolver(const MatrixType& matrix, bool computeEigenvectors = true)
+    template<typename InputType>
+    explicit ComplexEigenSolver(const EigenBase<InputType>& matrix, bool computeEigenvectors = true)
             : m_eivec(matrix.rows(),matrix.cols()),
               m_eivalues(matrix.cols()),
               m_schur(matrix.rows()),
@@ -130,7 +131,7 @@ template<typename _MatrixType> class ComplexEigenSolver
               m_eigenvectorsOk(false),
               m_matX(matrix.rows(),matrix.cols())
     {
-      compute(matrix, computeEigenvectors);
+      compute(matrix.derived(), computeEigenvectors);
     }
 
     /** \brief Returns the eigenvectors of given matrix.
@@ -208,7 +209,8 @@ template<typename _MatrixType> class ComplexEigenSolver
       * Example: \include ComplexEigenSolver_compute.cpp
       * Output: \verbinclude ComplexEigenSolver_compute.out
       */
-    ComplexEigenSolver& compute(const MatrixType& matrix, bool computeEigenvectors = true);
+    template<typename InputType>
+    ComplexEigenSolver& compute(const EigenBase<InputType>& matrix, bool computeEigenvectors = true);
 
     /** \brief Reports whether previous computation was successful.
       *
@@ -254,8 +256,9 @@ template<typename _MatrixType> class ComplexEigenSolver
 
 
 template<typename MatrixType>
+template<typename InputType>
 ComplexEigenSolver<MatrixType>& 
-ComplexEigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvectors)
+ComplexEigenSolver<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeEigenvectors)
 {
   check_template_parameters();
   
@@ -264,13 +267,13 @@ ComplexEigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEi
 
   // Do a complex Schur decomposition, A = U T U^*
   // The eigenvalues are on the diagonal of T.
-  m_schur.compute(matrix, computeEigenvectors);
+  m_schur.compute(matrix.derived(), computeEigenvectors);
 
   if(m_schur.info() == Success)
   {
     m_eivalues = m_schur.matrixT().diagonal();
     if(computeEigenvectors)
-      doComputeEigenvectors(matrix.norm());
+      doComputeEigenvectors(m_schur.matrixT().norm());
     sortEigenvalues(computeEigenvectors);
   }
 

Eigen/src/Eigenvalues/ComplexSchur.h

@@ -109,7 +109,8 @@ template<typename _MatrixType> class ComplexSchur
       *
       * \sa matrixT() and matrixU() for examples.
       */
-    explicit ComplexSchur(const MatrixType& matrix, bool computeU = true)
+    template<typename InputType>
+    explicit ComplexSchur(const EigenBase<InputType>& matrix, bool computeU = true)
       : m_matT(matrix.rows(),matrix.cols()),
         m_matU(matrix.rows(),matrix.cols()),
         m_hess(matrix.rows()),
@@ -117,7 +118,7 @@ template<typename _MatrixType> class ComplexSchur
         m_matUisUptodate(false),
         m_maxIters(-1)
     {
-      compute(matrix, computeU);
+      compute(matrix.derived(), computeU);
     }
 
     /** \brief Returns the unitary matrix in the Schur decomposition. 
@@ -186,7 +187,8 @@ template<typename _MatrixType> class ComplexSchur
       *
       * \sa compute(const MatrixType&, bool, Index)
       */
-    ComplexSchur& compute(const MatrixType& matrix, bool computeU = true);
+    template<typename InputType>
+    ComplexSchur& compute(const EigenBase<InputType>& matrix, bool computeU = true);
     
