bump to 3.3-alpha1

. Missing } and unprotected min/max calls and definitions.

.hgignore

@@ -30,3 +30,5 @@ log
 patch
 a
 a.*
+lapack/testing
+lapack/reference

CMakeLists.txt

@@ -1,6 +1,6 @@
 project(Eigen)
 
-cmake_minimum_required(VERSION 2.8.2)
+cmake_minimum_required(VERSION 2.8.4)
 
 # guard against in-source builds
 
@@ -147,6 +147,12 @@ if(NOT MSVC)
   ei_add_cxx_compiler_flag("-Wenum-conversion")
   ei_add_cxx_compiler_flag("-Wc++11-extensions")
   
+  # -Wshadow is insanely too strict with gcc, hopefully it will become usable with gcc 6
+  # if(NOT CMAKE_COMPILER_IS_GNUCXX OR (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER "5.0.0"))
+  if(NOT CMAKE_COMPILER_IS_GNUCXX)
+    ei_add_cxx_compiler_flag("-Wshadow")
+  endif()
+  
   ei_add_cxx_compiler_flag("-Wno-psabi")
   ei_add_cxx_compiler_flag("-Wno-variadic-macros")
   ei_add_cxx_compiler_flag("-Wno-long-long")
@@ -168,6 +174,11 @@ if(NOT MSVC)
   else()
     ei_add_cxx_compiler_flag("-ansi")
   endif()
+
+  if(ANDROID_NDK)
+    ei_add_cxx_compiler_flag("-pie")
+    ei_add_cxx_compiler_flag("-fPIE")
+  endif()
   
   set(CMAKE_REQUIRED_FLAGS "")
 
@@ -208,7 +219,7 @@ if(NOT MSVC)
   endif()
 
   option(EIGEN_TEST_FMA "Enable/Disable FMA in tests/examples" OFF)
-  if(EIGEN_TEST_FMA)
+  if(EIGEN_TEST_FMA AND NOT EIGEN_TEST_NEON)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfma")
     message(STATUS "Enabling FMA in tests/examples")
   endif()
@@ -219,12 +230,31 @@ if(NOT MSVC)
     message(STATUS "Enabling AltiVec in tests/examples")
   endif()
 
+  option(EIGEN_TEST_VSX "Enable/Disable VSX in tests/examples" OFF)
+  if(EIGEN_TEST_VSX)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64 -mvsx")
+    message(STATUS "Enabling VSX in tests/examples")
+  endif()
+
   option(EIGEN_TEST_NEON "Enable/Disable Neon in tests/examples" OFF)
   if(EIGEN_TEST_NEON)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon -mcpu=cortex-a8")
+    if(EIGEN_TEST_FMA)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon-vfpv4")
+    else()
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon")
+    endif()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfloat-abi=softfp")
     message(STATUS "Enabling NEON in tests/examples")
   endif()
 
+  option(EIGEN_TEST_NEON64 "Enable/Disable Neon in tests/examples" OFF)
+  if(EIGEN_TEST_NEON64)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    message(STATUS "Enabling NEON in tests/examples")
+  endif()
+
+
+
   check_cxx_compiler_flag("-fopenmp" COMPILER_SUPPORT_OPENMP)
   if(COMPILER_SUPPORT_OPENMP)
     option(EIGEN_TEST_OPENMP "Enable/Disable OpenMP in tests/examples" OFF)
@@ -307,7 +337,7 @@ if(EIGEN_TEST_NO_EXCEPTIONS)
   message(STATUS "Disabling exceptions in tests/examples")
 endif()
 
-option(EIGEN_TEST_C++0x "Enables all C++0x features." OFF)
+option(EIGEN_TEST_CXX11 "Enable testing with C++11 and C++11 features (e.g. Tensor module)." OFF)
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
 
@@ -434,6 +464,7 @@ if(cmake_generator_tolower MATCHES "makefile")
   message(STATUS "make check    | Build and run the unit-tests. Read this page:")
   message(STATUS "              |   http://eigen.tuxfamily.org/index.php?title=Tests")
   message(STATUS "make blas     | Build BLAS library (not the same thing as Eigen)")
+  message(STATUS "make uninstall| Removes files installed by make install")
   message(STATUS "--------------+--------------------------------------------------------------")
 else()
   message(STATUS "To build/run the unit tests, read this page:")
@@ -469,3 +500,7 @@ install ( FILES ${CMAKE_CURRENT_SOURCE_DIR}/cmake/UseEigen3.cmake
                 ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
           DESTINATION ${EIGEN_CONFIG_CMAKE_PATH}
         )
+
+# Add uninstall target
+add_custom_target ( uninstall
+    COMMAND ${CMAKE_COMMAND} -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/EigenUninstall.cmake)

Eigen/Cholesky

@@ -21,7 +21,6 @@
   * \endcode
   */
 
-#include "src/misc/Solve.h"
 #include "src/Cholesky/LLT.h"
 #include "src/Cholesky/LDLT.h"
 #ifdef EIGEN_USE_LAPACKE

Eigen/CholmodSupport

@@ -33,12 +33,8 @@ extern "C" {
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/CholmodSupport/CholmodSupport.h"
 
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_CHOLMODSUPPORT_MODULE_H

Eigen/Core

@@ -24,9 +24,15 @@
   #ifdef EIGEN_INTERNAL_DEBUGGING
   #undef EIGEN_INTERNAL_DEBUGGING
   #endif
-  
+
   // Do not try to vectorize on CUDA!
+  #ifndef EIGEN_DONT_VECTORIZE
   #define EIGEN_DONT_VECTORIZE
+  #endif
+
+  #ifdef EIGEN_EXCEPTIONS
+  #undef EIGEN_EXCEPTIONS
+  #endif
   
   // All functions callable from CUDA code must be qualified with __device__
   #define EIGEN_DEVICE_FUNC __host__ __device__
@@ -57,7 +63,7 @@
 
 // Disable the ipa-cp-clone optimization flag with MinGW 6.x or newer (enabled by default with -O3)
 // See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=556 for details.
-#if defined(__MINGW32__) && EIGEN_GNUC_AT_LEAST(4,6)
+#if EIGEN_COMP_MINGW && EIGEN_GNUC_AT_LEAST(4,6)
   #pragma GCC optimize ("-fno-ipa-cp-clone")
 #endif
 
@@ -67,26 +73,26 @@
 // and inclusion of their respective header files
 #include "src/Core/util/MKL_support.h"
 
-// if alignment is disabled, then disable vectorization. Note: EIGEN_ALIGN is the proper check, it takes into
-// account both the user's will (EIGEN_DONT_ALIGN) and our own platform checks
-#if !EIGEN_ALIGN
+// if alignment is disabled, then disable vectorization. Note: EIGEN_MAX_ALIGN_BYTES is the proper check, it takes into
+// account both the user's will (EIGEN_MAX_ALIGN_BYTES,EIGEN_DONT_ALIGN) and our own platform checks
+#if EIGEN_MAX_ALIGN_BYTES==0
   #ifndef EIGEN_DONT_VECTORIZE
     #define EIGEN_DONT_VECTORIZE
   #endif
 #endif
 
-#ifdef _MSC_VER
+#if EIGEN_COMP_MSVC
   #include <malloc.h> // for _aligned_malloc -- need it regardless of whether vectorization is enabled
-  #if (_MSC_VER >= 1500) // 2008 or later
+  #if (EIGEN_COMP_MSVC >= 1500) // 2008 or later
     // Remember that usage of defined() in a #define is undefined by the standard.
     // a user reported that in 64-bit mode, MSVC doesn't care to define _M_IX86_FP.
-    #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || defined(_M_X64)
+    #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || EIGEN_ARCH_x86_64
       #define EIGEN_SSE2_ON_MSVC_2008_OR_LATER
     #endif
   #endif
 #else
   // Remember that usage of defined() in a #define is undefined by the standard
-  #if (defined __SSE2__) && ( (!defined __GNUC__) || (defined __INTEL_COMPILER) || EIGEN_GNUC_AT_LEAST(4,2) )
+  #if (defined __SSE2__) && ( (!EIGEN_COMP_GNUC) || EIGEN_COMP_ICC || EIGEN_GNUC_AT_LEAST(4,2) )
     #define EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC
   #endif
 #endif
@@ -125,9 +131,13 @@
       #define EIGEN_VECTORIZE_SSE4_1
       #define EIGEN_VECTORIZE_SSE4_2
     #endif
+    #ifdef __AVX2__
+      #define EIGEN_VECTORIZE_AVX2
+    #endif
     #ifdef __FMA__
       #define EIGEN_VECTORIZE_FMA
     #endif
+
     // include files
 
     // This extern "C" works around a MINGW-w64 compilation issue
@@ -140,7 +150,7 @@
     extern "C" {
       // In theory we should only include immintrin.h and not the other *mmintrin.h header files directly.
       // Doing so triggers some issues with ICC. However old gcc versions seems to not have this file, thus:
-      #if defined(__INTEL_COMPILER) && __INTEL_COMPILER >= 1110
+      #if EIGEN_COMP_ICC >= 1110
         #include <immintrin.h>
       #else
         #include <emmintrin.h>
@@ -162,6 +172,15 @@
         #endif
       #endif
     } // end extern "C"
+  #elif defined __VSX__
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_VSX
+    #include <altivec.h>
+    // We need to #undef all these ugly tokens defined in <altivec.h>
+    // => use __vector instead of vector
+    #undef bool
+    #undef vector
+    #undef pixel
   #elif defined __ALTIVEC__
     #define EIGEN_VECTORIZE
     #define EIGEN_VECTORIZE_ALTIVEC
@@ -171,13 +190,18 @@
     #undef bool
     #undef vector
     #undef pixel
-  #elif defined  __ARM_NEON__
+  #elif (defined  __ARM_NEON) || (defined __ARM_NEON__)
     #define EIGEN_VECTORIZE
     #define EIGEN_VECTORIZE_NEON
     #include <arm_neon.h>
   #endif
 #endif
 
+#if defined __CUDACC__
+  #define EIGEN_VECTORIZE_CUDA
+  #include <vector_types.h>
+#endif
+
 #if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE)
   #define EIGEN_HAS_OPENMP
 #endif
@@ -187,7 +211,7 @@
 #endif
 
 // MSVC for windows mobile does not have the errno.h file
-#if !(defined(_MSC_VER) && defined(_WIN32_WCE)) && !defined(__ARMCC_VERSION)
+#if !(EIGEN_COMP_MSVC && EIGEN_OS_WINCE) && !EIGEN_COMP_ARM
 #define EIGEN_HAS_ERRNO
 #endif
 
@@ -213,7 +237,7 @@
 #endif
 
 // required for __cpuid, needs to be included after cmath
-#if defined(_MSC_VER) && (defined(_M_IX86)||defined(_M_X64)) && (!defined(_WIN32_WCE))
+#if EIGEN_COMP_MSVC && EIGEN_ARCH_i386_OR_x86_64 && !EIGEN_OS_WINCE
   #include <intrin.h>
 #endif
 
@@ -235,6 +259,8 @@ inline static const char *SimdInstructionSetsInUse(void) {
   return "SSE, SSE2";
 #elif defined(EIGEN_VECTORIZE_ALTIVEC)
   return "AltiVec";
+#elif defined(EIGEN_VECTORIZE_VSX)
+  return "VSX";
 #elif defined(EIGEN_VECTORIZE_NEON)
   return "ARM NEON";
 #else
@@ -266,8 +292,8 @@ using std::ptrdiff_t;
   */
 
 #include "src/Core/util/Constants.h"
-#include "src/Core/util/ForwardDeclarations.h"
 #include "src/Core/util/Meta.h"
+#include "src/Core/util/ForwardDeclarations.h"
 #include "src/Core/util/StaticAssert.h"
 #include "src/Core/util/XprHelper.h"
 #include "src/Core/util/Memory.h"
@@ -280,39 +306,47 @@ using std::ptrdiff_t;
   // Use AVX for floats and doubles, SSE for integers
   #include "src/Core/arch/SSE/PacketMath.h"
   #include "src/Core/arch/SSE/Complex.h"
+  #include "src/Core/arch/SSE/MathFunctions.h"
   #include "src/Core/arch/AVX/PacketMath.h"
+  #include "src/Core/arch/AVX/MathFunctions.h"
   #include "src/Core/arch/AVX/Complex.h"
+  #include "src/Core/arch/AVX/TypeCasting.h"
 #elif defined EIGEN_VECTORIZE_SSE
   #include "src/Core/arch/SSE/PacketMath.h"
   #include "src/Core/arch/SSE/MathFunctions.h"
   #include "src/Core/arch/SSE/Complex.h"
-#elif defined EIGEN_VECTORIZE_ALTIVEC
+  #include "src/Core/arch/SSE/TypeCasting.h"
+#elif defined(EIGEN_VECTORIZE_ALTIVEC) || defined(EIGEN_VECTORIZE_VSX)
   #include "src/Core/arch/AltiVec/PacketMath.h"
+  #include "src/Core/arch/AltiVec/MathFunctions.h"
   #include "src/Core/arch/AltiVec/Complex.h"
 #elif defined EIGEN_VECTORIZE_NEON
   #include "src/Core/arch/NEON/PacketMath.h"
+  #include "src/Core/arch/NEON/MathFunctions.h"
   #include "src/Core/arch/NEON/Complex.h"
 #endif
 
+#if defined EIGEN_VECTORIZE_CUDA
+  #include "src/Core/arch/CUDA/PacketMath.h"
+  #include "src/Core/arch/CUDA/MathFunctions.h"
+#endif
+
 #include "src/Core/arch/Default/Settings.h"
 
 #include "src/Core/functors/BinaryFunctors.h"
 #include "src/Core/functors/UnaryFunctors.h"
 #include "src/Core/functors/NullaryFunctors.h"
 #include "src/Core/functors/StlFunctors.h"
+#include "src/Core/functors/AssignmentFunctors.h"
 
 #include "src/Core/DenseCoeffsBase.h"
 #include "src/Core/DenseBase.h"
 #include "src/Core/MatrixBase.h"
 #include "src/Core/EigenBase.h"
 
-#ifdef EIGEN_ENABLE_EVALUATORS
-#include "src/Core/functors/AssignmentFunctors.h"
 #include "src/Core/Product.h"
 #include "src/Core/CoreEvaluators.h"
 #include "src/Core/AssignEvaluator.h"
-#include "src/Core/ProductEvaluators.h"
-#endif
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN // work around Doxygen bug triggered by Assign.h r814874
                                 // at least confirmed with Doxygen 1.5.5 and 1.5.6
@@ -323,7 +357,9 @@ using std::ptrdiff_t;
 #include "src/Core/util/BlasUtil.h"
 #include "src/Core/DenseStorage.h"
 #include "src/Core/NestByValue.h"
-#include "src/Core/ForceAlignedAccess.h"
+
+// #include "src/Core/ForceAlignedAccess.h"
+
 #include "src/Core/ReturnByValue.h"
 #include "src/Core/NoAlias.h"
 #include "src/Core/PlainObjectBase.h"
@@ -336,12 +372,12 @@ using std::ptrdiff_t;
 #include "src/Core/SelfCwiseBinaryOp.h"
 #include "src/Core/Dot.h"
 #include "src/Core/StableNorm.h"
-#include "src/Core/MapBase.h"
 #include "src/Core/Stride.h"
+#include "src/Core/MapBase.h"
 #include "src/Core/Map.h"
+#include "src/Core/Ref.h"
 #include "src/Core/Block.h"
 #include "src/Core/VectorBlock.h"
-#include "src/Core/Ref.h"
 #include "src/Core/Transpose.h"
 #include "src/Core/DiagonalMatrix.h"
 #include "src/Core/Diagonal.h"
@@ -354,14 +390,14 @@ using std::ptrdiff_t;
 #include "src/Core/IO.h"
 #include "src/Core/Swap.h"
 #include "src/Core/CommaInitializer.h"
-#include "src/Core/Flagged.h"
-#include "src/Core/ProductBase.h"
 #include "src/Core/GeneralProduct.h"
+#include "src/Core/Solve.h"
+#include "src/Core/Inverse.h"
 #include "src/Core/TriangularMatrix.h"
 #include "src/Core/SelfAdjointView.h"
 #include "src/Core/products/GeneralBlockPanelKernel.h"
 #include "src/Core/products/Parallelizer.h"
-#include "src/Core/products/CoeffBasedProduct.h"
+#include "src/Core/ProductEvaluators.h"
 #include "src/Core/products/GeneralMatrixVector.h"
 #include "src/Core/products/GeneralMatrixMatrix.h"
 #include "src/Core/SolveTriangular.h"

Eigen/Geometry

@@ -9,10 +9,6 @@
 #include "LU"
 #include <limits>
 
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
-#endif
-
 /** \defgroup Geometry_Module Geometry module
   *
   *

Eigen/IterativeLinearSolvers

@@ -12,26 +12,26 @@
   * This module currently provides iterative methods to solve problems of the form \c A \c x = \c b, where \c A is a squared matrix, usually very large and sparse.
   * Those solvers are accessible via the following classes:
   *  - ConjugateGradient for selfadjoint (hermitian) matrices,
+  *  - LeastSquaresConjugateGradient for rectangular least-square problems,
   *  - BiCGSTAB for general square matrices.
   *
   * These iterative solvers are associated with some preconditioners:
   *  - IdentityPreconditioner - not really useful
-  *  - DiagonalPreconditioner - also called JAcobi preconditioner, work very well on diagonal dominant matrices.
-  *  - IncompleteILUT - incomplete LU factorization with dual thresholding
+  *  - DiagonalPreconditioner - also called Jacobi preconditioner, work very well on diagonal dominant matrices.
+  *  - IncompleteLUT - incomplete LU factorization with dual thresholding
   *
   * Such problems can also be solved using the direct sparse decomposition modules: SparseCholesky, CholmodSupport, UmfPackSupport, SuperLUSupport.
   *
-  * \code
-  * #include <Eigen/IterativeLinearSolvers>
-  * \endcode
+    \code
+    #include <Eigen/IterativeLinearSolvers>
+    \endcode
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
+#include "src/IterativeLinearSolvers/SolveWithGuess.h"
 #include "src/IterativeLinearSolvers/IterativeSolverBase.h"
 #include "src/IterativeLinearSolvers/BasicPreconditioners.h"
 #include "src/IterativeLinearSolvers/ConjugateGradient.h"
+#include "src/IterativeLinearSolvers/LeastSquareConjugateGradient.h"
 #include "src/IterativeLinearSolvers/BiCGSTAB.h"
 #include "src/IterativeLinearSolvers/IncompleteLUT.h"
 

Eigen/LU

@@ -16,7 +16,6 @@
   * \endcode
   */
 
-#include "src/misc/Solve.h"
 #include "src/misc/Kernel.h"
 #include "src/misc/Image.h"
 #include "src/LU/FullPivLU.h"
@@ -25,7 +24,7 @@
 #include "src/LU/PartialPivLU_MKL.h"
 #endif
 #include "src/LU/Determinant.h"
-#include "src/LU/Inverse.h"
+#include "src/LU/InverseImpl.h"
 
 // Use the SSE optimized version whenever possible. At the moment the
 // SSE version doesn't compile when AVX is enabled

Eigen/PaStiXSupport

@@ -35,12 +35,8 @@ extern "C" {
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/PaStiXSupport/PaStiXSupport.h"
 