     /** \brief Compute Schur decomposition from a given Hessenberg matrix
      *  \param[in] matrixH Matrix in Hessenberg form H
@@ -313,14 +315,15 @@ typename ComplexSchur<MatrixType>::ComplexScalar ComplexSchur<MatrixType>::compu
 
 
 template<typename MatrixType>
-ComplexSchur<MatrixType>& ComplexSchur<MatrixType>::compute(const MatrixType& matrix, bool computeU)
+template<typename InputType>
+ComplexSchur<MatrixType>& ComplexSchur<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeU)
 {
   m_matUisUptodate = false;
   eigen_assert(matrix.cols() == matrix.rows());
 
   if(matrix.cols() == 1)
   {
-    m_matT = matrix.template cast<ComplexScalar>();
+    m_matT = matrix.derived().template cast<ComplexScalar>();
     if(computeU)  m_matU = ComplexMatrixType::Identity(1,1);
     m_info = Success;
     m_isInitialized = true;
@@ -328,7 +331,7 @@ ComplexSchur<MatrixType>& ComplexSchur<MatrixType>::compute(const MatrixType& ma
     return *this;
   }
 
-  internal::complex_schur_reduce_to_hessenberg<MatrixType, NumTraits<Scalar>::IsComplex>::run(*this, matrix, computeU);
+  internal::complex_schur_reduce_to_hessenberg<MatrixType, NumTraits<Scalar>::IsComplex>::run(*this, matrix.derived(), computeU);
   computeFromHessenberg(m_matT, m_matU, computeU);
   return *this;
 }

Eigen/src/Eigenvalues/ComplexSchur_MKL.h

@@ -40,9 +40,9 @@ namespace Eigen {
 /** \internal Specialization for the data types supported by MKL */
 
 #define EIGEN_MKL_SCHUR_COMPLEX(EIGTYPE, MKLTYPE, MKLPREFIX, MKLPREFIX_U, EIGCOLROW, MKLCOLROW) \
-template<> inline \
+template<> template<typename InputType> inline \
 ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
-ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW>& matrix, bool computeU) \
+ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const EigenBase<InputType>& matrix, bool computeU) \
 { \
   typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> MatrixType; \
   typedef MatrixType::RealScalar RealScalar; \
@@ -53,7 +53,7 @@ ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matri
   m_matUisUptodate = false; \
   if(matrix.cols() == 1) \
   { \
-    m_matT = matrix.cast<ComplexScalar>(); \
+    m_matT = matrix.derived().template cast<ComplexScalar>(); \
     if(computeU)  m_matU = ComplexMatrixType::Identity(1,1); \
       m_info = Success; \
       m_isInitialized = true; \
@@ -61,7 +61,6 @@ ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matri
       return *this; \
   } \
   lapack_int n = matrix.cols(), sdim, info; \
-  lapack_int lda = matrix.outerStride(); \
   lapack_int matrix_order = MKLCOLROW; \
   char jobvs, sort='N'; \
   LAPACK_##MKLPREFIX_U##_SELECT1 select = 0; \
@@ -69,6 +68,7 @@ ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matri
   m_matU.resize(n, n); \
   lapack_int ldvs  = m_matU.outerStride(); \
   m_matT = matrix; \
+  lapack_int lda = m_matT.outerStride(); \
   Matrix<EIGTYPE, Dynamic, Dynamic> w; \
   w.resize(n, 1);\
   info = LAPACKE_##MKLPREFIX##gees( matrix_order, jobvs, sort, select, n, (MKLTYPE*)m_matT.data(), lda, &sdim, (MKLTYPE*)w.data(), (MKLTYPE*)m_matU.data(), ldvs ); \

Eigen/src/Eigenvalues/EigenSolver.h

@@ -110,7 +110,7 @@ template<typename _MatrixType> class EigenSolver
       *
       * \sa compute() for an example.
       */
- EigenSolver() : m_eivec(), m_eivalues(), m_isInitialized(false), m_realSchur(), m_matT(), m_tmp() {}
+    EigenSolver() : m_eivec(), m_eivalues(), m_isInitialized(false), m_realSchur(), m_matT(), m_tmp() {}
 
     /** \brief Default constructor with memory preallocation
       *
@@ -143,7 +143,8 @@ template<typename _MatrixType> class EigenSolver
       *
       * \sa compute()
       */
-    explicit EigenSolver(const MatrixType& matrix, bool computeEigenvectors = true)
+    template<typename InputType>
+    explicit EigenSolver(const EigenBase<InputType>& matrix, bool computeEigenvectors = true)
       : m_eivec(matrix.rows(), matrix.cols()),
         m_eivalues(matrix.cols()),
         m_isInitialized(false),
@@ -152,7 +153,7 @@ template<typename _MatrixType> class EigenSolver
         m_matT(matrix.rows(), matrix.cols()), 
         m_tmp(matrix.cols())
     {
-      compute(matrix, computeEigenvectors);
+      compute(matrix.derived(), computeEigenvectors);
     }
 
     /** \brief Returns the eigenvectors of given matrix. 
@@ -273,7 +274,8 @@ template<typename _MatrixType> class EigenSolver
       * Example: \include EigenSolver_compute.cpp
       * Output: \verbinclude EigenSolver_compute.out
       */
-    EigenSolver& compute(const MatrixType& matrix, bool computeEigenvectors = true);
+    template<typename InputType>
+    EigenSolver& compute(const EigenBase<InputType>& matrix, bool computeEigenvectors = true);
 
     /** \returns NumericalIssue if the input contains INF or NaN values or overflow occured. Returns Success otherwise. */
     ComputationInfo info() const
@@ -370,8 +372,9 @@ typename EigenSolver<MatrixType>::EigenvectorsType EigenSolver<MatrixType>::eige
 }
 
 template<typename MatrixType>
+template<typename InputType>
 EigenSolver<MatrixType>& 
-EigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvectors)
+EigenSolver<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeEigenvectors)
 {
   check_template_parameters();
   
@@ -381,7 +384,7 @@ EigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvect
   eigen_assert(matrix.cols() == matrix.rows());
 
   // Reduce to real Schur form.
-  m_realSchur.compute(matrix, computeEigenvectors);
+  m_realSchur.compute(matrix.derived(), computeEigenvectors);
   
   m_info = m_realSchur.info();
 

Eigen/src/Eigenvalues/GeneralizedEigenSolver.h

@@ -145,7 +145,7 @@ template<typename _MatrixType> class GeneralizedEigenSolver
       *
       * \sa compute()
       */
-    explicit GeneralizedEigenSolver(const MatrixType& A, const MatrixType& B, bool computeEigenvectors = true)
+    GeneralizedEigenSolver(const MatrixType& A, const MatrixType& B, bool computeEigenvectors = true)
       : m_eivec(A.rows(), A.cols()),
         m_alphas(A.cols()),
         m_betas(A.cols()),

Eigen/src/Eigenvalues/HessenbergDecomposition.h

@@ -115,8 +115,9 @@ template<typename _MatrixType> class HessenbergDecomposition
       *
       * \sa matrixH() for an example.
       */
-    explicit HessenbergDecomposition(const MatrixType& matrix)
-      : m_matrix(matrix),
+    template<typename InputType>
+    explicit HessenbergDecomposition(const EigenBase<InputType>& matrix)
+      : m_matrix(matrix.derived()),
         m_temp(matrix.rows()),
         m_isInitialized(false)
     {
@@ -147,9 +148,10 @@ template<typename _MatrixType> class HessenbergDecomposition
       * Example: \include HessenbergDecomposition_compute.cpp
       * Output: \verbinclude HessenbergDecomposition_compute.out
       */
-    HessenbergDecomposition& compute(const MatrixType& matrix)
+    template<typename InputType>
+    HessenbergDecomposition& compute(const EigenBase<InputType>& matrix)
     {
-      m_matrix = matrix;
+      m_matrix = matrix.derived();
       if(matrix.rows()<2)
       {
         m_isInitialized = true;

Eigen/src/Eigenvalues/RealQZ.h

@@ -101,7 +101,7 @@ namespace Eigen {
        *
        * This constructor calls compute() to compute the QZ decomposition.
        */
-      explicit RealQZ(const MatrixType& A, const MatrixType& B, bool computeQZ = true) :
+      RealQZ(const MatrixType& A, const MatrixType& B, bool computeQZ = true) :
         m_S(A.rows(),A.cols()),
         m_T(A.rows(),A.cols()),
         m_Q(A.rows(),A.cols()),