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_PASTIXSUPPORT_MODULE_H

Eigen/QR

@@ -24,7 +24,6 @@
   * \endcode
   */
 
-#include "src/misc/Solve.h"
 #include "src/QR/HouseholderQR.h"
 #include "src/QR/FullPivHouseholderQR.h"
 #include "src/QR/ColPivHouseholderQR.h"

Eigen/SPQRSupport

@@ -21,8 +21,6 @@
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
 #include "src/CholmodSupport/CholmodSupport.h"
 #include "src/SPQRSupport/SuiteSparseQRSupport.h"
 

Eigen/SVD

@@ -12,21 +12,25 @@
   *
   *
   * This module provides SVD decomposition for matrices (both real and complex).
-  * This decomposition is accessible via the following MatrixBase method:
+  * Two decomposition algorithms are provided:
+  *  - JacobiSVD implementing two-sided Jacobi iterations is numerically very accurate, fast for small matrices, but very slow for larger ones.
+  *  - BDCSVD implementing a recursive divide & conquer strategy on top of an upper-bidiagonalization which remains fast for large problems.
+  * These decompositions are accessible via the respective classes and following MatrixBase methods:
   *  - MatrixBase::jacobiSvd()
+  *  - MatrixBase::bdcSvd()
   *
   * \code
   * #include <Eigen/SVD>
   * \endcode
   */
 
-#include "src/misc/Solve.h"
+#include "src/SVD/UpperBidiagonalization.h"
 #include "src/SVD/SVDBase.h"
 #include "src/SVD/JacobiSVD.h"
+#include "src/SVD/BDCSVD.h"
 #if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT)
 #include "src/SVD/JacobiSVD_MKL.h"
 #endif
-#include "src/SVD/UpperBidiagonalization.h"
 
 #include "src/Core/util/ReenableStupidWarnings.h"
 

Eigen/Sparse

@@ -11,9 +11,9 @@
   * - \ref SparseQR_Module
   * - \ref IterativeLinearSolvers_Module
   *
-  * \code
-  * #include <Eigen/Sparse>
-  * \endcode
+    \code
+    #include <Eigen/Sparse>
+    \endcode
   */
 
 #include "SparseCore"

Eigen/SparseCholesky

@@ -34,8 +34,6 @@
 #error The SparseCholesky module has nothing to offer in MPL2 only mode
 #endif
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
 #include "src/SparseCholesky/SimplicialCholesky.h"
 
 #ifndef EIGEN_MPL2_ONLY

Eigen/SparseCore

@@ -26,37 +26,35 @@
   * This module depends on: Core.
   */
 
-namespace Eigen {
-
-/** The type used to identify a general sparse storage. */
-struct Sparse {};
-
-}
-
 #include "src/SparseCore/SparseUtil.h"
 #include "src/SparseCore/SparseMatrixBase.h"
+#include "src/SparseCore/SparseAssign.h"
 #include "src/SparseCore/CompressedStorage.h"
 #include "src/SparseCore/AmbiVector.h"
+#include "src/SparseCore/SparseCompressedBase.h"
 #include "src/SparseCore/SparseMatrix.h"
+#include "src/SparseCore/SparseMap.h"
 #include "src/SparseCore/MappedSparseMatrix.h"
 #include "src/SparseCore/SparseVector.h"
-#include "src/SparseCore/SparseBlock.h"
-#include "src/SparseCore/SparseTranspose.h"
+#include "src/SparseCore/SparseRef.h"
 #include "src/SparseCore/SparseCwiseUnaryOp.h"
 #include "src/SparseCore/SparseCwiseBinaryOp.h"
+#include "src/SparseCore/SparseTranspose.h"
+#include "src/SparseCore/SparseBlock.h"
 #include "src/SparseCore/SparseDot.h"
-#include "src/SparseCore/SparsePermutation.h"
 #include "src/SparseCore/SparseRedux.h"
-#include "src/SparseCore/SparseFuzzy.h"
+#include "src/SparseCore/SparseView.h"
+#include "src/SparseCore/SparseDiagonalProduct.h"
 #include "src/SparseCore/ConservativeSparseSparseProduct.h"
 #include "src/SparseCore/SparseSparseProductWithPruning.h"
 #include "src/SparseCore/SparseProduct.h"
 #include "src/SparseCore/SparseDenseProduct.h"
-#include "src/SparseCore/SparseDiagonalProduct.h"
-#include "src/SparseCore/SparseTriangularView.h"
 #include "src/SparseCore/SparseSelfAdjointView.h"
+#include "src/SparseCore/SparseTriangularView.h"
 #include "src/SparseCore/TriangularSolver.h"
-#include "src/SparseCore/SparseView.h"
+#include "src/SparseCore/SparsePermutation.h"
+#include "src/SparseCore/SparseFuzzy.h"
+#include "src/SparseCore/SparseSolverBase.h"
 
 #include "src/Core/util/ReenableStupidWarnings.h"
 

Eigen/SparseLU

@@ -20,9 +20,6 @@
   * Please, see the documentation of the SparseLU class for more details.
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 // Ordering interface
 #include "OrderingMethods"
 

Eigen/SparseQR

@@ -21,9 +21,6 @@
   * 
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "OrderingMethods"
 #include "src/SparseCore/SparseColEtree.h"
 #include "src/SparseQR/SparseQR.h"

Eigen/StdDeque

@@ -14,7 +14,7 @@
 #include "Core"
 #include <deque>
 
-#if (defined(_MSC_VER) && defined(_WIN64)) /* MSVC auto aligns in 64 bit builds */
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */
 
 #define EIGEN_DEFINE_STL_DEQUE_SPECIALIZATION(...)
 

Eigen/StdList

@@ -13,7 +13,7 @@
 #include "Core"
 #include <list>
 
-#if (defined(_MSC_VER) && defined(_WIN64)) /* MSVC auto aligns in 64 bit builds */    
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */    
 
 #define EIGEN_DEFINE_STL_LIST_SPECIALIZATION(...)
 

Eigen/StdVector

@@ -14,7 +14,7 @@
 #include "Core"
 #include <vector>
 
-#if (defined(_MSC_VER) && defined(_WIN64)) /* MSVC auto aligns in 64 bit builds */
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */
 
 #define EIGEN_DEFINE_STL_VECTOR_SPECIALIZATION(...)
 

Eigen/SuperLUSupport

@@ -48,12 +48,8 @@ namespace Eigen { struct SluMatrix; }
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/SuperLUSupport/SuperLUSupport.h"
 
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_SUPERLUSUPPORT_MODULE_H

Eigen/UmfPackSupport

@@ -26,9 +26,6 @@ extern "C" {
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/UmfPackSupport/UmfPackSupport.h"
 
 #include "src/Core/util/ReenableStupidWarnings.h"

Eigen/src/Cholesky/LDLT.h

@@ -59,7 +59,8 @@ template<typename _MatrixType, int _UpLo> class LDLT
     };
     typedef typename MatrixType::Scalar Scalar;
     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
-    typedef typename MatrixType::Index Index;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename MatrixType::StorageIndex StorageIndex;
     typedef Matrix<Scalar, RowsAtCompileTime, 1, Options, MaxRowsAtCompileTime, 1> TmpMatrixType;
 
     typedef Transpositions<RowsAtCompileTime, MaxRowsAtCompileTime> TranspositionType;
@@ -85,7 +86,7 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * according to the specified problem \a size.
       * \sa LDLT()
       */
-    LDLT(Index size)
+    explicit LDLT(Index size)
       : m_matrix(size, size),
         m_transpositions(size),
         m_temporary(size),
@@ -98,7 +99,7 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * This calculates the decomposition for the input \a matrix.
       * \sa LDLT(Index size)
       */
-    LDLT(const MatrixType& matrix)
+    explicit LDLT(const MatrixType& matrix)
       : m_matrix(matrix.rows(), matrix.cols()),
         m_transpositions(matrix.rows()),
         m_temporary(matrix.rows()),
@@ -175,13 +176,13 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt()
       */
     template<typename Rhs>
-    inline const internal::solve_retval<LDLT, Rhs>
+    inline const Solve<LDLT, Rhs>
     solve(const MatrixBase<Rhs>& b) const
     {
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
       eigen_assert(m_matrix.rows()==b.rows()
                 && "LDLT::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::solve_retval<LDLT, Rhs>(*this, b.derived());
+      return Solve<LDLT, Rhs>(*this, b.derived());
     }
 
     template<typename Derived>
@@ -217,8 +218,19 @@ template<typename _MatrixType, int _UpLo> class LDLT
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
       return Success;
     }
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
 
     /** \internal
       * Used to compute and store the Cholesky decomposition A = L D L^* = U^* D U.
@@ -245,8 +257,7 @@ template<> struct ldlt_inplace<Lower>
     using std::abs;
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
-    typedef typename MatrixType::Index Index;
-    typedef typename TranspositionType::StorageIndexType IndexType;
+    typedef typename TranspositionType::StorageIndex IndexType;
     eigen_assert(mat.rows()==mat.cols());
     const Index size = mat.rows();
 
@@ -303,9 +314,9 @@ template<> struct ldlt_inplace<Lower>
       }
       
       // In some previous versions of Eigen (e.g., 3.2.1), the scaling was omitted if the pivot
-      // was smaller than the cutoff value. However, soince LDLT is not rank-revealing
-      // we should only make sure we do not introduce INF or NaN values.
-      // LAPACK also uses 0 as the cutoff value.
+      // was smaller than the cutoff value. However, since LDLT is not rank-revealing
+      // we should only make sure that we do not introduce INF or NaN values.
+      // Remark that LAPACK also uses 0 as the cutoff value.
       RealScalar realAkk = numext::real(mat.coeffRef(k,k));
       if((rs>0) && (abs(realAkk) > RealScalar(0)))
         A21 /= realAkk;
@@ -336,7 +347,6 @@ template<> struct ldlt_inplace<Lower>
     using numext::isfinite;
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
-    typedef typename MatrixType::Index Index;
 
     const Index size = mat.rows();
     eigen_assert(mat.cols() == size && w.size()==size);
@@ -400,16 +410,16 @@ template<typename MatrixType> struct LDLT_Traits<MatrixType,Lower>
 {
   typedef const TriangularView<const MatrixType, UnitLower> MatrixL;
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitUpper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m; }
-  static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
 };
 
 template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
 {
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitLower> MatrixL;
   typedef const TriangularView<const MatrixType, UnitUpper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
-  static inline MatrixU getU(const MatrixType& m) { return m; }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); }
 };
 
 } // end namespace internal
@@ -419,6 +429,8 @@ template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
 template<typename MatrixType, int _UpLo>
 LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
 {
+  check_template_parameters();
+  
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
 
@@ -427,6 +439,7 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
   m_transpositions.resize(size);
   m_isInitialized = false;
   m_temporary.resize(size);
+  m_sign = internal::ZeroSign;
 
   internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, m_sign);
 
@@ -441,9 +454,9 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
   */
 template<typename MatrixType, int _UpLo>
 template<typename Derived>
-LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w, const typename NumTraits<typename MatrixType::Scalar>::Real& sigma)
+LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w, const typename LDLT<MatrixType,_UpLo>::RealScalar& sigma)
 {
-  typedef typename TranspositionType::StorageIndexType IndexType;
+  typedef typename TranspositionType::StorageIndex IndexType;
   const Index size = w.rows();
   if (m_isInitialized)
   {
@@ -466,52 +479,45 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Deri
   return *this;
 }
 
-namespace internal {
-template<typename _MatrixType, int _UpLo, typename Rhs>
-struct solve_retval<LDLT<_MatrixType,_UpLo>, Rhs>
-  : solve_retval_base<LDLT<_MatrixType,_UpLo>, Rhs>
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType, int _UpLo>
+template<typename RhsType, typename DstType>
+void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
 {
-  typedef LDLT<_MatrixType,_UpLo> LDLTType;
-  EIGEN_MAKE_SOLVE_HELPERS(LDLTType,Rhs)
+  eigen_assert(rhs.rows() == rows());
+  // dst = P b
+  dst = m_transpositions * rhs;
 
-  template<typename Dest> void evalTo(Dest& dst) const
+  // dst = L^-1 (P b)
+  matrixL().solveInPlace(dst);
+
+  // dst = D^-1 (L^-1 P b)
+  // more precisely, use pseudo-inverse of D (see bug 241)
+  using std::abs;
+  const typename Diagonal<const MatrixType>::RealReturnType vecD(vectorD());
+  // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon
+  // as motivated by LAPACK's xGELSS:
+  // RealScalar tolerance = numext::maxi(vecD.array().abs().maxCoeff() * NumTraits<RealScalar>::epsilon(),RealScalar(1) / NumTraits<RealScalar>::highest());
+  // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest
+  // diagonal element is not well justified and leads to numerical issues in some cases.
+  // Moreover, Lapack's xSYTRS routines use 0 for the tolerance.
+  RealScalar tolerance = RealScalar(1) / NumTraits<RealScalar>::highest();
+  
+  for (Index i = 0; i < vecD.size(); ++i)
   {
-    eigen_assert(rhs().rows() == dec().matrixLDLT().rows());
-    // dst = P b
-    dst = dec().transpositionsP() * rhs();
-
-    // dst = L^-1 (P b)
-    dec().matrixL().solveInPlace(dst);
-
-    // dst = D^-1 (L^-1 P b)
-    // more precisely, use pseudo-inverse of D (see bug 241)
-    using std::abs;
-    EIGEN_USING_STD_MATH(max);
-    typedef typename LDLTType::MatrixType MatrixType;
-    typedef typename LDLTType::RealScalar RealScalar;
-    const typename Diagonal<const MatrixType>::RealReturnType vectorD(dec().vectorD());
-    // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon
-    // as motivated by LAPACK's xGELSS:
-    // RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() *NumTraits<RealScalar>::epsilon(),RealScalar(1) / NumTraits<RealScalar>::highest());
-    // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest
-    // diagonal element is not well justified and to numerical issues in some cases.
-    // Moreover, Lapack's xSYTRS routines use 0 for the tolerance.
-    RealScalar tolerance = RealScalar(1) / NumTraits<RealScalar>::highest();
-    for (Index i = 0; i < vectorD.size(); ++i) {
-      if(abs(vectorD(i)) > tolerance)
-        dst.row(i) /= vectorD(i);
-      else
-        dst.row(i).setZero();
-    }
-
-    // dst = L^-T (D^-1 L^-1 P b)
-    dec().matrixU().solveInPlace(dst);
-
-    // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b
-    dst = dec().transpositionsP().transpose() * dst;
+    if(abs(vecD(i)) > tolerance)
+      dst.row(i) /= vecD(i);
+    else
+      dst.row(i).setZero();
   }
-};
+
+  // dst = L^-T (D^-1 L^-1 P b)
+  matrixU().solveInPlace(dst);
+
+  // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b
+  dst = m_transpositions.transpose() * dst;
 }
+#endif
 
 /** \internal use x = ldlt_object.solve(x);
   *

Eigen/src/Cholesky/LLT.h

@@ -59,7 +59,8 @@ template<typename _MatrixType, int _UpLo> class LLT
     };
     typedef typename MatrixType::Scalar Scalar;
     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
-    typedef typename MatrixType::Index Index;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename MatrixType::StorageIndex StorageIndex;
 
     enum {
       PacketSize = internal::packet_traits<Scalar>::size,
@@ -83,10 +84,10 @@ template<typename _MatrixType, int _UpLo> class LLT
       * according to the specified problem \a size.
       * \sa LLT()
       */
-    LLT(Index size) : m_matrix(size, size),
+    explicit LLT(Index size) : m_matrix(size, size),
                     m_isInitialized(false) {}
 
-    LLT(const MatrixType& matrix)
+    explicit LLT(const MatrixType& matrix)
       : m_matrix(matrix.rows(), matrix.cols()),
         m_isInitialized(false)
     {
@@ -118,13 +119,13 @@ template<typename _MatrixType, int _UpLo> class LLT
       * \sa solveInPlace(), MatrixBase::llt(), SelfAdjointView::llt()
       */
     template<typename Rhs>
-    inline const internal::solve_retval<LLT, Rhs>
+    inline const Solve<LLT, Rhs>
     solve(const MatrixBase<Rhs>& b) const
     {
       eigen_assert(m_isInitialized && "LLT is not initialized.");
       eigen_assert(m_matrix.rows()==b.rows()
                 && "LLT::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::solve_retval<LLT, Rhs>(*this, b.derived());
+      return Solve<LLT, Rhs>(*this, b.derived());
     }
 
     template<typename Derived>
@@ -161,8 +162,20 @@ template<typename _MatrixType, int _UpLo> class LLT
 
     template<typename VectorType>
     LLT rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     /** \internal
       * Used to compute and store L
       * The strict upper part is not used and even not initialized.
@@ -177,12 +190,11 @@ namespace internal {
 template<typename Scalar, int UpLo> struct llt_inplace;
 
 template<typename MatrixType, typename VectorType>
-static typename MatrixType::Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma)
+static Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma)
 {
   using std::sqrt;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::ColXpr ColXpr;
   typedef typename internal::remove_all<ColXpr>::type ColXprCleaned;
   typedef typename ColXprCleaned::SegmentReturnType ColXprSegment;
@@ -251,10 +263,9 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
   template<typename MatrixType>
-  static typename MatrixType::Index unblocked(MatrixType& mat)
+  static Index unblocked(MatrixType& mat)
   {
     using std::sqrt;
-    typedef typename MatrixType::Index Index;
     
     eigen_assert(mat.rows()==mat.cols());
     const Index size = mat.rows();
@@ -278,9 +289,8 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
   }
 
   template<typename MatrixType>
-  static typename MatrixType::Index blocked(MatrixType& m)
+  static Index blocked(MatrixType& m)
   {
-    typedef typename MatrixType::Index Index;
     eigen_assert(m.rows()==m.cols());
     Index size = m.rows();
     if(size<32)
@@ -311,7 +321,7 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
   }
 
   template<typename MatrixType, typename VectorType>
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
   {
     return Eigen::internal::llt_rank_update_lower(mat, vec, sigma);
   }
@@ -322,19 +332,19 @@ template<typename Scalar> struct llt_inplace<Scalar, Upper>
   typedef typename NumTraits<Scalar>::Real RealScalar;
 
   template<typename MatrixType>
-  static EIGEN_STRONG_INLINE typename MatrixType::Index unblocked(MatrixType& mat)
+  static EIGEN_STRONG_INLINE Index unblocked(MatrixType& mat)
   {
     Transpose<MatrixType> matt(mat);
     return llt_inplace<Scalar, Lower>::unblocked(matt);
   }
   template<typename MatrixType>
-  static EIGEN_STRONG_INLINE typename MatrixType::Index blocked(MatrixType& mat)
+  static EIGEN_STRONG_INLINE Index blocked(MatrixType& mat)
   {
     Transpose<MatrixType> matt(mat);
     return llt_inplace<Scalar, Lower>::blocked(matt);
   }
   template<typename MatrixType, typename VectorType>
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
   {
     Transpose<MatrixType> matt(mat);
     return llt_inplace<Scalar, Lower>::rankUpdate(matt, vec.conjugate(), sigma);
@@ -345,8 +355,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Lower>
 {
   typedef const TriangularView<const MatrixType, Lower> MatrixL;
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, Upper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m; }
-  static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
   static bool inplace_decomposition(MatrixType& m)
   { return llt_inplace<typename MatrixType::Scalar, Lower>::blocked(m)==-1; }
 };
@@ -355,8 +365,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
 {
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, Lower> MatrixL;
   typedef const TriangularView<const MatrixType, Upper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
-  static inline MatrixU getU(const MatrixType& m) { return m; }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); }
   static bool inplace_decomposition(MatrixType& m)
   { return llt_inplace<typename MatrixType::Scalar, Upper>::blocked(m)==-1; }
 };
@@ -373,6 +383,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
 template<typename MatrixType, int _UpLo>
 LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
 {
+  check_template_parameters();
+  
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
   m_matrix.resize(size, size);
@@ -404,22 +416,16 @@ LLT<_MatrixType,_UpLo> LLT<_MatrixType,_UpLo>::rankUpdate(const VectorType& v, c
 
   return *this;
 }
-    
-namespace internal {
-template<typename _MatrixType, int UpLo, typename Rhs>
-struct solve_retval<LLT<_MatrixType, UpLo>, Rhs>
-  : solve_retval_base<LLT<_MatrixType, UpLo>, Rhs>
+ 
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType,int _UpLo>
+template<typename RhsType, typename DstType>
+void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
 {
-  typedef LLT<_MatrixType,UpLo> LLTType;
-  EIGEN_MAKE_SOLVE_HELPERS(LLTType,Rhs)
-
-  template<typename Dest> void evalTo(Dest& dst) const
-  {
-    dst = rhs();
-    dec().solveInPlace(dst);
-  }
-};
+  dst = rhs;
+  solveInPlace(dst);
 }
+#endif
 