Eigen/src/Eigenvalues/RealSchur.h

@@ -100,7 +100,8 @@ template<typename _MatrixType> class RealSchur
       * Example: \include RealSchur_RealSchur_MatrixType.cpp
       * Output: \verbinclude RealSchur_RealSchur_MatrixType.out
       */
-    explicit RealSchur(const MatrixType& matrix, bool computeU = true)
+    template<typename InputType>
+    explicit RealSchur(const EigenBase<InputType>& matrix, bool computeU = true)
             : m_matT(matrix.rows(),matrix.cols()),
               m_matU(matrix.rows(),matrix.cols()),
               m_workspaceVector(matrix.rows()),
@@ -109,7 +110,7 @@ template<typename _MatrixType> class RealSchur
               m_matUisUptodate(false),
               m_maxIters(-1)
     {
-      compute(matrix, computeU);
+      compute(matrix.derived(), computeU);
     }
 
     /** \brief Returns the orthogonal matrix in the Schur decomposition. 
@@ -165,7 +166,8 @@ template<typename _MatrixType> class RealSchur
       *
       * \sa compute(const MatrixType&, bool, Index)
       */
-    RealSchur& compute(const MatrixType& matrix, bool computeU = true);
+    template<typename InputType>
+    RealSchur& compute(const EigenBase<InputType>& matrix, bool computeU = true);
 
     /** \brief Computes Schur decomposition of a Hessenberg matrix H = Z T Z^T
      *  \param[in] matrixH Matrix in Hessenberg form H
@@ -243,7 +245,8 @@ template<typename _MatrixType> class RealSchur
 
 
 template<typename MatrixType>
-RealSchur<MatrixType>& RealSchur<MatrixType>::compute(const MatrixType& matrix, bool computeU)
+template<typename InputType>
+RealSchur<MatrixType>& RealSchur<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeU)
 {
   eigen_assert(matrix.cols() == matrix.rows());
   Index maxIters = m_maxIters;
@@ -251,7 +254,7 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::compute(const MatrixType& matrix,
     maxIters = m_maxIterationsPerRow * matrix.rows();
 
   // Step 1. Reduce to Hessenberg form
-  m_hess.compute(matrix);
+  m_hess.compute(matrix.derived());
 
   // Step 2. Reduce to real Schur form  
   computeFromHessenberg(m_hess.matrixH(), m_hess.matrixQ(), computeU);

Eigen/src/Eigenvalues/RealSchur_MKL.h

@@ -40,14 +40,13 @@ namespace Eigen {
 /** \internal Specialization for the data types supported by MKL */
 
 #define EIGEN_MKL_SCHUR_REAL(EIGTYPE, MKLTYPE, MKLPREFIX, MKLPREFIX_U, EIGCOLROW, MKLCOLROW) \
-template<> inline \
+template<> template<typename InputType> inline \
 RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
-RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW>& matrix, bool computeU) \
+RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const EigenBase<InputType>& matrix, bool computeU) \
 { \
   eigen_assert(matrix.cols() == matrix.rows()); \
 \
   lapack_int n = matrix.cols(), sdim, info; \
-  lapack_int lda = matrix.outerStride(); \
   lapack_int matrix_order = MKLCOLROW; \
   char jobvs, sort='N'; \
   LAPACK_##MKLPREFIX_U##_SELECT2 select = 0; \
@@ -55,6 +54,7 @@ RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matrix<E
   m_matU.resize(n, n); \
   lapack_int ldvs  = m_matU.outerStride(); \
   m_matT = matrix; \
+  lapack_int lda = m_matT.outerStride(); \
   Matrix<EIGTYPE, Dynamic, Dynamic> wr, wi; \
   wr.resize(n, 1); wi.resize(n, 1); \
   info = LAPACKE_##MKLPREFIX##gees( matrix_order, jobvs, sort, select, n, (MKLTYPE*)m_matT.data(), lda, &sdim, (MKLTYPE*)wr.data(), (MKLTYPE*)wi.data(), (MKLTYPE*)m_matU.data(), ldvs ); \

Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h

@@ -157,14 +157,15 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
       *
       * \sa compute(const MatrixType&, int)
       */
+    template<typename InputType>
     EIGEN_DEVICE_FUNC
-    explicit SelfAdjointEigenSolver(const MatrixType& matrix, int options = ComputeEigenvectors)
+    explicit SelfAdjointEigenSolver(const EigenBase<InputType>& matrix, int options = ComputeEigenvectors)
       : m_eivec(matrix.rows(), matrix.cols()),
         m_eivalues(matrix.cols()),
         m_subdiag(matrix.rows() > 1 ? matrix.rows() - 1 : 1),
         m_isInitialized(false)
     {
-      compute(matrix, options);
+      compute(matrix.derived(), options);
     }
 
     /** \brief Computes eigendecomposition of given matrix.
@@ -197,8 +198,9 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
       *
       * \sa SelfAdjointEigenSolver(const MatrixType&, int)
       */
+    template<typename InputType>
     EIGEN_DEVICE_FUNC
-    SelfAdjointEigenSolver& compute(const MatrixType& matrix, int options = ComputeEigenvectors);
+    SelfAdjointEigenSolver& compute(const EigenBase<InputType>& matrix, int options = ComputeEigenvectors);
     
     /** \brief Computes eigendecomposition of given matrix using a closed-form algorithm
       *
@@ -389,12 +391,15 @@ static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index sta
 }
 
 template<typename MatrixType>
+template<typename InputType>
 EIGEN_DEVICE_FUNC
 SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
-::compute(const MatrixType& matrix, int options)
+::compute(const EigenBase<InputType>& a_matrix, int options)
 {
   check_template_parameters();
   
+  const InputType &matrix(a_matrix.derived());
+  
   using std::abs;
   eigen_assert(matrix.cols() == matrix.rows());
   eigen_assert((options&~(EigVecMask|GenEigMask))==0
@@ -406,7 +411,7 @@ SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
 
   if(n==1)
   {
-    m_eivalues.coeffRef(0,0) = numext::real(matrix.coeff(0,0));
+    m_eivalues.coeffRef(0,0) = numext::real(matrix(0,0));
     if(computeEigenvectors)
       m_eivec.setOnes(n,n);
     m_info = Success;

Eigen/src/Eigenvalues/SelfAdjointEigenSolver_MKL.h

@@ -40,9 +40,9 @@ namespace Eigen {
 /** \internal Specialization for the data types supported by MKL */
 
 #define EIGEN_MKL_EIG_SELFADJ(EIGTYPE, MKLTYPE, MKLRTYPE, MKLNAME, EIGCOLROW, MKLCOLROW ) \
-template<> inline \
+template<> template<typename InputType> inline \
 SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
-SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW>& matrix, int options) \
+SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const EigenBase<InputType>& matrix, int options) \
 { \
   eigen_assert(matrix.cols() == matrix.rows()); \
   eigen_assert((options&~(EigVecMask|GenEigMask))==0 \
@@ -56,7 +56,7 @@ SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(c
 \
   if(n==1) \
   { \
-    m_eivalues.coeffRef(0,0) = numext::real(matrix.coeff(0,0)); \
+    m_eivalues.coeffRef(0,0) = numext::real(m_eivec.coeff(0,0)); \
     if(computeEigenvectors) m_eivec.setOnes(n,n); \
     m_info = Success; \
     m_isInitialized = true; \
@@ -64,7 +64,7 @@ SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(c
     return *this; \
   } \
 \
-  lda = matrix.outerStride(); \
+  lda = m_eivec.outerStride(); \
   matrix_order=MKLCOLROW; \
   char jobz, uplo='L'/*, range='A'*/; \
   jobz = computeEigenvectors ? 'V' : 'N'; \

Eigen/src/Eigenvalues/Tridiagonalization.h

@@ -126,8 +126,9 @@ template<typename _MatrixType> class Tridiagonalization
       * Example: \include Tridiagonalization_Tridiagonalization_MatrixType.cpp
       * Output: \verbinclude Tridiagonalization_Tridiagonalization_MatrixType.out
       */
-    explicit Tridiagonalization(const MatrixType& matrix)
-      : m_matrix(matrix),
+    template<typename InputType>
+    explicit Tridiagonalization(const EigenBase<InputType>& matrix)
+      : m_matrix(matrix.derived()),
         m_hCoeffs(matrix.cols() > 1 ? matrix.cols()-1 : 1),
         m_isInitialized(false)
     {
@@ -152,9 +153,10 @@ template<typename _MatrixType> class Tridiagonalization
       * Example: \include Tridiagonalization_compute.cpp
       * Output: \verbinclude Tridiagonalization_compute.out
       */
-    Tridiagonalization& compute(const MatrixType& matrix)
+    template<typename InputType>
+    Tridiagonalization& compute(const EigenBase<InputType>& matrix)
     {
-      m_matrix = matrix;
+      m_matrix = matrix.derived();
       m_hCoeffs.resize(matrix.rows()-1, 1);
       internal::tridiagonalization_inplace(m_matrix, m_hCoeffs);
       m_isInitialized = true;

Eigen/src/Geometry/AlignedBox.h

@@ -163,7 +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)

Eigen/src/Geometry/AngleAxis.h

@@ -85,10 +85,17 @@ 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 */
@@ -133,7 +140,7 @@ 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.
@@ -170,8 +177,8 @@ AngleAxis<Scalar>& AngleAxis<Scalar>::operator=(const QuaternionBase<QuatDerived
   }
   else
   {
-    m_angle = 0;
-    m_axis << 1, 0, 0;
+    m_angle = Scalar(0);
+    m_axis << Scalar(1), Scalar(0), Scalar(0);
   }
   return *this;
 }

Eigen/src/Geometry/Homogeneous.h

@@ -445,6 +445,11 @@ struct generic_product_impl<Transform<Scalar,Dim,Mode,Options>, Homogeneous<RhsA
   }
 };
 
+template<typename ExpressionType, int Side, bool Transposed>
+struct permutation_matrix_product<ExpressionType, Side, Transposed, HomogeneousShape>
+  : public permutation_matrix_product<ExpressionType, Side, Transposed, DenseShape>
+{};
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Geometry/Quaternion.h

@@ -278,6 +278,10 @@ public:
   inline const Coefficients& coeffs() const { return m_coeffs;}
 
   EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsAlignment)
+  
+#ifdef EIGEN_QUATERNION_PLUGIN
+# include EIGEN_QUATERNION_PLUGIN
+#endif
 
 protected:
   Coefficients m_coeffs;
@@ -735,8 +739,9 @@ template<typename Other>
 struct quaternionbase_assign_impl<Other,3,3>
 {
   typedef typename Other::Scalar Scalar;
-  template<class Derived> static inline void run(QuaternionBase<Derived>& q, const Other& mat)
+  template<class Derived> static inline void run(QuaternionBase<Derived>& q, const Other& a_mat)
   {
+    const typename internal::nested_eval<Other,2>::type mat(a_mat);
     using std::sqrt;
     // This algorithm comes from  "Quaternion Calculus and Fast Animation",
     // Ken Shoemake, 1987 SIGGRAPH course notes

Eigen/src/Geometry/Rotation2D.h

@@ -64,6 +64,16 @@ public:
   /** Default constructor wihtout initialization. The represented rotation is undefined. */
   Rotation2D() {}
 
+  /** Construct a 2D rotation from a 2x2 rotation matrix \a mat.
+    *
+    * \sa fromRotationMatrix()
+    */
+  template<typename Derived>
+  explicit Rotation2D(const MatrixBase<Derived>& m)
+  {
+    fromRotationMatrix(m.derived());
+  }
+
   /** \returns the rotation angle */
   inline Scalar angle() const { return m_angle; }
 
@@ -103,6 +113,17 @@ public:
   Rotation2D& fromRotationMatrix(const MatrixBase<Derived>& m);
   Matrix2 toRotationMatrix() const;
 