 /** \internal use x = llt_object.solve(x);
   * 

Eigen/src/Cholesky/LLT_MKL.h

@@ -46,7 +46,7 @@ template<typename Scalar> struct mkl_llt;
 template<> struct mkl_llt<EIGTYPE> \
 { \
   template<typename MatrixType> \
-  static inline typename MatrixType::Index potrf(MatrixType& m, char uplo) \
+  static inline Index potrf(MatrixType& m, char uplo) \
   { \
     lapack_int matrix_order; \
     lapack_int size, lda, info, StorageOrder; \
@@ -60,30 +60,30 @@ template<> struct mkl_llt<EIGTYPE> \
     lda = m.outerStride(); \
 \
     info = LAPACKE_##MKLPREFIX##potrf( matrix_order, uplo, size, (MKLTYPE*)a, lda ); \
-    info = (info==0) ? Success : NumericalIssue; \
+    info = (info==0) ? -1 : info>0 ? info-1 : size; \
     return info; \
   } \
 }; \
 template<> struct llt_inplace<EIGTYPE, Lower> \
 { \
   template<typename MatrixType> \
-  static typename MatrixType::Index blocked(MatrixType& m) \
+  static Index blocked(MatrixType& m) \
   { \
     return mkl_llt<EIGTYPE>::potrf(m, 'L'); \
   } \
   template<typename MatrixType, typename VectorType> \
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
   { return Eigen::internal::llt_rank_update_lower(mat, vec, sigma); } \
 }; \
 template<> struct llt_inplace<EIGTYPE, Upper> \
 { \
   template<typename MatrixType> \
-  static typename MatrixType::Index blocked(MatrixType& m) \
+  static Index blocked(MatrixType& m) \
   { \
     return mkl_llt<EIGTYPE>::potrf(m, 'U'); \
   } \
   template<typename MatrixType, typename VectorType> \
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
   { \
     Transpose<MatrixType> matt(mat); \
     return llt_inplace<EIGTYPE, Lower>::rankUpdate(matt, vec.conjugate(), sigma); \

Eigen/src/CholmodSupport/CholmodSupport.h

@@ -48,8 +48,8 @@ void cholmod_configure_matrix(CholmodType& mat)
 /** Wraps the Eigen sparse matrix \a mat into a Cholmod sparse matrix object.
   * Note that the data are shared.
   */
-template<typename _Scalar, int _Options, typename _Index>
-cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat)
+template<typename _Scalar, int _Options, typename _StorageIndex>
+cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_StorageIndex>& mat)
 {
   cholmod_sparse res;
   res.nzmax   = mat.nonZeros();
@@ -74,11 +74,11 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat)
   res.dtype   = 0;
   res.stype   = -1;
   
-  if (internal::is_same<_Index,int>::value)
+  if (internal::is_same<_StorageIndex,int>::value)
   {
     res.itype = CHOLMOD_INT;
   }
-  else if (internal::is_same<_Index,UF_long>::value)
+  else if (internal::is_same<_StorageIndex,UF_long>::value)
   {
     res.itype = CHOLMOD_LONG;
   }
@@ -105,7 +105,7 @@ const cholmod_sparse viewAsCholmod(const SparseMatrix<_Scalar,_Options,_Index>&
 /** Returns a view of the Eigen sparse matrix \a mat as Cholmod sparse matrix.
   * The data are not copied but shared. */
 template<typename _Scalar, int _Options, typename _Index, unsigned int UpLo>
-cholmod_sparse viewAsCholmod(const SparseSelfAdjointView<SparseMatrix<_Scalar,_Options,_Index>, UpLo>& mat)
+cholmod_sparse viewAsCholmod(const SparseSelfAdjointView<const SparseMatrix<_Scalar,_Options,_Index>, UpLo>& mat)
 {
   cholmod_sparse res = viewAsCholmod(mat.matrix().const_cast_derived());
   
@@ -138,12 +138,12 @@ cholmod_dense viewAsCholmod(MatrixBase<Derived>& mat)
 
 /** Returns a view of the Cholmod sparse matrix \a cm as an Eigen sparse matrix.
   * The data are not copied but shared. */
-template<typename Scalar, int Flags, typename Index>
-MappedSparseMatrix<Scalar,Flags,Index> viewAsEigen(cholmod_sparse& cm)
+template<typename Scalar, int Flags, typename StorageIndex>
+MappedSparseMatrix<Scalar,Flags,StorageIndex> viewAsEigen(cholmod_sparse& cm)
 {
-  return MappedSparseMatrix<Scalar,Flags,Index>
-         (cm.nrow, cm.ncol, static_cast<Index*>(cm.p)[cm.ncol],
-          static_cast<Index*>(cm.p), static_cast<Index*>(cm.i),static_cast<Scalar*>(cm.x) );
+  return MappedSparseMatrix<Scalar,Flags,StorageIndex>
+         (cm.nrow, cm.ncol, static_cast<StorageIndex*>(cm.p)[cm.ncol],
+          static_cast<StorageIndex*>(cm.p), static_cast<StorageIndex*>(cm.i),static_cast<Scalar*>(cm.x) );
 }
 
 enum CholmodMode {
@@ -157,27 +157,31 @@ enum CholmodMode {
   * \sa class CholmodSupernodalLLT, class CholmodSimplicialLDLT, class CholmodSimplicialLLT
   */
 template<typename _MatrixType, int _UpLo, typename Derived>
-class CholmodBase : internal::noncopyable
+class CholmodBase : public SparseSolverBase<Derived>
 {
+  protected:
+    typedef SparseSolverBase<Derived> Base;
+    using Base::derived;
+    using Base::m_isInitialized;
   public:
     typedef _MatrixType MatrixType;
     enum { UpLo = _UpLo };
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
     typedef MatrixType CholMatrixType;
-    typedef typename MatrixType::Index Index;
+    typedef typename MatrixType::StorageIndex StorageIndex;
 
   public:
 
     CholmodBase()
-      : m_cholmodFactor(0), m_info(Success), m_isInitialized(false)
+      : m_cholmodFactor(0), m_info(Success)
     {
       m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0);
       cholmod_start(&m_cholmod);
     }
 
-    CholmodBase(const MatrixType& matrix)
-      : m_cholmodFactor(0), m_info(Success), m_isInitialized(false)
+    explicit CholmodBase(const MatrixType& matrix)
+      : m_cholmodFactor(0), m_info(Success)
     {
       m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0);
       cholmod_start(&m_cholmod);
@@ -191,11 +195,8 @@ class CholmodBase : internal::noncopyable
       cholmod_finish(&m_cholmod);
     }
     
-    inline Index cols() const { return m_cholmodFactor->n; }
-    inline Index rows() const { return m_cholmodFactor->n; }
-    
-    Derived& derived() { return *static_cast<Derived*>(this); }
-    const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    inline StorageIndex cols() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
+    inline StorageIndex rows() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
     
     /** \brief Reports whether previous computation was successful.
       *
@@ -216,34 +217,6 @@ class CholmodBase : internal::noncopyable
       return derived();
     }
     
-    /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A.
-      *
-      * \sa compute()
-      */
-    template<typename Rhs>
-    inline const internal::solve_retval<CholmodBase, Rhs>
-    solve(const MatrixBase<Rhs>& b) const
-    {
-      eigen_assert(m_isInitialized && "LLT is not initialized.");
-      eigen_assert(rows()==b.rows()
-                && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::solve_retval<CholmodBase, Rhs>(*this, b.derived());
-    }
-    
-    /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A.
-      *
-      * \sa compute()
-      */
-    template<typename Rhs>
-    inline const internal::sparse_solve_retval<CholmodBase, Rhs>
-    solve(const SparseMatrixBase<Rhs>& b) const
-    {
-      eigen_assert(m_isInitialized && "LLT is not initialized.");
-      eigen_assert(rows()==b.rows()
-                && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::sparse_solve_retval<CholmodBase, Rhs>(*this, b.derived());
-    }
-    
     /** Performs a symbolic decomposition on the sparsity pattern of \a matrix.
       *
       * This function is particularly useful when solving for several problems having the same structure.
@@ -290,7 +263,7 @@ class CholmodBase : internal::noncopyable
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal */
     template<typename Rhs,typename Dest>
-    void _solve(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
     {
       eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
       const Index size = m_cholmodFactor->n;
@@ -304,6 +277,7 @@ class CholmodBase : internal::noncopyable
       if(!x_cd)
       {
         this->m_info = NumericalIssue;
+        return;
       }
       // TODO optimize this copy by swapping when possible (be careful with alignment, etc.)
       dest = Matrix<Scalar,Dest::RowsAtCompileTime,Dest::ColsAtCompileTime>::Map(reinterpret_cast<Scalar*>(x_cd->x),b.rows(),b.cols());
@@ -312,7 +286,7 @@ class CholmodBase : internal::noncopyable
     
     /** \internal */
     template<typename RhsScalar, int RhsOptions, typename RhsIndex, typename DestScalar, int DestOptions, typename DestIndex>
-    void _solve(const SparseMatrix<RhsScalar,RhsOptions,RhsIndex> &b, SparseMatrix<DestScalar,DestOptions,DestIndex> &dest) const
+    void _solve_impl(const SparseMatrix<RhsScalar,RhsOptions,RhsIndex> &b, SparseMatrix<DestScalar,DestOptions,DestIndex> &dest) const
     {
       eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
       const Index size = m_cholmodFactor->n;
@@ -325,6 +299,7 @@ class CholmodBase : internal::noncopyable
       if(!x_cs)
       {
         this->m_info = NumericalIssue;
+        return;
       }
       // TODO optimize this copy by swapping when possible (be careful with alignment, etc.)
       dest = viewAsEigen<DestScalar,DestOptions,DestIndex>(*x_cs);
@@ -357,7 +332,6 @@ class CholmodBase : internal::noncopyable
     cholmod_factor* m_cholmodFactor;
     RealScalar m_shiftOffset[2];
     mutable ComputationInfo m_info;
-    bool m_isInitialized;
     int m_factorizationIsOk;
     int m_analysisIsOk;
 };
@@ -395,7 +369,7 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl
     CholmodSimplicialLLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodSimplicialLLT() {}
@@ -442,7 +416,7 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp
     CholmodSimplicialLDLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodSimplicialLDLT() {}
@@ -487,7 +461,7 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper
     CholmodSupernodalLLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodSupernodalLLT() {}
@@ -534,7 +508,7 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom
     CholmodDecomposition(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodDecomposition() {}
@@ -572,36 +546,6 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom
     }
 };
 
-namespace internal {
-  
-template<typename _MatrixType, int _UpLo, typename Derived, typename Rhs>
-struct solve_retval<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-  : solve_retval_base<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-{
-  typedef CholmodBase<_MatrixType,_UpLo,Derived> Dec;
-  EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs)
-
-  template<typename Dest> void evalTo(Dest& dst) const
-  {
-    dec()._solve(rhs(),dst);
-  }
-};
-
-template<typename _MatrixType, int _UpLo, typename Derived, typename Rhs>
-struct sparse_solve_retval<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-  : sparse_solve_retval_base<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-{
-  typedef CholmodBase<_MatrixType,_UpLo,Derived> Dec;
-  EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs)
-
-  template<typename Dest> void evalTo(Dest& dst) const
-  {
-    dec()._solve(rhs(),dst);
-  }
-};
-
-} // end namespace internal
-
 } // end namespace Eigen
 
 #endif // EIGEN_CHOLMODSUPPORT_H

Eigen/src/Core/Array.h

@@ -24,6 +24,9 @@ namespace Eigen {
   * API for the %Matrix class provides easy access to linear-algebra
   * operations.
   *
+  * See documentation of class Matrix for detailed information on the template parameters
+  * storage layout.
+  * 
   * This class can be extended with the help of the plugin mechanism described on the page
   * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN.
   *
@@ -75,6 +78,21 @@ class Array
       return Base::operator=(other);
     }
 
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill()
+      */
+    /* This overload is needed because the usage of
+      *   using Base::operator=;
+      * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped
+      * the usage of 'using'. This should be done only for operator=.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const Scalar &value)
+    {
+      Base::setConstant(value);
+      return *this;
+    }
+
     /** Copies the value of the expression \a other into \c *this with automatic resizing.
       *
       * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized),
@@ -86,7 +104,7 @@ class Array
       */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Array& operator=(const ArrayBase<OtherDerived>& other)
+    EIGEN_STRONG_INLINE Array& operator=(const DenseBase<OtherDerived>& other)
     {
       return Base::_set(other);
     }
@@ -99,7 +117,7 @@ class Array
     {
       return Base::_set(other);
     }
-
+    
     /** Default constructor.
       *
       * For fixed-size matrices, does nothing.
@@ -130,6 +148,7 @@ class Array
 #endif
 
 #ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
     Array(Array&& other)
       : Base(std::move(other))
     {
@@ -137,6 +156,7 @@ class Array
       if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
         Base::_set_noalias(other);
     }
+    EIGEN_DEVICE_FUNC
     Array& operator=(Array&& other)
     {
       other.swap(*this);
@@ -144,7 +164,6 @@ class Array
     }
 #endif
 
-
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename T>
     EIGEN_DEVICE_FUNC
@@ -206,50 +225,18 @@ class Array
       m_storage.data()[3] = val3;
     }
 
-    /** Constructor copying the value of the expression \a other */
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Array(const ArrayBase<OtherDerived>& other)
-             : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
     /** Copy constructor */
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const Array& other)
-            : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
-    /** Copy constructor with in-place evaluation */
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Array(const ReturnByValue<OtherDerived>& other)
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      other.evalTo(*this);
-    }
+            : Base(other)
+    { }
 
     /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const EigenBase<OtherDerived> &other)
-      : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
-    {
-      Base::_check_template_params();
-      Base::_resize_to_match(other);
-      *this = other;
-    }
-
-    /** Override MatrixBase::swap() since for dynamic-sized matrices of same type it is enough to swap the
-      * data pointers.
-      */
-    template<typename OtherDerived>
-    void swap(ArrayBase<OtherDerived> const & other)
-    { this->_swap(other.derived()); }
+      : Base(other.derived())
+    { }
 
     EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
     EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }

Eigen/src/Core/ArrayBase.h

@@ -50,7 +50,6 @@ template<typename Derived> class ArrayBase
                 typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>::operator*;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -64,8 +63,7 @@ template<typename Derived> class ArrayBase
     using Base::MaxSizeAtCompileTime;
     using Base::IsVectorAtCompileTime;
     using Base::Flags;
-    using Base::CoeffReadCost;
-
+    
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -85,22 +83,10 @@ template<typename Derived> class ArrayBase
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-    /** \internal the plain matrix type corresponding to this expression. Note that is not necessarily
-      * exactly the return type of eval(): in the case of plain matrices, the return type of eval() is a const
-      * reference to a matrix, not a matrix! It is however guaranteed that the return type of eval() is either
-      * PlainObject or const PlainObject&.
-      */
-    typedef Array<typename internal::traits<Derived>::Scalar,
-                internal::traits<Derived>::RowsAtCompileTime,
-                internal::traits<Derived>::ColsAtCompileTime,
-                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
-                internal::traits<Derived>::MaxRowsAtCompileTime,
-                internal::traits<Derived>::MaxColsAtCompileTime
-          > PlainObject;
-
+    typedef typename Base::PlainObject PlainObject;
 
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::ArrayBase
@@ -121,8 +107,15 @@ template<typename Derived> class ArrayBase
     EIGEN_DEVICE_FUNC
     Derived& operator=(const ArrayBase& other)
     {
-      return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+      internal::call_assignment(derived(), other.derived());
+      return derived();
     }
+    
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill() */
+    EIGEN_DEVICE_FUNC
+    Derived& operator=(const Scalar &value)
+    { Base::setConstant(value); return derived(); }
 
     EIGEN_DEVICE_FUNC
     Derived& operator+=(const Scalar& scalar);
@@ -153,9 +146,9 @@ template<typename Derived> class ArrayBase
     /** \returns an \link Eigen::MatrixBase Matrix \endlink expression of this array
       * \sa MatrixBase::array() */
     EIGEN_DEVICE_FUNC
-    MatrixWrapper<Derived> matrix() { return derived(); }
+    MatrixWrapper<Derived> matrix() { return MatrixWrapper<Derived>(derived()); }
     EIGEN_DEVICE_FUNC
-    const MatrixWrapper<const Derived> matrix() const { return derived(); }
+    const MatrixWrapper<const Derived> matrix() const { return MatrixWrapper<const Derived>(derived()); }
 
 //     template<typename Dest>
 //     inline void evalTo(Dest& dst) const { dst = matrix(); }
@@ -186,8 +179,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other)
 {
-  SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar>());
   return derived();
 }
 
@@ -200,8 +192,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar>());
   return derived();
 }
 
@@ -214,8 +205,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_product_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::mul_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
@@ -228,8 +218,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator/=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::div_assign_op<Scalar>());
   return derived();
 }
 

Eigen/src/Core/ArrayWrapper.h

@@ -44,6 +44,7 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     typedef ArrayBase<ArrayWrapper> Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(ArrayWrapper)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(ArrayWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
 
     typedef typename internal::conditional<
                        internal::is_lvalue<ExpressionType>::value,
@@ -51,10 +52,10 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
+    typedef typename internal::ref_selector<ExpressionType>::type NestedExpressionType;
 
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
+    explicit EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
     EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_expression.rows(); }
@@ -148,7 +149,7 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index,Index)*/
     EIGEN_DEVICE_FUNC
-    void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); }
+    void resize(Index rows, Index cols) { m_expression.const_cast_derived().resize(rows,cols); }
 
   protected:
     NestedExpressionType m_expression;
@@ -186,6 +187,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
     typedef MatrixBase<MatrixWrapper<ExpressionType> > Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(MatrixWrapper)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(MatrixWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
 
     typedef typename internal::conditional<
                        internal::is_lvalue<ExpressionType>::value,
@@ -193,10 +195,10 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
+    typedef typename internal::ref_selector<ExpressionType>::type NestedExpressionType;
 
     EIGEN_DEVICE_FUNC
-    inline MatrixWrapper(ExpressionType& a_matrix) : m_expression(a_matrix) {}
+    explicit inline MatrixWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
     EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_expression.rows(); }
@@ -286,7 +288,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index,Index)*/
     EIGEN_DEVICE_FUNC
-    void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); }
+    void resize(Index rows, Index cols) { m_expression.const_cast_derived().resize(rows,cols); }
 
   protected:
     NestedExpressionType m_expression;

Eigen/src/Core/Assign.h

@@ -14,485 +14,6 @@
 
 namespace Eigen {
 
-namespace internal {
-
-/***************************************************************************
-* Part 1 : the logic deciding a strategy for traversal and unrolling       *
-***************************************************************************/
-
-template <typename Derived, typename OtherDerived>
-struct assign_traits
-{
-public:
-  enum {
-    DstIsAligned = Derived::Flags & AlignedBit,
-    DstHasDirectAccess = Derived::Flags & DirectAccessBit,
-    SrcIsAligned = OtherDerived::Flags & AlignedBit,
-    JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned
-  };
-
-private:
-  enum {
-    InnerSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::SizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::ColsAtCompileTime)
-              : int(Derived::RowsAtCompileTime),
-    InnerMaxSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::MaxSizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::MaxColsAtCompileTime)
-              : int(Derived::MaxRowsAtCompileTime),
-    MaxSizeAtCompileTime = Derived::SizeAtCompileTime,
-    PacketSize = packet_traits<typename Derived::Scalar>::size
-  };
-
-  enum {
-    StorageOrdersAgree = (int(Derived::IsRowMajor) == int(OtherDerived::IsRowMajor)),
-    MightVectorize = StorageOrdersAgree
-                  && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit),
-    MayInnerVectorize  = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0
-                       && int(DstIsAligned) && int(SrcIsAligned),
-    MayLinearize = StorageOrdersAgree && (int(Derived::Flags) & int(OtherDerived::Flags) & LinearAccessBit),
-    MayLinearVectorize = MightVectorize && MayLinearize && DstHasDirectAccess
-                       && (DstIsAligned || MaxSizeAtCompileTime == Dynamic),
-      /* If the destination isn't aligned, we have to do runtime checks and we don't unroll,
-         so it's only good for large enough sizes. */
-    MaySliceVectorize  = MightVectorize && DstHasDirectAccess
-                       && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=3*PacketSize)
-      /* slice vectorization can be slow, so we only want it if the slices are big, which is
-         indicated by InnerMaxSize rather than InnerSize, think of the case of a dynamic block
-         in a fixed-size matrix */
-  };
-
-public:
-  enum {
-    Traversal = int(MayInnerVectorize)  ? int(InnerVectorizedTraversal)
-              : int(MayLinearVectorize) ? int(LinearVectorizedTraversal)
-              : int(MaySliceVectorize)  ? int(SliceVectorizedTraversal)
-              : int(MayLinearize)       ? int(LinearTraversal)
-                                        : int(DefaultTraversal),
-    Vectorized = int(Traversal) == InnerVectorizedTraversal
-              || int(Traversal) == LinearVectorizedTraversal
-              || int(Traversal) == SliceVectorizedTraversal
-  };
-
-private:
-  enum {
-    UnrollingLimit      = EIGEN_UNROLLING_LIMIT * (Vectorized ? int(PacketSize) : 1),
-    MayUnrollCompletely = int(Derived::SizeAtCompileTime) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(Derived::SizeAtCompileTime) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit),
-    MayUnrollInner      = int(InnerSize) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(InnerSize) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit)
-  };
-
-public:
-  enum {
-    Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal))
-                ? (
-                    int(MayUnrollCompletely) ? int(CompleteUnrolling)
-                  : int(MayUnrollInner)      ? int(InnerUnrolling)
-                                             : int(NoUnrolling)
-                  )
-              : int(Traversal) == int(LinearVectorizedTraversal)
-                ? ( bool(MayUnrollCompletely) && bool(DstIsAligned) ? int(CompleteUnrolling) : int(NoUnrolling) )
-              : int(Traversal) == int(LinearTraversal)
-                ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) : int(NoUnrolling) )
-              : int(NoUnrolling)
-  };
-
-#ifdef EIGEN_DEBUG_ASSIGN
-  static void debug()
-  {
-    EIGEN_DEBUG_VAR(DstIsAligned)
-    EIGEN_DEBUG_VAR(SrcIsAligned)
-    EIGEN_DEBUG_VAR(JointAlignment)
-    EIGEN_DEBUG_VAR(Derived::SizeAtCompileTime)
-    EIGEN_DEBUG_VAR(OtherDerived::CoeffReadCost)
-    EIGEN_DEBUG_VAR(InnerSize)
-    EIGEN_DEBUG_VAR(InnerMaxSize)
-    EIGEN_DEBUG_VAR(PacketSize)
-    EIGEN_DEBUG_VAR(StorageOrdersAgree)
-    EIGEN_DEBUG_VAR(MightVectorize)
-    EIGEN_DEBUG_VAR(MayLinearize)
-    EIGEN_DEBUG_VAR(MayInnerVectorize)
-    EIGEN_DEBUG_VAR(MayLinearVectorize)
-    EIGEN_DEBUG_VAR(MaySliceVectorize)
-    EIGEN_DEBUG_VAR(Traversal)
-    EIGEN_DEBUG_VAR(UnrollingLimit)
-    EIGEN_DEBUG_VAR(MayUnrollCompletely)
-    EIGEN_DEBUG_VAR(MayUnrollInner)
-    EIGEN_DEBUG_VAR(Unrolling)
-  }
-#endif
-};
-
-/***************************************************************************
-* Part 2 : meta-unrollers
-***************************************************************************/
-
-/************************
-*** Default traversal ***
-************************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_DefaultTraversal_CompleteUnrolling
-{
-  enum {
-    outer = Index / Derived1::InnerSizeAtCompileTime,
-    inner = Index % Derived1::InnerSizeAtCompileTime
-  };
-
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.copyCoeffByOuterInner(outer, inner, src);
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  EIGEN_DEVICE_FUNC 
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_DefaultTraversal_InnerUnrolling
-{
-  EIGEN_DEVICE_FUNC 
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer)
-  {
-    dst.copyCoeffByOuterInner(outer, Index, src);
-    assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src, outer);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  EIGEN_DEVICE_FUNC 
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {}
-};
-
-/***********************
-*** Linear traversal ***
-***********************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_LinearTraversal_CompleteUnrolling
-{
-  EIGEN_DEVICE_FUNC 
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.copyCoeff(Index, src);
-    assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  EIGEN_DEVICE_FUNC 
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-/**************************
-*** Inner vectorization ***
-**************************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_innervec_CompleteUnrolling
-{
-  enum {
-    outer = Index / Derived1::InnerSizeAtCompileTime,
-    inner = Index % Derived1::InnerSizeAtCompileTime,
-    JointAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-  };
-
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.template copyPacketByOuterInner<Derived2, Aligned, JointAlignment>(outer, inner, src);
-    assign_innervec_CompleteUnrolling<Derived1, Derived2,
-      Index+packet_traits<typename Derived1::Scalar>::size, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_innervec_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_innervec_InnerUnrolling
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer)
-  {
-    dst.template copyPacketByOuterInner<Derived2, Aligned, Aligned>(outer, Index, src);
-    assign_innervec_InnerUnrolling<Derived1, Derived2,
-      Index+packet_traits<typename Derived1::Scalar>::size, Stop>::run(dst, src, outer);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_innervec_InnerUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {}
-};
-
-/***************************************************************************
-* Part 3 : implementation of all cases
-***************************************************************************/
-
-template<typename Derived1, typename Derived2,
-         int Traversal = assign_traits<Derived1, Derived2>::Traversal,
-         int Unrolling = assign_traits<Derived1, Derived2>::Unrolling,
-         int Version = Specialized>
-struct assign_impl;
-
-/************************
-*** Default traversal ***
-************************/
-
-template<typename Derived1, typename Derived2, int Unrolling, int Version>
-struct assign_impl<Derived1, Derived2, InvalidTraversal, Unrolling, Version>
-{
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &, const Derived2 &) { }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  EIGEN_DEVICE_FUNC 
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      for(Index inner = 0; inner < innerSize; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, CompleteUnrolling, Version>
-{
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, InnerUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  EIGEN_DEVICE_FUNC 
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, 0, Derived1::InnerSizeAtCompileTime>
-        ::run(dst, src, outer);
-  }
-};
-
-/***********************
-*** Linear traversal ***
-***********************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index size = dst.size();
-    for(Index i = 0; i < size; ++i)
-      dst.copyCoeff(i, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearTraversal, CompleteUnrolling, Version>
-{
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-/**************************
-*** Inner vectorization ***
-**************************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    const Index packetSize = packet_traits<typename Derived1::Scalar>::size;
-    for(Index outer = 0; outer < outerSize; ++outer)
-      for(Index inner = 0; inner < innerSize; inner+=packetSize)
-        dst.template copyPacketByOuterInner<Derived2, Aligned, Aligned>(outer, inner, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, CompleteUnrolling, Version>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_innervec_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, InnerUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      assign_innervec_InnerUnrolling<Derived1, Derived2, 0, Derived1::InnerSizeAtCompileTime>
-        ::run(dst, src, outer);
-  }
-};
-
-/***************************
-*** Linear vectorization ***
-***************************/
-
-template <bool IsAligned = false>
-struct unaligned_assign_impl
-{
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_STRONG_INLINE void run(const Derived&, OtherDerived&, typename Derived::Index, typename Derived::Index) {}
-};
-
-template <>
-struct unaligned_assign_impl<false>
-{
-  // MSVC must not inline this functions. If it does, it fails to optimize the
-  // packet access path.
-#ifdef _MSC_VER
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_DONT_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end)
-#else
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_STRONG_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end)
-#endif
-  {
-    for (typename Derived::Index index = start; index < end; ++index)
-      dst.copyCoeff(index, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index size = dst.size();
-    typedef packet_traits<typename Derived1::Scalar> PacketTraits;
-    enum {
-      packetSize = PacketTraits::size,
-      dstAlignment = PacketTraits::AlignedOnScalar ? Aligned : int(assign_traits<Derived1,Derived2>::DstIsAligned) ,
-      srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-    };
-    const Index alignedStart = assign_traits<Derived1,Derived2>::DstIsAligned ? 0
-                             : internal::first_aligned(&dst.coeffRef(0), size);
-    const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
-
-    unaligned_assign_impl<assign_traits<Derived1,Derived2>::DstIsAligned!=0>::run(src,dst,0,alignedStart);
-
-    for(Index index = alignedStart; index < alignedEnd; index += packetSize)
-    {
-      dst.template copyPacket<Derived2, dstAlignment, srcAlignment>(index, src);
-    }
-
-    unaligned_assign_impl<>::run(src,dst,alignedEnd,size);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, CompleteUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    enum { size = Derived1::SizeAtCompileTime,
-           packetSize = packet_traits<typename Derived1::Scalar>::size,
-           alignedSize = (size/packetSize)*packetSize };
-
-    assign_innervec_CompleteUnrolling<Derived1, Derived2, 0, alignedSize>::run(dst, src);
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, alignedSize, size>::run(dst, src);
-  }
-};
-
-/**************************
-*** Slice vectorization ***
-***************************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, SliceVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    typedef packet_traits<typename Derived1::Scalar> PacketTraits;
-    enum {
-      packetSize = PacketTraits::size,
-      alignable = PacketTraits::AlignedOnScalar,
-      dstAlignment = alignable ? Aligned : int(assign_traits<Derived1,Derived2>::DstIsAligned) ,
-      srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-    };
-    const Index packetAlignedMask = packetSize - 1;
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    const Index alignedStep = alignable ? (packetSize - dst.outerStride() % packetSize) & packetAlignedMask : 0;
-    Index alignedStart = ((!alignable) || assign_traits<Derived1,Derived2>::DstIsAligned) ? 0
-                       : internal::first_aligned(&dst.coeffRef(0,0), innerSize);
-
-    for(Index outer = 0; outer < outerSize; ++outer)
-    {
-      const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask);
-      // do the non-vectorizable part of the assignment
-      for(Index inner = 0; inner<alignedStart ; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-
-      // do the vectorizable part of the assignment
-      for(Index inner = alignedStart; inner<alignedEnd; inner+=packetSize)
-        dst.template copyPacketByOuterInner<Derived2, dstAlignment, Unaligned>(outer, inner, src);
-
-      // do the non-vectorizable part of the assignment
-      for(Index inner = alignedEnd; inner<innerSize ; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-
-      alignedStart = std::min<Index>((alignedStart+alignedStep)%packetSize, innerSize);
-    }
-  }
-};
-
-} // end namespace internal
-
-/***************************************************************************
-* Part 4 : implementation of DenseBase methods
-***************************************************************************/
-
 template<typename Derived>
 template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
@@ -506,91 +27,35 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
   EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Derived,OtherDerived)
   EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
 
-#ifdef EIGEN_TEST_EVALUATORS
-  
-#ifdef EIGEN_DEBUG_ASSIGN
-  internal::copy_using_evaluator_traits<Derived, OtherDerived>::debug();
-#endif
   eigen_assert(rows() == other.rows() && cols() == other.cols());
-  internal::call_dense_assignment_loop(derived(),other.derived());
+  internal::call_assignment_no_alias(derived(),other.derived());
   
-#else // EIGEN_TEST_EVALUATORS
-
-#ifdef EIGEN_DEBUG_ASSIGN
-  internal::assign_traits<Derived, OtherDerived>::debug();
-#endif
-  eigen_assert(rows() == other.rows() && cols() == other.cols());
-  internal::assign_impl<Derived, OtherDerived, int(SameType) ? int(internal::assign_traits<Derived, OtherDerived>::Traversal)
-                                                             : int(InvalidTraversal)>::run(derived(),other.derived());
-  
-#endif // EIGEN_TEST_EVALUATORS
-  
-#ifndef EIGEN_NO_DEBUG
-  checkTransposeAliasing(other.derived());
-#endif
   return derived();
 }
 
-namespace internal {
-
-template<typename Derived, typename OtherDerived,
-         bool EvalBeforeAssigning = (int(internal::traits<OtherDerived>::Flags) & EvalBeforeAssigningBit) != 0,
-         bool NeedToTranspose = ((int(Derived::RowsAtCompileTime) == 1 && int(OtherDerived::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(Derived::ColsAtCompileTime) == 1 && int(OtherDerived::RowsAtCompileTime) == 1))
-                              && int(Derived::SizeAtCompileTime) != 1>
-struct assign_selector;
-
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,false,false> {
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); }
-  template<typename ActualDerived, typename ActualOtherDerived>
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,true,false> {
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.eval()); }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,false,true> {
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); }
-  template<typename ActualDerived, typename ActualOtherDerived>
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { Transpose<ActualDerived> dstTrans(dst); other.evalTo(dstTrans); return dst; }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,true,true> {
-  EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); }
-};
-
-} // end namespace internal
-
 template<typename Derived>
 template<typename OtherDerived>
 EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase& other)
 {
-  return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const MatrixBase& other)
 {
-  return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
@@ -598,7 +63,8 @@ template <typename OtherDerived>
 EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
@@ -606,7 +72,8 @@ template <typename OtherDerived>
 EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const EigenBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived,false>::evalTo(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
@@ -614,7 +81,8 @@ template<typename OtherDerived>
 EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived,false>::evalTo(derived(), other.derived());
+  other.derived().evalTo(derived());
+  return derived();
 }
 
 } // end namespace Eigen

Eigen/src/Core/AssignEvaluator.h

@@ -2,7 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2011 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2011-2013 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2011-2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -24,39 +24,52 @@ namespace internal {
 
 // copy_using_evaluator_traits is based on assign_traits
 
-template <typename Derived, typename OtherDerived>
+template <typename DstEvaluator, typename SrcEvaluator, typename AssignFunc>
 struct copy_using_evaluator_traits
 {
+  typedef typename DstEvaluator::XprType Dst;
+  typedef typename Dst::Scalar DstScalar;
+  // TODO distinguish between linear traversal and inner-traversals
+  typedef typename find_best_packet<DstScalar,Dst::SizeAtCompileTime>::type PacketType; 
+  
+  enum {
+    DstFlags = DstEvaluator::Flags,
+    SrcFlags = SrcEvaluator::Flags,
+    RequiredAlignment = unpacket_traits<PacketType>::alignment
+  };
+  
 public:
   enum {
-    DstIsAligned = Derived::Flags & AlignedBit,
-    DstHasDirectAccess = Derived::Flags & DirectAccessBit,
-    SrcIsAligned = OtherDerived::Flags & AlignedBit,
-    JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned,
-    SrcEvalBeforeAssign = (evaluator_traits<OtherDerived>::HasEvalTo == 1)
+    DstAlignment = DstEvaluator::Alignment,
+    SrcAlignment = SrcEvaluator::Alignment,
+    DstHasDirectAccess = DstFlags & DirectAccessBit,
+    JointAlignment = EIGEN_PLAIN_ENUM_MIN(DstAlignment,SrcAlignment)
   };
 
 private:
   enum {
-    InnerSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::SizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::ColsAtCompileTime)
-              : int(Derived::RowsAtCompileTime),
-    InnerMaxSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::MaxSizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::MaxColsAtCompileTime)
-              : int(Derived::MaxRowsAtCompileTime),
-    MaxSizeAtCompileTime = Derived::SizeAtCompileTime,
-    PacketSize = packet_traits<typename Derived::Scalar>::size
+    InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
+              : int(DstFlags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
+              : int(Dst::RowsAtCompileTime),
+    InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime)
+              : int(DstFlags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime)
+              : int(Dst::MaxRowsAtCompileTime),
+    MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
+    PacketSize = unpacket_traits<PacketType>::size
   };
 
   enum {
-    StorageOrdersAgree = (int(Derived::IsRowMajor) == int(OtherDerived::IsRowMajor)),
+    DstIsRowMajor = DstFlags&RowMajorBit,
+    SrcIsRowMajor = SrcFlags&RowMajorBit,
+    StorageOrdersAgree = (int(DstIsRowMajor) == int(SrcIsRowMajor)),
     MightVectorize = StorageOrdersAgree
-                  && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit),
+                  && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit)
+                  && (functor_traits<AssignFunc>::PacketAccess),
     MayInnerVectorize  = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0
-                       && int(DstIsAligned) && int(SrcIsAligned),
-    MayLinearize = StorageOrdersAgree && (int(Derived::Flags) & int(OtherDerived::Flags) & LinearAccessBit),
+                       && int(JointAlignment)>=int(RequiredAlignment),
+    MayLinearize = StorageOrdersAgree && (int(DstFlags) & int(SrcFlags) & LinearAccessBit),
     MayLinearVectorize = MightVectorize && MayLinearize && DstHasDirectAccess
-                       && (DstIsAligned || MaxSizeAtCompileTime == Dynamic),
+                       && ((int(DstAlignment)>=int(RequiredAlignment)) || MaxSizeAtCompileTime == Dynamic),
       /* If the destination isn't aligned, we have to do runtime checks and we don't unroll,
          so it's only good for large enough sizes. */
     MaySliceVectorize  = MightVectorize && DstHasDirectAccess
@@ -68,8 +81,7 @@ private:
 