+  /** Set \c *this from a 2x2 rotation matrix \a mat.
+    * In other words, this function extract the rotation angle from the rotation matrix.
+    *
+    * This method is an alias for fromRotationMatrix()
+    *
+    * \sa fromRotationMatrix()
+    */
+  template<typename Derived>
+  Rotation2D& operator=(const  MatrixBase<Derived>& m)
+  { return fromRotationMatrix(m.derived()); }
+
   /** \returns the spherical interpolation between \c *this and \a other using
     * parameter \a t. It is in fact equivalent to a linear interpolation.
     */

Eigen/src/Geometry/Transform.h

@@ -118,15 +118,15 @@ 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
-  * one before doing the matrix product. The convertions to homogeneous representations
+  * to points, the latters are automatically promoted to homogeneous vectors
+  * before doing the matrix product. The conventions to homogeneous representations
   * are performed as follow:
   *
   * \b Translation t (Dim)x(1):
@@ -140,7 +140,7 @@ 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\\
@@ -152,14 +152,20 @@ 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
@@ -404,26 +410,39 @@ 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:
-    * \li a vector of size Dim,
+    * The right-hand-side \a other can be either:
     * \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 Dynamic,
-    * \li a linear transformation matrix of size Dim x Dim,
-    * \li an affine transformation matrix of size Dim x Dim+1,
+    * \li a set of homogeneous vectors of size Dim+1 x N,
     * \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/IterativeLinearSolvers/BasicPreconditioners.h

@@ -23,6 +23,8 @@ namespace Eigen {
   *
   * \tparam _Scalar the type of the scalar.
   *
+  * \implsparsesolverconcept
+  *
   * This preconditioner is suitable for both selfadjoint and general problems.
   * The diagonal entries are pre-inverted and stored into a dense vector.
   *
@@ -37,8 +39,10 @@ class DiagonalPreconditioner
     typedef Matrix<Scalar,Dynamic,1> Vector;
   public:
     typedef typename Vector::StorageIndex StorageIndex;
-    // this typedef is only to export the scalar type and compile-time dimensions to solve_retval
-    typedef Matrix<Scalar,Dynamic,Dynamic> MatrixType;
+    enum {
+      ColsAtCompileTime = Dynamic,
+      MaxColsAtCompileTime = Dynamic
+    };
 
     DiagonalPreconditioner() : m_isInitialized(false) {}
 
@@ -114,6 +118,8 @@ class DiagonalPreconditioner
   *
   * \tparam _Scalar the type of the scalar.
   *
+  * \implsparsesolverconcept
+  *
   * The diagonal entries are pre-inverted and stored into a dense vector.
   * 
   * \sa class LeastSquaresConjugateGradient, class DiagonalPreconditioner
@@ -172,6 +178,8 @@ class LeastSquareDiagonalPreconditioner : public DiagonalPreconditioner<_Scalar>
 /** \ingroup IterativeLinearSolvers_Module
   * \brief A naive preconditioner which approximates any matrix as the identity matrix
   *
+  * \implsparsesolverconcept
+  *
   * \sa class DiagonalPreconditioner
   */
 class IdentityPreconditioner

Eigen/src/IterativeLinearSolvers/BiCGSTAB.h

@@ -132,6 +132,8 @@ struct traits<BiCGSTAB<_MatrixType,_Preconditioner> >
   * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
   * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
   *
+  * \implsparsesolverconcept
+  *
   * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
   * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
   * and NumTraits<Scalar>::epsilon() for the tolerance.
@@ -148,13 +150,15 @@ struct traits<BiCGSTAB<_MatrixType,_Preconditioner> >
   * By default the iterations start with x=0 as an initial guess of the solution.
   * One can control the start using the solveWithGuess() method.
   * 
+  * BiCGSTAB can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+  *
   * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
   */
 template< typename _MatrixType, typename _Preconditioner>
 class BiCGSTAB : public IterativeSolverBase<BiCGSTAB<_MatrixType,_Preconditioner> >
 {
   typedef IterativeSolverBase<BiCGSTAB> Base;
-  using Base::mp_matrix;
+  using Base::matrix;
   using Base::m_error;
   using Base::m_iterations;
   using Base::m_info;
@@ -180,7 +184,8 @@ public:
     * this class becomes invalid. Call compute() to update it with the new
     * matrix A, or modify a copy of A.
     */
-  explicit BiCGSTAB(const MatrixType& A) : Base(A) {}
+  template<typename MatrixDerived>
+  explicit BiCGSTAB(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
 