 public:
   enum {
-    Traversal = int(SrcEvalBeforeAssign) ? int(AllAtOnceTraversal) 
-              : int(MayInnerVectorize)   ? int(InnerVectorizedTraversal)
+    Traversal = int(MayInnerVectorize)   ? int(InnerVectorizedTraversal)
               : int(MayLinearVectorize)  ? int(LinearVectorizedTraversal)
               : int(MaySliceVectorize)   ? int(SliceVectorizedTraversal)
               : int(MayLinearize)        ? int(LinearTraversal)
@@ -82,12 +94,12 @@ public:
 private:
   enum {
     UnrollingLimit      = EIGEN_UNROLLING_LIMIT * (Vectorized ? int(PacketSize) : 1),
-    MayUnrollCompletely = int(Derived::SizeAtCompileTime) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(Derived::SizeAtCompileTime) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit),
+    MayUnrollCompletely = int(Dst::SizeAtCompileTime) != Dynamic
+                       && int(SrcEvaluator::CoeffReadCost) != Dynamic
+                       && int(Dst::SizeAtCompileTime) * int(SrcEvaluator::CoeffReadCost) <= int(UnrollingLimit),
     MayUnrollInner      = int(InnerSize) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(InnerSize) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit)
+                       && int(SrcEvaluator::CoeffReadCost) != Dynamic
+                       && int(InnerSize) * int(SrcEvaluator::CoeffReadCost) <= int(UnrollingLimit)
   };
 
 public:
@@ -99,8 +111,8 @@ public:
                                              : int(NoUnrolling)
                   )
               : int(Traversal) == int(LinearVectorizedTraversal)
-                ? ( bool(MayUnrollCompletely) && bool(DstIsAligned) ? int(CompleteUnrolling) 
-                                                                    : int(NoUnrolling) )
+                ? ( bool(MayUnrollCompletely) && (int(DstAlignment)>=int(RequiredAlignment)) ? int(CompleteUnrolling)
+                                                                                             : int(NoUnrolling) )
               : int(Traversal) == int(LinearTraversal)
                 ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) 
                                               : int(NoUnrolling) )
@@ -110,8 +122,15 @@ public:
 #ifdef EIGEN_DEBUG_ASSIGN
   static void debug()
   {
-    EIGEN_DEBUG_VAR(DstIsAligned)
-    EIGEN_DEBUG_VAR(SrcIsAligned)
+    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.unsetf(std::ios::hex);
+    EIGEN_DEBUG_VAR(DstAlignment)
+    EIGEN_DEBUG_VAR(SrcAlignment)
+    EIGEN_DEBUG_VAR(RequiredAlignment)
     EIGEN_DEBUG_VAR(JointAlignment)
     EIGEN_DEBUG_VAR(InnerSize)
     EIGEN_DEBUG_VAR(InnerMaxSize)
@@ -127,6 +146,7 @@ public:
     EIGEN_DEBUG_VAR(MayUnrollCompletely)
     EIGEN_DEBUG_VAR(MayUnrollInner)
     EIGEN_DEBUG_VAR(Unrolling)
+    std::cerr << std::endl;
   }
 #endif
 };
@@ -142,6 +162,7 @@ public:
 template<typename Kernel, int Index, int Stop>
 struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling
 {
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
   typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
   typedef typename DstEvaluatorType::XprType DstXprType;
   
@@ -150,7 +171,7 @@ struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling
     inner = Index % DstXprType::InnerSizeAtCompileTime
   };
 
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     kernel.assignCoeffByOuterInner(outer, inner);
     copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel);
@@ -160,23 +181,23 @@ struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling
 template<typename Kernel, int Stop>
 struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Stop, Stop>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel&) { }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
 };
 
-template<typename Kernel, int Index, int Stop>
+template<typename Kernel, int Index_, int Stop>
 struct copy_using_evaluator_DefaultTraversal_InnerUnrolling
 {
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel, int outer)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
   {
-    kernel.assignCoeffByOuterInner(outer, Index);
-    copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Index+1, Stop>::run(kernel, outer);
+    kernel.assignCoeffByOuterInner(outer, Index_);
+    copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Index_+1, Stop>::run(kernel, outer);
   }
 };
 
 template<typename Kernel, int Stop>
 struct copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Stop, Stop>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel&, int) { }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index) { }
 };
 
 /***********************
@@ -186,7 +207,7 @@ struct copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Stop, Stop>
 template<typename Kernel, int Index, int Stop>
 struct copy_using_evaluator_LinearTraversal_CompleteUnrolling
 {
-  static EIGEN_STRONG_INLINE void run(Kernel& kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel& kernel)
   {
     kernel.assignCoeff(Index);
     copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel);
@@ -196,7 +217,7 @@ struct copy_using_evaluator_LinearTraversal_CompleteUnrolling
 template<typename Kernel, int Stop>
 struct copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Stop, Stop>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel&) { }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
 };
 
 /**************************
@@ -206,19 +227,21 @@ struct copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Stop, Stop
 template<typename Kernel, int Index, int Stop>
 struct copy_using_evaluator_innervec_CompleteUnrolling
 {
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
   typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
   typedef typename DstEvaluatorType::XprType DstXprType;
-
+  typedef typename Kernel::PacketType PacketType;
+  
   enum {
     outer = Index / DstXprType::InnerSizeAtCompileTime,
     inner = Index % DstXprType::InnerSizeAtCompileTime,
     JointAlignment = Kernel::AssignmentTraits::JointAlignment
   };
 
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
-    kernel.template assignPacketByOuterInner<Aligned, JointAlignment>(outer, inner);
-    enum { NextIndex = Index + packet_traits<typename DstXprType::Scalar>::size };
+    kernel.template assignPacketByOuterInner<Aligned, JointAlignment, PacketType>(outer, inner);
+    enum { NextIndex = Index + unpacket_traits<PacketType>::size };
     copy_using_evaluator_innervec_CompleteUnrolling<Kernel, NextIndex, Stop>::run(kernel);
   }
 };
@@ -226,17 +249,17 @@ struct copy_using_evaluator_innervec_CompleteUnrolling
 template<typename Kernel, int Stop>
 struct copy_using_evaluator_innervec_CompleteUnrolling<Kernel, Stop, Stop>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel&) { }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
 };
 
-template<typename Kernel, int Index, int Stop>
+template<typename Kernel, int Index_, int Stop>
 struct copy_using_evaluator_innervec_InnerUnrolling
 {
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel, int outer)
+  typedef typename Kernel::PacketType PacketType;
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
   {
-    kernel.template assignPacketByOuterInner<Aligned, Aligned>(outer, Index);
-    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
-    enum { NextIndex = Index + packet_traits<typename DstXprType::Scalar>::size };
+    kernel.template assignPacketByOuterInner<Aligned, Aligned, PacketType>(outer, Index_);
+    enum { NextIndex = Index_ + unpacket_traits<PacketType>::size };
     copy_using_evaluator_innervec_InnerUnrolling<Kernel, NextIndex, Stop>::run(kernel, outer);
   }
 };
@@ -244,7 +267,7 @@ struct copy_using_evaluator_innervec_InnerUnrolling
 template<typename Kernel, int Stop>
 struct copy_using_evaluator_innervec_InnerUnrolling<Kernel, Stop, Stop>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel &, int) { }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &, Index) { }
 };
 
 /***************************************************************************
@@ -265,10 +288,8 @@ struct dense_assignment_loop;
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, DefaultTraversal, NoUnrolling>
 {
-  static void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static void run(Kernel &kernel)
   {
-    typedef typename Kernel::Index Index;
-    
     for(Index outer = 0; outer < kernel.outerSize(); ++outer) {
       for(Index inner = 0; inner < kernel.innerSize(); ++inner) {
         kernel.assignCoeffByOuterInner(outer, inner);
@@ -280,7 +301,7 @@ struct dense_assignment_loop<Kernel, DefaultTraversal, NoUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, DefaultTraversal, CompleteUnrolling>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
     copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
@@ -290,8 +311,8 @@ struct dense_assignment_loop<Kernel, DefaultTraversal, CompleteUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, DefaultTraversal, InnerUnrolling>
 {
-  typedef typename Kernel::Index Index;
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  typedef typename Kernel::StorageIndex StorageIndex;
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
 
@@ -314,7 +335,7 @@ struct unaligned_dense_assignment_loop
 {
   // if IsAligned = true, then do nothing
   template <typename Kernel>
-  static EIGEN_STRONG_INLINE void run(Kernel&, typename Kernel::Index, typename Kernel::Index) {}
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index, Index) {}
 };
 
 template <>
@@ -323,19 +344,19 @@ struct unaligned_dense_assignment_loop<false>
   // MSVC must not inline this functions. If it does, it fails to optimize the
   // packet access path.
   // FIXME check which version exhibits this issue
-#ifdef _MSC_VER
+#if EIGEN_COMP_MSVC
   template <typename Kernel>
   static EIGEN_DONT_INLINE void run(Kernel &kernel,
-                                    typename Kernel::Index start,
-                                    typename Kernel::Index end)
+                                    Index start,
+                                    Index end)
 #else
   template <typename Kernel>
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel,
-                                      typename Kernel::Index start,
-                                      typename Kernel::Index end)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel,
+                                      Index start,
+                                      Index end)
 #endif
   {
-    for (typename Kernel::Index index = start; index < end; ++index)
+    for (Index index = start; index < end; ++index)
       kernel.assignCoeff(index);
   }
 };
@@ -343,25 +364,26 @@ struct unaligned_dense_assignment_loop<false>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, NoUnrolling>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
-    typedef typename Kernel::Index Index;
-
     const Index size = kernel.size();
-    typedef packet_traits<typename Kernel::Scalar> PacketTraits;
+    typedef typename Kernel::Scalar Scalar;
+    typedef typename Kernel::PacketType PacketType;
     enum {
-      packetSize = PacketTraits::size,
-      dstIsAligned = int(Kernel::AssignmentTraits::DstIsAligned),
-      dstAlignment = PacketTraits::AlignedOnScalar ? Aligned : dstIsAligned,
+      requestedAlignment = Kernel::AssignmentTraits::RequiredAlignment,
+      packetSize = unpacket_traits<PacketType>::size,
+      dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment),
+      dstAlignment = packet_traits<Scalar>::AlignedOnScalar ? int(requestedAlignment)
+                                                            : int(Kernel::AssignmentTraits::DstAlignment),
       srcAlignment = Kernel::AssignmentTraits::JointAlignment
     };
-    const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned(&kernel.dstEvaluator().coeffRef(0), size);
+    const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned<requestedAlignment>(&kernel.dstEvaluator().coeffRef(0), size);
     const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
 
     unaligned_dense_assignment_loop<dstIsAligned!=0>::run(kernel, 0, alignedStart);
 
     for(Index index = alignedStart; index < alignedEnd; index += packetSize)
-      kernel.template assignPacket<dstAlignment, srcAlignment>(index);
+      kernel.template assignPacket<dstAlignment, srcAlignment, PacketType>(index);
 
     unaligned_dense_assignment_loop<>::run(kernel, alignedEnd, size);
   }
@@ -370,8 +392,8 @@ struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, NoUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, CompleteUnrolling>
 {
-  typedef typename Kernel::Index Index;
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  typedef typename Kernel::StorageIndex StorageIndex;
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
     
@@ -391,23 +413,22 @@ struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, CompleteUnrollin
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, NoUnrolling>
 {
-  static inline void run(Kernel &kernel)
+  typedef typename Kernel::PacketType PacketType;
+  EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel)
   {
-    typedef typename Kernel::Index Index;
-
     const Index innerSize = kernel.innerSize();
     const Index outerSize = kernel.outerSize();
-    const Index packetSize = packet_traits<typename Kernel::Scalar>::size;
+    const Index packetSize = unpacket_traits<PacketType>::size;
     for(Index outer = 0; outer < outerSize; ++outer)
       for(Index inner = 0; inner < innerSize; inner+=packetSize)
-        kernel.template assignPacketByOuterInner<Aligned, Aligned>(outer, inner);
+        kernel.template assignPacketByOuterInner<Aligned, Aligned, PacketType>(outer, inner);
   }
 };
 
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, CompleteUnrolling>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
     copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
@@ -417,8 +438,8 @@ struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, CompleteUnrolling
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling>
 {
-  typedef typename Kernel::Index Index;
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  typedef typename Kernel::StorageIndex StorageIndex;
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
     const Index outerSize = kernel.outerSize();
@@ -434,9 +455,8 @@ struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, LinearTraversal, NoUnrolling>
 {
-  static inline void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel)
   {
-    typedef typename Kernel::Index Index;
     const Index size = kernel.size();
     for(Index i = 0; i < size; ++i)
       kernel.assignCoeff(i);
@@ -446,7 +466,7 @@ struct dense_assignment_loop<Kernel, LinearTraversal, NoUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, LinearTraversal, CompleteUnrolling>
 {
-  static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
     copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
@@ -460,21 +480,29 @@ struct dense_assignment_loop<Kernel, LinearTraversal, CompleteUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
 {
-  static inline void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel)
   {
-    typedef typename Kernel::Index Index;
-    typedef packet_traits<typename Kernel::Scalar> PacketTraits;
+    typedef typename Kernel::Scalar Scalar;
+    typedef typename Kernel::PacketType PacketType;
     enum {
-      packetSize = PacketTraits::size,
-      alignable = PacketTraits::AlignedOnScalar,
-      dstAlignment = alignable ? Aligned : int(Kernel::AssignmentTraits::DstIsAligned)
+      packetSize = unpacket_traits<PacketType>::size,
+      requestedAlignment = int(Kernel::AssignmentTraits::RequiredAlignment),
+      alignable = packet_traits<Scalar>::AlignedOnScalar || int(Kernel::AssignmentTraits::DstAlignment)>=sizeof(Scalar),
+      dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment),
+      dstAlignment = alignable ? int(requestedAlignment)
+                               : int(Kernel::AssignmentTraits::DstAlignment)
     };
+    const Scalar *dst_ptr = &kernel.dstEvaluator().coeffRef(0,0);
+    if((!bool(dstIsAligned)) && (size_t(dst_ptr) % sizeof(Scalar))>0)
+    {
+      // the pointer is not aligend-on scalar, so alignment is not possible
+      return dense_assignment_loop<Kernel,DefaultTraversal,NoUnrolling>::run(kernel);
+    }
     const Index packetAlignedMask = packetSize - 1;
     const Index innerSize = kernel.innerSize();
     const Index outerSize = kernel.outerSize();
     const Index alignedStep = alignable ? (packetSize - kernel.outerStride() % packetSize) & packetAlignedMask : 0;
-    Index alignedStart = ((!alignable) || Kernel::AssignmentTraits::DstIsAligned) ? 0
-                       : internal::first_aligned(&kernel.dstEvaluator().coeffRef(0,0), innerSize);
+    Index alignedStart = ((!alignable) || bool(dstIsAligned)) ? 0 : internal::first_aligned<requestedAlignment>(dst_ptr, innerSize);
 
     for(Index outer = 0; outer < outerSize; ++outer)
     {
@@ -485,7 +513,7 @@ struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
 
       // do the vectorizable part of the assignment
       for(Index inner = alignedStart; inner<alignedEnd; inner+=packetSize)
-        kernel.template assignPacketByOuterInner<dstAlignment, Unaligned>(outer, inner);
+        kernel.template assignPacketByOuterInner<dstAlignment, Unaligned, PacketType>(outer, inner);
 
       // do the non-vectorizable part of the assignment
       for(Index inner = alignedEnd; inner<innerSize ; ++inner)
@@ -496,25 +524,8 @@ struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
   }
 };
 
-/****************************
-*** All-at-once traversal ***
-****************************/
-
-// TODO: this 'AllAtOnceTraversal' should be dropped or caught earlier (Gael)
-// Indeed, what to do with the kernel's functor??
-template<typename Kernel>
-struct dense_assignment_loop<Kernel, AllAtOnceTraversal, NoUnrolling>
-{
-  static inline void run(Kernel & kernel)
-  {
-    // Evaluate rhs in temporary to prevent aliasing problems in a = a * a;
-    // TODO: Do not pass the xpr object to evalTo() (Jitse)
-    kernel.srcEvaluator().evalTo(kernel.dstEvaluator(), kernel.dstExpression());
-  }
-};
-
 /***************************************************************************
-* Part 4 : Generic Assignment routine
+* Part 4 : Generic dense assignment kernel
 ***************************************************************************/
 
 // This class generalize the assignment of a coefficient (or packet) from one dense evaluator
@@ -523,7 +534,7 @@ struct dense_assignment_loop<Kernel, AllAtOnceTraversal, NoUnrolling>
 // This abstraction level permits to keep the evaluation loops as simple and as generic as possible.
 // One can customize the assignment using this generic dense_assignment_kernel with different
 // functors, or by completely overloading it, by-passing a functor.
-template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor>
+template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version = Specialized>
 class generic_dense_assignment_kernel
 {
 protected:
@@ -534,36 +545,46 @@ public:
   typedef DstEvaluatorTypeT DstEvaluatorType;
   typedef SrcEvaluatorTypeT SrcEvaluatorType;
   typedef typename DstEvaluatorType::Scalar Scalar;
-  typedef typename DstEvaluatorType::Index Index;
-  typedef copy_using_evaluator_traits<DstXprType, SrcXprType> AssignmentTraits;
+  typedef typename DstEvaluatorType::StorageIndex StorageIndex;
+  typedef copy_using_evaluator_traits<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor> AssignmentTraits;
+  typedef typename AssignmentTraits::PacketType PacketType;
   
   
-  generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+  EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
     : m_dst(dst), m_src(src), m_functor(func), m_dstExpr(dstExpr)
-  {}
+  {
+    #ifdef EIGEN_DEBUG_ASSIGN
+    AssignmentTraits::debug();
+    #endif
+  }
   
-  Index size() const        { return m_dstExpr.size(); }
-  Index innerSize() const   { return m_dstExpr.innerSize(); }
-  Index outerSize() const   { return m_dstExpr.outerSize(); }
-  Index outerStride() const { return m_dstExpr.outerStride(); }
+  EIGEN_DEVICE_FUNC Index size() const        { return m_dstExpr.size(); }
+  EIGEN_DEVICE_FUNC Index innerSize() const   { return m_dstExpr.innerSize(); }
+  EIGEN_DEVICE_FUNC Index outerSize() const   { return m_dstExpr.outerSize(); }
+  EIGEN_DEVICE_FUNC Index rows() const        { return m_dstExpr.rows(); }
+  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:
-  DstXprType& dstExpression() const { return m_dstExpr; }
+  EIGEN_DEVICE_FUNC DstXprType& dstExpression() const { return m_dstExpr; }
   
-  DstEvaluatorType& dstEvaluator() { return m_dst; }
-  const SrcEvaluatorType& srcEvaluator() const { return m_src; }
+  EIGEN_DEVICE_FUNC DstEvaluatorType& dstEvaluator() { return m_dst; }
+  EIGEN_DEVICE_FUNC const SrcEvaluatorType& srcEvaluator() const { return m_src; }
   
-  void assignCoeff(Index row, Index col)
+  /// Assign src(row,col) to dst(row,col) through the assignment functor.
+  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
   {
     m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col));
   }
   
-  void assignCoeff(Index index)
+  /// \sa assignCoeff(Index,Index)
+  EIGEN_DEVICE_FUNC void assignCoeff(Index index)
   {
     m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index));
   }
   
-  void assignCoeffByOuterInner(Index outer, Index inner)
+  /// \sa assignCoeff(Index,Index)
+  EIGEN_DEVICE_FUNC void assignCoeffByOuterInner(Index outer, Index inner)
   {
     Index row = rowIndexByOuterInner(outer, inner); 
     Index col = colIndexByOuterInner(outer, inner); 
@@ -571,41 +592,41 @@ public:
   }
   
   
-  template<int StoreMode, int LoadMode>
-  void assignPacket(Index row, Index col)
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC void assignPacket(Index row, Index col)
   {
-    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(row,col), m_src.template packet<LoadMode>(row,col));
+    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(row,col), m_src.template packet<LoadMode,PacketType>(row,col));
   }
   
-  template<int StoreMode, int LoadMode>
-  void assignPacket(Index index)
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC void assignPacket(Index index)
   {
-    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(index), m_src.template packet<LoadMode>(index));
+    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(index), m_src.template packet<LoadMode,PacketType>(index));
   }
   
-  template<int StoreMode, int LoadMode>
-  void assignPacketByOuterInner(Index outer, Index inner)
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC void assignPacketByOuterInner(Index outer, Index inner)
   {
     Index row = rowIndexByOuterInner(outer, inner); 
     Index col = colIndexByOuterInner(outer, inner);
-    assignPacket<StoreMode,LoadMode>(row, col);
+    assignPacket<StoreMode,LoadMode,PacketType>(row, col);
   }
   
-  static Index rowIndexByOuterInner(Index outer, Index inner)
+  EIGEN_DEVICE_FUNC static Index rowIndexByOuterInner(Index outer, Index inner)
   {
     typedef typename DstEvaluatorType::ExpressionTraits Traits;
     return int(Traits::RowsAtCompileTime) == 1 ? 0
       : int(Traits::ColsAtCompileTime) == 1 ? inner
-      : int(Traits::Flags)&RowMajorBit ? outer
+      : int(DstEvaluatorType::Flags)&RowMajorBit ? outer
       : inner;
   }
 
-  static Index colIndexByOuterInner(Index outer, Index inner)
+  EIGEN_DEVICE_FUNC static Index colIndexByOuterInner(Index outer, Index inner)
   {
     typedef typename DstEvaluatorType::ExpressionTraits Traits;
     return int(Traits::ColsAtCompileTime) == 1 ? 0
       : int(Traits::RowsAtCompileTime) == 1 ? inner
-      : int(Traits::Flags)&RowMajorBit ? inner
+      : int(DstEvaluatorType::Flags)&RowMajorBit ? inner
       : outer;
   }
   
@@ -617,17 +638,17 @@ protected:
   DstXprType& m_dstExpr;
 };
 
+/***************************************************************************
+* Part 5 : Entry point for dense rectangular assignment
+***************************************************************************/
+
 template<typename DstXprType, typename SrcXprType, typename Functor>
-void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func)
+EIGEN_DEVICE_FUNC void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func)
 {
-#ifdef EIGEN_DEBUG_ASSIGN
-  // TODO these traits should be computed from information provided by the evaluators
-  internal::copy_using_evaluator_traits<DstXprType, SrcXprType>::debug();
-#endif
   eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
   
-  typedef typename evaluator<DstXprType>::type DstEvaluatorType;
-  typedef typename evaluator<SrcXprType>::type SrcEvaluatorType;
+  typedef evaluator<DstXprType> DstEvaluatorType;
+  typedef evaluator<SrcXprType> SrcEvaluatorType;
 
   DstEvaluatorType dstEvaluator(dst);
   SrcEvaluatorType srcEvaluator(src);
@@ -639,201 +660,166 @@ void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, co
 }
 
 template<typename DstXprType, typename SrcXprType>
-void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src)
+EIGEN_DEVICE_FUNC void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src)
 {
   call_dense_assignment_loop(dst, src, internal::assign_op<typename DstXprType::Scalar>());
 }
 
 /***************************************************************************
-* Part 5 : Entry points
+* Part 6 : Generic assignment
 ***************************************************************************/
 
-// Based on DenseBase::LazyAssign()
-// The following functions are just for testing and they are meant to be moved to operator= and the likes.
+// Based on the respective shapes of the destination and source,
+// the class AssignmentKind determine the kind of assignment mechanism.
+// AssignmentKind must define a Kind typedef.
+template<typename DstShape, typename SrcShape> struct AssignmentKind;
 
-template<typename DstXprType, template <typename> class StorageBase, typename SrcXprType>
-EIGEN_STRONG_INLINE
-const DstXprType& copy_using_evaluator(const NoAlias<DstXprType, StorageBase>& dst, 
-                                       const EigenBase<SrcXprType>& src)
-{
-  return noalias_copy_using_evaluator(dst.expression(), src.derived(), internal::assign_op<typename DstXprType::Scalar>());
-}
+// Assignement kind defined in this file:
+struct Dense2Dense {};
+struct EigenBase2EigenBase {};
 
-template<typename XprType, int AssumeAliasing = evaluator_traits<XprType>::AssumeAliasing>
-struct AddEvalIfAssumingAliasing;
-
-template<typename XprType>
-struct AddEvalIfAssumingAliasing<XprType, 0>
-{
-  static const XprType& run(const XprType& xpr) 
-  {
-    return xpr;
-  }
-};
-
-template<typename XprType>
-struct AddEvalIfAssumingAliasing<XprType, 1>
-{
-  static const EvalToTemp<XprType> run(const XprType& xpr)
-  {
-    return EvalToTemp<XprType>(xpr);
-  }
-};
-
-template<typename DstXprType, typename SrcXprType, typename Functor>
-EIGEN_STRONG_INLINE
-const DstXprType& copy_using_evaluator(const EigenBase<DstXprType>& dst, const EigenBase<SrcXprType>& src, const Functor &func)
-{
-  return noalias_copy_using_evaluator(dst.const_cast_derived(), 
-                                      AddEvalIfAssumingAliasing<SrcXprType>::run(src.derived()),
-                                      func
-                                     );
-}
-
-// this mimics operator=
-template<typename DstXprType, typename SrcXprType>
-EIGEN_STRONG_INLINE
-const DstXprType& copy_using_evaluator(const EigenBase<DstXprType>& dst, const EigenBase<SrcXprType>& src)
-{
-  return copy_using_evaluator(dst.const_cast_derived(), src.derived(), internal::assign_op<typename DstXprType::Scalar>());
-}
-
-template<typename DstXprType, typename SrcXprType, typename Functor>
-EIGEN_STRONG_INLINE
-const DstXprType& noalias_copy_using_evaluator(const PlainObjectBase<DstXprType>& dst, const EigenBase<SrcXprType>& src, const Functor &func)
-{
-#ifdef EIGEN_DEBUG_ASSIGN
-  internal::copy_using_evaluator_traits<DstXprType, SrcXprType>::debug();
-#endif
-#ifdef EIGEN_NO_AUTOMATIC_RESIZING
-  eigen_assert((dst.size()==0 || (IsVectorAtCompileTime ? (dst.size() == src.size())
-                                                        : (dst.rows() == src.rows() && dst.cols() == src.cols())))
-              && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined");
-#else
-  dst.const_cast_derived().resizeLike(src.derived());
-#endif
-  call_dense_assignment_loop(dst.const_cast_derived(), src.derived(), func);
-  return dst.derived();
-}
-
-template<typename DstXprType, typename SrcXprType, typename Functor>
-EIGEN_STRONG_INLINE
-const DstXprType& noalias_copy_using_evaluator(const EigenBase<DstXprType>& dst, const EigenBase<SrcXprType>& src, const Functor &func)
-{
-  call_dense_assignment_loop(dst.const_cast_derived(), src.derived(), func);
-  return dst.derived();
-}
-
-// Based on DenseBase::swap()
-// TODO: Check whether we need to do something special for swapping two
-//       Arrays or Matrices. (Jitse)
-
-// Overload default assignPacket behavior for swapping them
-template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT>
-class swap_kernel : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar> >
-{
-  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar> > Base;
-  typedef typename DstEvaluatorTypeT::PacketScalar PacketScalar;
-  using Base::m_dst;
-  using Base::m_src;
-  using Base::m_functor;
-  
-public:
-  typedef typename Base::Scalar Scalar;
-  typedef typename Base::Index Index;
-  typedef typename Base::DstXprType DstXprType;
-  
-  swap_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, DstXprType& dstExpr)
-    : Base(dst, src, swap_assign_op<Scalar>(), dstExpr)
-  {}
-  
-  template<int StoreMode, int LoadMode>
-  void assignPacket(Index row, Index col)
-  {
-    m_functor.template swapPacket<StoreMode,LoadMode,PacketScalar>(&m_dst.coeffRef(row,col), &const_cast<SrcEvaluatorTypeT&>(m_src).coeffRef(row,col));
-  }
-  
-  template<int StoreMode, int LoadMode>
-  void assignPacket(Index index)
-  {
-    m_functor.template swapPacket<StoreMode,LoadMode,PacketScalar>(&m_dst.coeffRef(index), &const_cast<SrcEvaluatorTypeT&>(m_src).coeffRef(index));
-  }
-  
-  // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael)
-  template<int StoreMode, int LoadMode>
-  void assignPacketByOuterInner(Index outer, Index inner)
-  {
-    Index row = Base::rowIndexByOuterInner(outer, inner); 
-    Index col = Base::colIndexByOuterInner(outer, inner);
-    assignPacket<StoreMode,LoadMode>(row, col);
-  }
-};
-  
-template<typename DstXprType, typename SrcXprType>
-void swap_using_evaluator(const DstXprType& dst, const SrcXprType& src)
-{
-  // TODO there is too much redundancy with call_dense_assignment_loop
-  
-  eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
-  
-  typedef typename evaluator<DstXprType>::type DstEvaluatorType;
-  typedef typename evaluator<SrcXprType>::type SrcEvaluatorType;
-
-  DstEvaluatorType dstEvaluator(dst);
-  SrcEvaluatorType srcEvaluator(src);
+template<typename,typename> struct AssignmentKind { typedef EigenBase2EigenBase Kind; };
+template<> struct AssignmentKind<DenseShape,DenseShape> { typedef Dense2Dense Kind; };
     
-  typedef swap_kernel<DstEvaluatorType,SrcEvaluatorType> Kernel;
-  Kernel kernel(dstEvaluator, srcEvaluator, dst.const_cast_derived());
+// This is the main assignment class
+template< typename DstXprType, typename SrcXprType, typename Functor,
+          typename Kind = typename AssignmentKind< typename evaluator_traits<DstXprType>::Shape , typename evaluator_traits<SrcXprType>::Shape >::Kind,
+          typename Scalar = typename DstXprType::Scalar>
+struct Assignment;
+
+
+// The only purpose of this call_assignment() function is to deal with noalias() / AssumeAliasing and automatic transposition.
+// Indeed, I (Gael) think that this concept of AssumeAliasing was a mistake, and it makes thing quite complicated.
+// So this intermediate function removes everything related to AssumeAliasing such that Assignment
+// does not has to bother about these annoying details.
+
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC void call_assignment(Dst& dst, const Src& src)
+{
+  call_assignment(dst, src, internal::assign_op<typename Dst::Scalar>());
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC void call_assignment(const Dst& dst, const Src& src)
+{
+  call_assignment(dst, src, internal::assign_op<typename Dst::Scalar>());
+}
+                     
+// Deal with AssumeAliasing
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if<evaluator_traits<Src>::AssumeAliasing==1, void*>::type = 0)
+{
+  typename plain_matrix_type<Src>::type tmp(src);
+  call_assignment_no_alias(dst, tmp, func);
+}
+
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if<evaluator_traits<Src>::AssumeAliasing==0, void*>::type = 0)
+{
+  call_assignment_no_alias(dst, src, func);
+}
+
+// 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);
+}
+
+
+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
+  };
+
+  Index dstRows = NeedToTranspose ? src.cols() : src.rows();
+  Index dstCols = NeedToTranspose ? src.rows() : src.cols();
+  if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+    dst.resize(dstRows, dstCols);
   
-  dense_assignment_loop<Kernel>::run(kernel);
-}
+  typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst>::type ActualDstTypeCleaned;
+  typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst&>::type ActualDstType;
+  ActualDstType actualDst(dst);
+  
+  // 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)
 
-// Based on MatrixBase::operator+= (in CwiseBinaryOp.h)
-template<typename DstXprType, typename SrcXprType>
-void add_assign_using_evaluator(const MatrixBase<DstXprType>& dst, const MatrixBase<SrcXprType>& 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);
+  
+  Assignment<ActualDstTypeCleaned,Src,Func>::run(actualDst, src, func);
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC void call_assignment_no_alias(Dst& dst, const Src& src)
 {
-  typedef typename DstXprType::Scalar Scalar;
-  copy_using_evaluator(dst.derived(), src.derived(), add_assign_op<Scalar>());
+  call_assignment_no_alias(dst, src, internal::assign_op<typename Dst::Scalar>());
 }
 
-// Based on ArrayBase::operator+=
-template<typename DstXprType, typename SrcXprType>
-void add_assign_using_evaluator(const ArrayBase<DstXprType>& dst, const ArrayBase<SrcXprType>& src)
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src, const Func& func)
 {
-  typedef typename DstXprType::Scalar Scalar;
-  copy_using_evaluator(dst.derived(), src.derived(), add_assign_op<Scalar>());
+  Index dstRows = src.rows();
+  Index dstCols = src.cols();
+  if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+    dst.resize(dstRows, dstCols);
+  
+  // TODO check whether this is the right place to perform these checks:
+  EIGEN_STATIC_ASSERT_LVALUE(Dst)
+  EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Dst,Src)
+  
+  Assignment<Dst,Src,Func>::run(dst, src, func);
 }
-
-// TODO: Add add_assign_using_evaluator for EigenBase ? (Jitse)
-
-template<typename DstXprType, typename SrcXprType>
-void subtract_assign_using_evaluator(const MatrixBase<DstXprType>& dst, const MatrixBase<SrcXprType>& src)
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src)
 {
-  typedef typename DstXprType::Scalar Scalar;
-  copy_using_evaluator(dst.derived(), src.derived(), sub_assign_op<Scalar>());
+  call_assignment_no_alias_no_transpose(dst, src, internal::assign_op<typename Dst::Scalar>());
 }
 
-template<typename DstXprType, typename SrcXprType>
-void subtract_assign_using_evaluator(const ArrayBase<DstXprType>& dst, const ArrayBase<SrcXprType>& src)
+// forward declaration
+template<typename Dst, typename Src> void check_for_aliasing(const Dst &dst, const Src &src);
+
+// Generic Dense to Dense assignment
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar>
+struct Assignment<DstXprType, SrcXprType, Functor, Dense2Dense, Scalar>
 {
-  typedef typename DstXprType::Scalar Scalar;
-  copy_using_evaluator(dst.derived(), src.derived(), sub_assign_op<Scalar>());
-}
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    
+#ifndef EIGEN_NO_DEBUG
+    internal::check_for_aliasing(dst, src);
+#endif
+    
+    call_dense_assignment_loop(dst, src, func);
+  }
+};
 
-template<typename DstXprType, typename SrcXprType>
-void multiply_assign_using_evaluator(const ArrayBase<DstXprType>& dst, const ArrayBase<SrcXprType>& src)
+// Generic assignment through evalTo.
+// TODO: not sure we have to keep that one, but it helps porting current code to new evaluator mechanism.
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar>
+struct Assignment<DstXprType, SrcXprType, Functor, EigenBase2EigenBase, Scalar>
 {
-  typedef typename DstXprType::Scalar Scalar;
-  copy_using_evaluator(dst.derived(), src.derived(), mul_assign_op<Scalar>());
-}
-
-template<typename DstXprType, typename SrcXprType>
-void divide_assign_using_evaluator(const ArrayBase<DstXprType>& dst, const ArrayBase<SrcXprType>& src)
-{
-  typedef typename DstXprType::Scalar Scalar;
-  copy_using_evaluator(dst.derived(), src.derived(), div_assign_op<Scalar>());
-}
-
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar> &/*func*/)
+  {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.evalTo(dst);
+  }
+};
 
 } // namespace internal
 

Eigen/src/Core/Assign_MKL.h

@@ -1,6 +1,7 @@
 /*
  Copyright (c) 2011, Intel Corporation. All rights reserved.
-
+ Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+ 
  Redistribution and use in source and binary forms, with or without modification,
  are permitted provided that the following conditions are met:
 
@@ -37,17 +38,13 @@ namespace Eigen {
 
 namespace internal {
 
-template<typename Op> struct vml_call
-{ enum { IsSupported = 0 }; };
-
-template<typename Dst, typename Src, typename UnaryOp>
+template<typename Dst, typename Src>
 class vml_assign_traits
 {
   private:
     enum {
       DstHasDirectAccess = Dst::Flags & DirectAccessBit,
       SrcHasDirectAccess = Src::Flags & DirectAccessBit,
-
       StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)),
       InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
                 : int(Dst::Flags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
@@ -57,166 +54,118 @@ class vml_assign_traits
                     : int(Dst::MaxRowsAtCompileTime),
       MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
 
-      MightEnableVml =  vml_call<UnaryOp>::IsSupported && StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess
-                     && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
+      MightEnableVml = StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
       MightLinearize = MightEnableVml && (int(Dst::Flags) & int(Src::Flags) & LinearAccessBit),
       VmlSize = MightLinearize ? MaxSizeAtCompileTime : InnerMaxSize,
-      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD,
-      MayEnableVml = MightEnableVml && LargeEnough,
-      MayLinearize = MayEnableVml && MightLinearize
+      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD
     };
   public:
     enum {
-      Traversal = MayLinearize ? LinearVectorizedTraversal
-                : MayEnableVml ? InnerVectorizedTraversal
-                : DefaultTraversal
+      EnableVml = MightEnableVml && LargeEnough,
+      Traversal = MightLinearize ? LinearTraversal : DefaultTraversal
     };
 };
 
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling,
-         int VmlTraversal = vml_assign_traits<Derived1, Derived2, UnaryOp>::Traversal >
-struct vml_assign_impl
-  : assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>
-{
-};
-
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
-struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, InnerVectorizedTraversal>
-{
-  typedef typename Derived1::Scalar Scalar;
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
-  {
-    // in case we want to (or have to) skip VML at runtime we can call:
-    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer) {
-      const Scalar *src_ptr = src.IsRowMajor ?  &(src.nestedExpression().coeffRef(outer,0)) :
-                                                &(src.nestedExpression().coeffRef(0, outer));
-      Scalar *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));
-      vml_call<UnaryOp>::run(src.functor(), innerSize, src_ptr, dst_ptr );
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
-struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, LinearVectorizedTraversal>
-{
-  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
-  {
-    // in case we want to (or have to) skip VML at runtime we can call:
-    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
-    vml_call<UnaryOp>::run(src.functor(), dst.size(), src.nestedExpression().data(), dst.data() );
-  }
-};
-
-// Macroses
-
-#define EIGEN_MKL_VML_SPECIALIZE_ASSIGN(TRAVERSAL,UNROLLING) \
-  template<typename Derived1, typename Derived2, typename UnaryOp> \
-  struct assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>, TRAVERSAL, UNROLLING, Specialized>  {  \
-    static inline void run(Derived1 &dst, const Eigen::CwiseUnaryOp<UnaryOp, Derived2> &src) { \
-      vml_assign_impl<Derived1,Derived2,UnaryOp,TRAVERSAL,UNROLLING>::run(dst, src); \
-    } \
-  };
-
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,InnerUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,InnerUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(SliceVectorizedTraversal,NoUnrolling)
-
-
+#define EIGEN_PP_EXPAND(ARG) ARG
 #if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1)
-#define  EIGEN_MKL_VML_MODE VML_HA
+#define EIGEN_VMLMODE_EXPAND_LA , VML_HA
 #else
-#define  EIGEN_MKL_VML_MODE VML_LA
+#define EIGEN_VMLMODE_EXPAND_LA , VML_LA
 #endif
 