   ~BiCGSTAB() {}
 
@@ -195,7 +200,7 @@ public:
       m_error = Base::m_tolerance;
       
       typename Dest::ColXpr xj(x,j);
-      if(!internal::bicgstab(mp_matrix, b.col(j), xj, Base::m_preconditioner, m_iterations, m_error))
+      if(!internal::bicgstab(matrix(), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error))
         failed = true;
     }
     m_info = failed ? NumericalIssue

Eigen/src/IterativeLinearSolvers/ConjugateGradient.h

@@ -118,6 +118,8 @@ struct traits<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
   *               Default is \c Lower, best performance is \c Lower|Upper.
   * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
   *
+  * \implsparsesolverconcept
+  *
   * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
   * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
   * and NumTraits<Scalar>::epsilon() for the tolerance.
@@ -147,13 +149,15 @@ struct traits<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
   * By default the iterations start with x=0 as an initial guess of the solution.
   * One can control the start using the solveWithGuess() method.
   * 
+  * ConjugateGradient can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+  *
   * \sa class LeastSquaresConjugateGradient, class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
   */
 template< typename _MatrixType, int _UpLo, typename _Preconditioner>
 class ConjugateGradient : public IterativeSolverBase<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
 {
   typedef IterativeSolverBase<ConjugateGradient> Base;
-  using Base::mp_matrix;
+  using Base::matrix;
   using Base::m_error;
   using Base::m_iterations;
   using Base::m_info;
@@ -183,7 +187,8 @@ public:
     * this class becomes invalid. Call compute() to update it with the new
     * matrix A, or modify a copy of A.
     */
-  explicit ConjugateGradient(const MatrixType& A) : Base(A) {}
+  template<typename MatrixDerived>
+  explicit ConjugateGradient(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
 
   ~ConjugateGradient() {}
 
@@ -191,12 +196,19 @@ public:
   template<typename Rhs,typename Dest>
   void _solve_with_guess_impl(const Rhs& b, Dest& x) const
   {
-    typedef Ref<const MatrixType> MatRef;
-    typedef typename internal::conditional<UpLo==(Lower|Upper) && (!MatrixType::IsRowMajor) && (!NumTraits<Scalar>::IsComplex),
-                                           Transpose<const MatRef>, MatRef const&>::type RowMajorWrapper;
+    typedef typename Base::MatrixWrapper MatrixWrapper;
+    typedef typename Base::ActualMatrixType ActualMatrixType;
+    enum {
+      TransposeInput  =   (!MatrixWrapper::MatrixFree)
+                      &&  (UpLo==(Lower|Upper))
+                      &&  (!MatrixType::IsRowMajor)
+                      &&  (!NumTraits<Scalar>::IsComplex)
+    };
+    typedef typename internal::conditional<TransposeInput,Transpose<const ActualMatrixType>, ActualMatrixType const&>::type RowMajorWrapper;
+    EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(MatrixWrapper::MatrixFree,UpLo==(Lower|Upper)),MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY);
     typedef typename internal::conditional<UpLo==(Lower|Upper),
                                            RowMajorWrapper,
-                                           typename MatRef::template ConstSelfAdjointViewReturnType<UpLo>::Type
+                                           typename MatrixWrapper::template ConstSelfAdjointViewReturnType<UpLo>::Type
                                           >::type SelfAdjointWrapper;
     m_iterations = Base::maxIterations();
     m_error = Base::m_tolerance;
@@ -207,7 +219,7 @@ public:
       m_error = Base::m_tolerance;
 
       typename Dest::ColXpr xj(x,j);
-      RowMajorWrapper row_mat(mp_matrix);
+      RowMajorWrapper row_mat(matrix());
       internal::conjugate_gradient(SelfAdjointWrapper(row_mat), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error);
     }