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)     \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/,        \
-                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
-      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst);                           \
-    }                                                                            \
+#define EIGEN_VMLMODE_EXPAND__ 
+
+#define EIGEN_VMLMODE_PREFIX_LA vm
+#define EIGEN_VMLMODE_PREFIX__  v
+#define EIGEN_VMLMODE_PREFIX(VMLMODE) EIGEN_CAT(EIGEN_VMLMODE_PREFIX_,VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested>                                                                         \
+  struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE>,             \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml,EIGENTYPE>::type> {    \
+    typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType;                                            \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE> &/*func*/) {                             \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                       \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal) {                                              \
+        VMLOP(dst.size(), (const VMLTYPE*)src.nestedExpression().data(),                                                        \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                           \
+      } else {                                                                                                                  \
+        const Index outerSize = dst.outerSize();                                                                                \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                      \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) :                             \
+                                                      &(src.nestedExpression().coeffRef(0, outer));                             \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                           \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr,                                                                      \
+                (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                             \
+        }                                                                                                                       \
+      }                                                                                                                         \
+    }                                                                                                                           \
+  };                                                                                                                            \
+
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),s##VMLOP), float, float, VMLMODE)           \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),d##VMLOP), double, double, VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),c##VMLOP), scomplex, MKL_Complex8, VMLMODE) \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),z##VMLOP), dcomplex, MKL_Complex16, VMLMODE)
+  
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP, VMLMODE)                                                              \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                               \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)
+
+  
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sin,   Sin,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(asin,  Asin,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sinh,  Sinh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cos,   Cos,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(acos,  Acos,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cosh,  Cosh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tan,   Tan,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(atan,  Atan,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tanh,  Tanh,  LA)
+// EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,   Abs,    _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(exp,   Exp,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log,   Ln,    LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log10, Log10, LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sqrt,  Sqrt,  _)
+
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr,   _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(arg, Arg,      _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(round, Round,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(floor, Floor,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil,  Ceil,   _)
+
+#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested>                                                                       \
+  struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE>,           \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml,EIGENTYPE>::type> {  \
+    typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType;                                          \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE> &/*func*/) {                           \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                     \
+      VMLTYPE exponent = reinterpret_cast<const VMLTYPE&>(src.functor().m_exponent);                                          \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal)                                              \
+      {                                                                                                                       \
+        VMLOP( dst.size(), (const VMLTYPE*)src.nestedExpression().data(), exponent,                                           \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                         \
+      } else {                                                                                                                \
+        const Index outerSize = dst.outerSize();                                                                              \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                    \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) :                           \
+                                                      &(src.nestedExpression().coeffRef(0, outer));                           \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                         \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr, exponent,                                                          \
+                 (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                          \
+        }                                                                                                                     \
+      }                                                                                                                       \
+    }                                                                                                                         \
   };
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)  \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/,        \
-                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
-      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
-      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst, vmlMode);                  \
-    }                                                                            \
-  };
-
-#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)       \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& func,        \
-                          int size, const EIGENTYPE* src, EIGENTYPE* dst) {      \
-      EIGENTYPE exponent = func.m_exponent;                                      \
-      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
-      VMLOP(&size, (const VMLTYPE*)src, (const VMLTYPE*)&exponent,               \
-                        (VMLTYPE*)dst, &vmlMode);                                \
-    }                                                                            \
-  };
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                   \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vs##VMLOP, float, float)             \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vd##VMLOP, double, double)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)                \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vc##VMLOP, scomplex, MKL_Complex8)   \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vz##VMLOP, dcomplex, MKL_Complex16)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP)                        \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                         \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)
-
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vms##VMLOP, float, float)         \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmd##VMLOP, double, double)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)             \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmc##VMLOP, scomplex, MKL_Complex8)  \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmz##VMLOP, dcomplex, MKL_Complex16)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(EIGENOP, VMLOP)                     \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                      \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)
-
-
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sin,  Sin)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(asin, Asin)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(cos,  Cos)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(acos, Acos)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(tan,  Tan)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(atan,  Atan)
-//EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,  Abs)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(exp,  Exp)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(log,  Ln)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sqrt, Sqrt)
-
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr)
-
-// The vm*powx functions are not avaibale in the windows version of MKL.
-#ifndef _WIN32
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmspowx_, float, float)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdpowx_, double, double)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcpowx_, scomplex, MKL_Complex8)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzpowx_, dcomplex, MKL_Complex16)
-#endif
+  
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmsPowx, float,    float,         LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdPowx, double,   double,        LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcPowx, scomplex, MKL_Complex8,  LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzPowx, dcomplex, MKL_Complex16, LA)
 
 } // end namespace internal
 

Eigen/src/Core/BandMatrix.h

@@ -32,7 +32,7 @@ class BandMatrixBase : public EigenBase<Derived>
     };
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime> DenseMatrixType;
-    typedef typename DenseMatrixType::Index Index;
+    typedef typename DenseMatrixType::StorageIndex StorageIndex;
     typedef typename internal::traits<Derived>::CoefficientsType CoefficientsType;
     typedef EigenBase<Derived> Base;
 
@@ -179,7 +179,7 @@ struct traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
 {
   typedef _Scalar Scalar;
   typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef Eigen::Index StorageIndex;
   enum {
     CoeffReadCost = NumTraits<Scalar>::ReadCost,
     RowsAtCompileTime = _Rows,
@@ -201,10 +201,10 @@ class BandMatrix : public BandMatrixBase<BandMatrix<_Scalar,Rows,Cols,Supers,Sub
   public:
 
     typedef typename internal::traits<BandMatrix>::Scalar Scalar;
-    typedef typename internal::traits<BandMatrix>::Index Index;
+    typedef typename internal::traits<BandMatrix>::StorageIndex StorageIndex;
     typedef typename internal::traits<BandMatrix>::CoefficientsType CoefficientsType;
 
-    inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs)
+    explicit inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs)
       : m_coeffs(1+supers+subs,cols),
         m_rows(rows), m_supers(supers), m_subs(subs)
     {
@@ -241,7 +241,7 @@ struct traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Opt
 {
   typedef typename _CoefficientsType::Scalar Scalar;
   typedef typename _CoefficientsType::StorageKind StorageKind;
-  typedef typename _CoefficientsType::Index Index;
+  typedef typename _CoefficientsType::StorageIndex StorageIndex;
   enum {
     CoeffReadCost = internal::traits<_CoefficientsType>::CoeffReadCost,
     RowsAtCompileTime = _Rows,
@@ -264,9 +264,9 @@ class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<_CoefficientsT
 
     typedef typename internal::traits<BandMatrixWrapper>::Scalar Scalar;
     typedef typename internal::traits<BandMatrixWrapper>::CoefficientsType CoefficientsType;
-    typedef typename internal::traits<BandMatrixWrapper>::Index Index;
+    typedef typename internal::traits<BandMatrixWrapper>::StorageIndex StorageIndex;
 
-    inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs)
+    explicit inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs)
       : m_coeffs(coeffs),
         m_rows(rows), m_supers(supers), m_subs(subs)
     {
@@ -312,9 +312,9 @@ template<typename Scalar, int Size, int Options>
 class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor>
 {
     typedef BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor> Base;
-    typedef typename Base::Index Index;
+    typedef typename Base::StorageIndex StorageIndex;
   public:
-    TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {}
+    explicit TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {}
 
     inline typename Base::template DiagonalIntReturnType<1>::Type super()
     { return Base::template diagonal<1>(); }
@@ -327,6 +327,25 @@ class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint
   protected:
 };
 
+
+struct BandShape {};
+
+template<typename _Scalar, int _Rows, int _Cols, int _Supers, int _Subs, int _Options>
+struct evaluator_traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options>
+struct evaluator_traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<> struct AssignmentKind<DenseShape,BandShape> { typedef EigenBase2EigenBase Kind; };
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/Block.h

@@ -55,7 +55,7 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
   typedef typename traits<XprType>::Scalar Scalar;
   typedef typename traits<XprType>::StorageKind StorageKind;
   typedef typename traits<XprType>::XprKind XprKind;
-  typedef typename nested<XprType>::type XprTypeNested;
+  typedef typename ref_selector<XprType>::type XprTypeNested;
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   enum{
     MatrixRows = traits<XprType>::RowsAtCompileTime,
@@ -68,6 +68,7 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
     MaxColsAtCompileTime = BlockCols==0 ? 0
                          : ColsAtCompileTime != Dynamic ? int(ColsAtCompileTime)
                          : int(traits<XprType>::MaxColsAtCompileTime),
+
     XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0,
     IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
                : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
@@ -80,18 +81,16 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
     OuterStrideAtCompileTime = HasSameStorageOrderAsXprType
                              ? int(outer_stride_at_compile_time<XprType>::ret)
                              : int(inner_stride_at_compile_time<XprType>::ret),
-    MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits<Scalar>::size) == 0)
-                       && (InnerStrideAtCompileTime == 1)
-                        ? PacketAccessBit : 0,
-    MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) ? AlignedBit : 0,
-    FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (traits<XprType>::Flags&LinearAccessBit))) ? LinearAccessBit : 0,
+
+    // FIXME, this traits is rather specialized for dense object and it needs to be cleaned further
     FlagsLvalueBit = is_lvalue<XprType>::value ? LvalueBit : 0,
     FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0,
-    Flags0 = traits<XprType>::Flags & ( (HereditaryBits & ~RowMajorBit) |
-                                        DirectAccessBit |
-                                        MaskPacketAccessBit |
-                                        MaskAlignedBit),
-    Flags = Flags0 | FlagsLinearAccessBit | FlagsLvalueBit | FlagsRowMajorBit
+    Flags = (traits<XprType>::Flags & (DirectAccessBit | (InnerPanel?CompressedAccessBit:0))) | FlagsLvalueBit | FlagsRowMajorBit,
+    // FIXME DirectAccessBit should not be handled by expressions
+    // 
+    // Alignment is needed by MapBase's assertions
+    // We can sefely set it to false here. Internal alignment errors will be detected by an eigen_internal_assert in the respective evaluator
+    Alignment = 0
   };
 };
 
@@ -111,6 +110,8 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class
     typedef Impl Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Block)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Block)
+    
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
   
     /** Column or Row constructor
       */
@@ -125,26 +126,26 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class
     /** Fixed-size constructor
       */
     EIGEN_DEVICE_FUNC
-    inline Block(XprType& xpr, Index a_startRow, Index a_startCol)
-      : Impl(xpr, a_startRow, a_startCol)
+    inline Block(XprType& xpr, Index startRow, Index startCol)
+      : Impl(xpr, startRow, startCol)
     {
       EIGEN_STATIC_ASSERT(RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic,THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE)
-      eigen_assert(a_startRow >= 0 && BlockRows >= 1 && a_startRow + BlockRows <= xpr.rows()
-             && a_startCol >= 0 && BlockCols >= 1 && a_startCol + BlockCols <= xpr.cols());
+      eigen_assert(startRow >= 0 && BlockRows >= 1 && startRow + BlockRows <= xpr.rows()
+             && startCol >= 0 && BlockCols >= 1 && startCol + BlockCols <= xpr.cols());
     }
 
     /** Dynamic-size constructor
       */
     EIGEN_DEVICE_FUNC
     inline Block(XprType& xpr,
-          Index a_startRow, Index a_startCol,
+          Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : Impl(xpr, a_startRow, a_startCol, blockRows, blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols)
     {
       eigen_assert((RowsAtCompileTime==Dynamic || RowsAtCompileTime==blockRows)
           && (ColsAtCompileTime==Dynamic || ColsAtCompileTime==blockCols));
-      eigen_assert(a_startRow >= 0 && blockRows >= 0 && a_startRow  <= xpr.rows() - blockRows
-          && a_startCol >= 0 && blockCols >= 0 && a_startCol <= xpr.cols() - blockCols);
+      eigen_assert(startRow >= 0 && blockRows >= 0 && startRow  <= xpr.rows() - blockRows
+          && startCol >= 0 && blockCols >= 0 && startCol <= xpr.cols() - blockCols);
     }
 };
          
@@ -155,15 +156,15 @@ class BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, Dense>
   : public internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel>
 {
     typedef internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel> Impl;
-    typedef typename XprType::Index Index;
+    typedef typename XprType::StorageIndex StorageIndex;
   public:
     typedef Impl Base;
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl)
     EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {}
-    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol) : Impl(xpr, a_startRow, a_startCol) {}
+    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index startRow, Index startCol) : Impl(xpr, startRow, startCol) {}
     EIGEN_DEVICE_FUNC
-    inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol, Index blockRows, Index blockCols)
-      : Impl(xpr, a_startRow, a_startCol, blockRows, blockCols) {}
+    inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols) {}
 };
 
 namespace internal {
@@ -179,7 +180,7 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
     EIGEN_DENSE_PUBLIC_INTERFACE(BlockType)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl_dense)
 
-    class InnerIterator;
+    // class InnerIterator; // FIXME apparently never used
 
     /** Column or Row constructor
       */
@@ -199,8 +200,8 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
     /** Fixed-size constructor
       */
     EIGEN_DEVICE_FUNC
-    inline BlockImpl_dense(XprType& xpr, Index a_startRow, Index a_startCol)
-      : m_xpr(xpr), m_startRow(a_startRow), m_startCol(a_startCol),
+    inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
                     m_blockRows(BlockRows), m_blockCols(BlockCols)
     {}
 
@@ -208,9 +209,9 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
       */
     EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr,
-          Index a_startRow, Index a_startCol,
+          Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : m_xpr(xpr), m_startRow(a_startRow), m_startCol(a_startCol),
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
                     m_blockRows(blockRows), m_blockCols(blockCols)
     {}
 
@@ -307,13 +308,13 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
     }
       
     EIGEN_DEVICE_FUNC
-    Index startRow() const
+    StorageIndex startRow() const
     { 
       return m_startRow.value(); 
     }
       
     EIGEN_DEVICE_FUNC
-    Index startCol() const 
+    StorageIndex startCol() const
     { 
       return m_startCol.value(); 
     }
@@ -321,10 +322,10 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
   protected:
 
     const typename XprType::Nested m_xpr;
-    const internal::variable_if_dynamic<Index, XprType::RowsAtCompileTime == 1 ? 0 : Dynamic> m_startRow;
-    const internal::variable_if_dynamic<Index, XprType::ColsAtCompileTime == 1 ? 0 : Dynamic> m_startCol;
-    const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_blockRows;
-    const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_blockCols;
+    const internal::variable_if_dynamic<StorageIndex, XprType::RowsAtCompileTime == 1 ? 0 : Dynamic> m_startRow;
+    const internal::variable_if_dynamic<StorageIndex, XprType::ColsAtCompileTime == 1 ? 0 : Dynamic> m_startCol;
+    const internal::variable_if_dynamic<StorageIndex, RowsAtCompileTime> m_blockRows;
+    const internal::variable_if_dynamic<StorageIndex, ColsAtCompileTime> m_blockCols;
 };
 
 /** \internal Internal implementation of dense Blocks in the direct access case.*/
@@ -333,6 +334,9 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
   : public MapBase<Block<XprType, BlockRows, BlockCols, InnerPanel> >
 {
     typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
+    enum {
+      XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0
+    };
   public:
 
     typedef MapBase<BlockType> Base;
@@ -343,9 +347,8 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
       */
     EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index i)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(
-              (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0,
-              (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0)),
+      : Base(xpr.data() + i * (    ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && (!XprTypeIsRowMajor)) 
+                                || ((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && ( XprTypeIsRowMajor)) ? xpr.innerStride() : xpr.outerStride()),
              BlockRows==1 ? 1 : xpr.rows(),
              BlockCols==1 ? 1 : xpr.cols()),
         m_xpr(xpr)
@@ -357,7 +360,8 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
       */
     EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol))), m_xpr(xpr)
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol)),
+        m_xpr(xpr)
     {
       init();
     }
@@ -368,7 +372,7 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
     inline BlockImpl_dense(XprType& xpr,
           Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol)), blockRows, blockCols),
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol), blockRows, blockCols),
         m_xpr(xpr)
     {
       init();

Eigen/src/Core/BooleanRedux.h

@@ -17,9 +17,10 @@ namespace internal {
 template<typename Derived, int UnrollCount>
 struct all_unroller
 {
+  typedef typename Derived::ExpressionTraits Traits;
   enum {
-    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+    col = (UnrollCount-1) / Traits::RowsAtCompileTime,
+    row = (UnrollCount-1) % Traits::RowsAtCompileTime
   };
 
   static inline bool run(const Derived &mat)
@@ -43,11 +44,12 @@ struct all_unroller<Derived, Dynamic>
 template<typename Derived, int UnrollCount>
 struct any_unroller
 {
+  typedef typename Derived::ExpressionTraits Traits;
   enum {
-    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+    col = (UnrollCount-1) / Traits::RowsAtCompileTime,
+    row = (UnrollCount-1) % Traits::RowsAtCompileTime
   };
-
+  
   static inline bool run(const Derived &mat)
   {
     return any_unroller<Derived, UnrollCount-1>::run(mat) || mat.coeff(row, col);
@@ -78,19 +80,21 @@ struct any_unroller<Derived, Dynamic>
 template<typename Derived>
 inline bool DenseBase<Derived>::all() const
 {
+  typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && CoeffReadCost != Dynamic
+          && Evaluator::CoeffReadCost != Dynamic
           && NumTraits<Scalar>::AddCost != Dynamic
-          && SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
+  Evaluator evaluator(derived());
   if(unroll)
-    return internal::all_unroller<Derived, unroll ? int(SizeAtCompileTime) : Dynamic>::run(derived());
+    return internal::all_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator);
   else
   {
     for(Index j = 0; j < cols(); ++j)
       for(Index i = 0; i < rows(); ++i)
-        if (!coeff(i, j)) return false;
+        if (!evaluator.coeff(i, j)) return false;
     return true;
   }
 }
@@ -102,19 +106,21 @@ inline bool DenseBase<Derived>::all() const
 template<typename Derived>
 inline bool DenseBase<Derived>::any() const
 {
+  typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && CoeffReadCost != Dynamic
+          && Evaluator::CoeffReadCost != Dynamic
           && NumTraits<Scalar>::AddCost != Dynamic
-          && SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
+  Evaluator evaluator(derived());
   if(unroll)
-    return internal::any_unroller<Derived, unroll ? int(SizeAtCompileTime) : Dynamic>::run(derived());
+    return internal::any_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator);
   else
   {
     for(Index j = 0; j < cols(); ++j)
       for(Index i = 0; i < rows(); ++i)
-        if (coeff(i, j)) return true;
+        if (evaluator.coeff(i, j)) return true;
     return false;
   }
 }
@@ -124,7 +130,7 @@ inline bool DenseBase<Derived>::any() const
   * \sa all(), any()
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::Index DenseBase<Derived>::count() const
+inline Eigen::Index DenseBase<Derived>::count() const
 {
   return derived().template cast<bool>().template cast<Index>().sum();
 }

Eigen/src/Core/CommaInitializer.h

@@ -28,7 +28,6 @@ template<typename XprType>
 struct CommaInitializer
 {
   typedef typename XprType::Scalar Scalar;
-  typedef typename XprType::Index Index;
 
   EIGEN_DEVICE_FUNC
   inline CommaInitializer(XprType& xpr, const Scalar& s)
@@ -106,6 +105,9 @@ 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)
+#endif
   {
     eigen_assert((m_row+m_currentBlockRows) == m_xpr.rows()
          && m_col == m_xpr.cols()

Eigen/src/Core/CoreIterators.h

@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 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
@@ -15,47 +15,113 @@ namespace Eigen {
 /* This file contains the respective InnerIterator definition of the expressions defined in Eigen/Core
  */
 
-/** \ingroup SparseCore_Module
-  * \class InnerIterator
-  * \brief An InnerIterator allows to loop over the element of a sparse (or dense) matrix or expression
-  *
-  * todo
+namespace internal {
+
+template<typename XprType, typename EvaluatorKind>
+class inner_iterator_selector;
+
+}
+
+/** \class InnerIterator
+  * \brief An InnerIterator allows to loop over the element of any matrix expression.
+  * 
+  * \warning To be used with care because an evaluator is constructed every time an InnerIterator iterator is constructed.
+  * 
+  * TODO: add a usage example
   */
-
-// generic version for dense matrix and expressions
-template<typename Derived> class DenseBase<Derived>::InnerIterator
+template<typename XprType>
+class InnerIterator
 {
-  protected:
-    typedef typename Derived::Scalar Scalar;
-    typedef typename Derived::Index Index;
-
-    enum { IsRowMajor = (Derived::Flags&RowMajorBit)==RowMajorBit };
-  public:
-    EIGEN_STRONG_INLINE InnerIterator(const Derived& expr, Index outer)
-      : m_expression(expr), m_inner(0), m_outer(outer), m_end(expr.innerSize())
-    {}
-
-    EIGEN_STRONG_INLINE Scalar value() const
-    {
-      return (IsRowMajor) ? m_expression.coeff(m_outer, m_inner)
-                          : m_expression.coeff(m_inner, m_outer);
-    }
-
-    EIGEN_STRONG_INLINE InnerIterator& operator++() { m_inner++; return *this; }
-
-    EIGEN_STRONG_INLINE Index index() const { return m_inner; }
-    inline Index row() const { return IsRowMajor ? m_outer : index(); }
-    inline Index col() const { return IsRowMajor ? index() : m_outer; }
-
-    EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
-
-  protected:
-    const Derived& m_expression;
-    Index m_inner;
-    const Index m_outer;
-    const Index m_end;
+protected:
+  typedef internal::inner_iterator_selector<XprType, typename internal::evaluator_traits<XprType>::Kind> IteratorType;
+  typedef internal::evaluator<XprType> EvaluatorType;
+  typedef typename internal::traits<XprType>::Scalar Scalar;
+public:
+  /** Construct an iterator over the \a outerId -th row or column of \a xpr */
+  InnerIterator(const XprType &xpr, const Index &outerId)
+    : m_eval(xpr), m_iter(m_eval, outerId, xpr.innerSize())
+  {}
+  
+  /// \returns the value of the current coefficient.
+  EIGEN_STRONG_INLINE Scalar value() const          { return m_iter.value(); }
+  /** Increment the iterator \c *this to the next non-zero coefficient.
+    * Explicit zeros are not skipped over. To skip explicit zeros, see class SparseView
+    */
+  EIGEN_STRONG_INLINE InnerIterator& operator++()   { m_iter.operator++(); return *this; }
+  /// \returns the column or row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index index() const           { return m_iter.index(); }
+  /// \returns the row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index row() const             { return m_iter.row(); }
+  /// \returns the column index of the current coefficient.
+  EIGEN_STRONG_INLINE Index col() const             { return m_iter.col(); }
+  /// \returns \c true if the iterator \c *this still references a valid coefficient.
+  EIGEN_STRONG_INLINE operator bool() const         { return m_iter; }
+  
+protected:
+  EvaluatorType m_eval;
+  IteratorType m_iter;
+private:
+  // If you get here, then you're not using the right InnerIterator type, e.g.:
+  //   SparseMatrix<double,RowMajor> A;
+  //   SparseMatrix<double>::InnerIterator it(A,0);
+  template<typename T> InnerIterator(const EigenBase<T>&,Index outer);
 };
 
+namespace internal {
+
+// Generic inner iterator implementation for dense objects
+template<typename XprType>
+class inner_iterator_selector<XprType, IndexBased>
+{
+protected:
+  typedef evaluator<XprType> EvaluatorType;
+  typedef typename traits<XprType>::Scalar Scalar;
+  enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit };
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &innerSize)
+    : m_eval(eval), m_inner(0), m_outer(outerId), m_end(innerSize)
+  {}
+
+  EIGEN_STRONG_INLINE Scalar value() const
+  {
+    return (IsRowMajor) ? m_eval.coeff(m_outer, m_inner)
+                        : m_eval.coeff(m_inner, m_outer);
+  }
+
+  EIGEN_STRONG_INLINE inner_iterator_selector& operator++() { m_inner++; return *this; }
+
+  EIGEN_STRONG_INLINE Index index() const { return m_inner; }
+  inline Index row() const { return IsRowMajor ? m_outer : index(); }
+  inline Index col() const { return IsRowMajor ? index() : m_outer; }
+
+  EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
+
+protected:
+  const EvaluatorType& m_eval;
+  Index m_inner;
+  const Index m_outer;
+  const Index m_end;
+};
+
+// For iterator-based evaluator, inner-iterator is already implemented as
+// evaluator<>::InnerIterator
+template<typename XprType>
+class inner_iterator_selector<XprType, IteratorBased>
+ : public evaluator<XprType>::InnerIterator
+{
+protected:
+  typedef typename evaluator<XprType>::InnerIterator Base;
+  typedef evaluator<XprType> EvaluatorType;
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &/*innerSize*/)
+    : Base(eval, outerId)
+  {}  
+};
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_COREITERATORS_H

Eigen/src/Core/CwiseBinaryOp.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) 2008-2014 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
@@ -56,69 +56,47 @@ struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
                        typename Rhs::Scalar
                      )
                    >::type Scalar;
-  typedef typename promote_storage_type<typename traits<Lhs>::StorageKind,
-                                           typename traits<Rhs>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<Lhs>::Index,
-                                         typename traits<Rhs>::Index>::type Index;
+  typedef typename cwise_promote_storage_type<typename traits<Lhs>::StorageKind,
+                                              typename traits<Rhs>::StorageKind,
+                                              BinaryOp>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<Lhs>::StorageIndex,
+                                      typename traits<Rhs>::StorageIndex>::type StorageIndex;
   typedef typename Lhs::Nested LhsNested;
   typedef typename Rhs::Nested RhsNested;
   typedef typename remove_reference<LhsNested>::type _LhsNested;
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   enum {
-    LhsCoeffReadCost = _LhsNested::CoeffReadCost,
-    RhsCoeffReadCost = _RhsNested::CoeffReadCost,
-    LhsFlags = _LhsNested::Flags,
-    RhsFlags = _RhsNested::Flags,
-    SameType = is_same<typename _LhsNested::Scalar,typename _RhsNested::Scalar>::value,
-    StorageOrdersAgree = (int(Lhs::Flags)&RowMajorBit)==(int(Rhs::Flags)&RowMajorBit),
-    Flags0 = (int(LhsFlags) | int(RhsFlags)) & (
-        HereditaryBits
-      | (int(LhsFlags) & int(RhsFlags) &
-           ( AlignedBit
-           | (StorageOrdersAgree ? LinearAccessBit : 0)
-           | (functor_traits<BinaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0)
-           )
-        )
-     ),
-    Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit),
-    CoeffReadCost = LhsCoeffReadCost + RhsCoeffReadCost + functor_traits<BinaryOp>::Cost
+    Flags = _LhsNested::Flags & RowMajorBit
   };
 };
 } // end namespace internal
 
-// we require Lhs and Rhs to have the same scalar type. Currently there is no example of a binary functor
-// that would take two operands of different types. If there were such an example, then this check should be
-// moved to the BinaryOp functors, on a per-case basis. This would however require a change in the BinaryOp functors, as
-// currently they take only one typename Scalar template parameter.
-// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths.
-// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to
-// add together a float matrix and a double matrix.
-#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)), \
-    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
 template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
 class CwiseBinaryOpImpl;
 
-template<typename BinaryOp, typename Lhs, typename Rhs>
-class CwiseBinaryOp : internal::no_assignment_operator,
+template<typename BinaryOp, typename LhsType, typename RhsType>
+class CwiseBinaryOp : 
   public CwiseBinaryOpImpl<
-          BinaryOp, Lhs, Rhs,
-          typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                           typename internal::traits<Rhs>::StorageKind>::ret>
+          BinaryOp, LhsType, RhsType,
+          typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                        typename internal::traits<RhsType>::StorageKind,
+                                                        BinaryOp>::ret>,
+  internal::no_assignment_operator
 {
   public:
+    
+    typedef typename internal::remove_all<LhsType>::type Lhs;
+    typedef typename internal::remove_all<RhsType>::type Rhs;
 
     typedef typename CwiseBinaryOpImpl<
-        BinaryOp, Lhs, Rhs,
-        typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                         typename internal::traits<Rhs>::StorageKind>::ret>::Base Base;
+        BinaryOp, LhsType, RhsType,
+        typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                      typename internal::traits<Rhs>::StorageKind,
+                                                      BinaryOp>::ret>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseBinaryOp)
 
-    typedef typename internal::nested<Lhs>::type LhsNested;
-    typedef typename internal::nested<Rhs>::type RhsNested;
+    typedef typename internal::ref_selector<LhsType>::type LhsNested;
+    typedef typename internal::ref_selector<RhsType>::type RhsNested;
     typedef typename internal::remove_reference<LhsNested>::type _LhsNested;
     typedef typename internal::remove_reference<RhsNested>::type _RhsNested;
 
@@ -165,43 +143,13 @@ class CwiseBinaryOp : internal::no_assignment_operator,
     const BinaryOp m_functor;
 };
 
-template<typename BinaryOp, typename Lhs, typename Rhs>
-class CwiseBinaryOpImpl<BinaryOp, Lhs, Rhs, Dense>
-  : public internal::dense_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
+// Generic API dispatcher
+template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
+class CwiseBinaryOpImpl
+  : public internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
 {
-    typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> Derived;
-  public:
-
-    typedef typename internal::dense_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE( Derived )
-
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
-    {
-      return derived().functor()(derived().lhs().coeff(rowId, colId),
-                                 derived().rhs().coeff(rowId, colId));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().functor().packetOp(derived().lhs().template packet<LoadMode>(rowId, colId),
-                                          derived().rhs().template packet<LoadMode>(rowId, colId));
-    }
-
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return derived().functor()(derived().lhs().coeff(index),
-                                 derived().rhs().coeff(index));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return derived().functor().packetOp(derived().lhs().template packet<LoadMode>(index),
-                                          derived().rhs().template packet<LoadMode>(index));
-    }
+public:
+  typedef typename internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base;
 };
 
 /** replaces \c *this by \c *this - \a other.
@@ -213,8 +161,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other)
 {
-  SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar>());
   return derived();
 }
 
@@ -227,8 +174,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator+=(const MatrixBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar>());
   return derived();
 }
 

Eigen/src/Core/CwiseNullaryOp.h

@@ -35,19 +35,13 @@ template<typename NullaryOp, typename PlainObjectType>
 struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectType>
 {
   enum {
-    Flags = (traits<PlainObjectType>::Flags
-      & (  HereditaryBits
-         | (functor_has_linear_access<NullaryOp>::ret ? LinearAccessBit : 0)
-         | (functor_traits<NullaryOp>::PacketAccess ? PacketAccessBit : 0)))
-      | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit),
-    CoeffReadCost = functor_traits<NullaryOp>::Cost
+    Flags = traits<PlainObjectType>::Flags & RowMajorBit
   };
 };
 }
 
 template<typename NullaryOp, typename PlainObjectType>
-class CwiseNullaryOp : internal::no_assignment_operator,
-  public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type
+class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type, internal::no_assignment_operator
 {
   public:
 
@@ -55,13 +49,13 @@ class CwiseNullaryOp : internal::no_assignment_operator,
     EIGEN_DENSE_PUBLIC_INTERFACE(CwiseNullaryOp)
 
     EIGEN_DEVICE_FUNC
-    CwiseNullaryOp(Index nbRows, Index nbCols, const NullaryOp& func = NullaryOp())
-      : m_rows(nbRows), m_cols(nbCols), m_functor(func)
+    CwiseNullaryOp(Index rows, Index cols, const NullaryOp& func = NullaryOp())
+      : m_rows(rows), m_cols(cols), m_functor(func)
     {
-      eigen_assert(nbRows >= 0
-            && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == nbRows)
-            &&  nbCols >= 0
-            && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == nbCols));
+      eigen_assert(rows >= 0
+            && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+            &&  cols >= 0
+            && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols));
     }
 
     EIGEN_DEVICE_FUNC
@@ -119,10 +113,10 @@ class CwiseNullaryOp : internal::no_assignment_operator,
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func)
 {
-  return CwiseNullaryOp<CustomNullaryOp, Derived>(rows, cols, func);
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(rows, cols, func);
 }
 
 /** \returns an expression of a matrix defined by a custom functor \a func
@@ -145,12 +139,12 @@ DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& f
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, Derived>(1, size, func);
-  else return CwiseNullaryOp<CustomNullaryOp, Derived>(size, 1, func);
+  if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, PlainObject>(1, size, func);
+  else return CwiseNullaryOp<CustomNullaryOp, PlainObject>(size, 1, func);
 }
 
 /** \returns an expression of a matrix defined by a custom functor \a func
@@ -164,19 +158,19 @@ DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
 {
-  return CwiseNullaryOp<CustomNullaryOp, Derived>(RowsAtCompileTime, ColsAtCompileTime, func);
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(RowsAtCompileTime, ColsAtCompileTime, func);
 }
 
 /** \returns an expression of a constant matrix of value \a value
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this DenseBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
-  * it is redundant to pass \a nbRows and \a nbCols as arguments, so Zero() should be used
+  * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used
   * instead.
   *
   * The template parameter \a CustomNullaryOp is the type of the functor.
@@ -185,9 +179,9 @@ DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Constant(Index nbRows, Index nbCols, const Scalar& value)
+DenseBase<Derived>::Constant(Index rows, Index cols, const Scalar& value)
 {
-  return DenseBase<Derived>::NullaryExpr(nbRows, nbCols, internal::scalar_constant_op<Scalar>(value));
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_constant_op<Scalar>(value));
 }
 
 /** \returns an expression of a constant matrix of value \a value
@@ -251,7 +245,7 @@ EIGEN_STRONG_INLINE const typename DenseBase<Derived>::SequentialLinSpacedReturn
 DenseBase<Derived>::LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,false>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar,false>(low,high,size));
 }
 
 /**
@@ -264,7 +258,7 @@ DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& hig
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
-  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,false>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar,false>(low,high,Derived::SizeAtCompileTime));
 }
 
 /**
@@ -285,7 +279,7 @@ EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedRetu
 DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,true>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar,true>(low,high,size));
 }
 
 /**
@@ -298,7 +292,7 @@ DenseBase<Derived>::LinSpaced(const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
-  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,true>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar,true>(low,high,Derived::SizeAtCompileTime));
 }
 
 /** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */
@@ -306,9 +300,10 @@ template<typename Derived>
 bool DenseBase<Derived>::isApproxToConstant
 (const Scalar& val, const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
     for(Index i = 0; i < rows(); ++i)
-      if(!internal::isApprox(this->coeff(i, j), val, prec))
+      if(!internal::isApprox(self.coeff(i, j), val, prec))
         return false;
   return true;
 }
@@ -362,8 +357,8 @@ PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
 
 /** Resizes to the given size, and sets all coefficients in this expression to the given \a value.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   * \param val the value to which all coefficients are set
   *
   * Example: \include Matrix_setConstant_int_int.cpp
@@ -373,9 +368,9 @@ PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setConstant(Index nbRows, Index nbCols, const Scalar& val)
+PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(val);
 }
 
@@ -396,7 +391,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,false>(low,high,newSize));
+  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,PacketScalar,false>(low,high,newSize));
 }
 
 /**
@@ -434,9 +429,9 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low,
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Zero(Index nbRows, Index nbCols)
+DenseBase<Derived>::Zero(Index rows, Index cols)
 {
-  return Constant(nbRows, nbCols, Scalar(0));
+  return Constant(rows, cols, Scalar(0));
 }
 
 /** \returns an expression of a zero vector.
@@ -490,9 +485,10 @@ DenseBase<Derived>::Zero()
 template<typename Derived>
 bool DenseBase<Derived>::isZero(const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
     for(Index i = 0; i < rows(); ++i)
-      if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<Scalar>(1), prec))
+      if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<Scalar>(1), prec))
         return false;
   return true;
 }
@@ -529,8 +525,8 @@ PlainObjectBase<Derived>::setZero(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to zero.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setZero_int_int.cpp
   * Output: \verbinclude Matrix_setZero_int_int.out
@@ -539,9 +535,9 @@ PlainObjectBase<Derived>::setZero(Index newSize)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
+PlainObjectBase<Derived>::setZero(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(Scalar(0));
 }
 
@@ -549,7 +545,7 @@ PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
 
 /** \returns an expression of a matrix where all coefficients equal one.
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
@@ -563,9 +559,9 @@ PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Ones(Index nbRows, Index nbCols)
+DenseBase<Derived>::Ones(Index rows, Index cols)
 {
-  return Constant(nbRows, nbCols, Scalar(1));
+  return Constant(rows, cols, Scalar(1));
 }
 
 /** \returns an expression of a vector where all coefficients equal one.
@@ -655,8 +651,8 @@ PlainObjectBase<Derived>::setOnes(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to one.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setOnes_int_int.cpp
   * Output: \verbinclude Matrix_setOnes_int_int.out
@@ -665,9 +661,9 @@ PlainObjectBase<Derived>::setOnes(Index newSize)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
+PlainObjectBase<Derived>::setOnes(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(Scalar(1));
 }
 
@@ -675,7 +671,7 @@ PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
 
 /** \returns an expression of the identity matrix (not necessarily square).
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
@@ -689,9 +685,9 @@ PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
-MatrixBase<Derived>::Identity(Index nbRows, Index nbCols)
+MatrixBase<Derived>::Identity(Index rows, Index cols)
 {
-  return DenseBase<Derived>::NullaryExpr(nbRows, nbCols, internal::scalar_identity_op<Scalar>());
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_identity_op<Scalar>());
 }
 
 /** \returns an expression of the identity matrix (not necessarily square).
@@ -725,18 +721,19 @@ template<typename Derived>
 bool MatrixBase<Derived>::isIdentity
 (const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
   {
     for(Index i = 0; i < rows(); ++i)
     {
       if(i == j)
       {
-        if(!internal::isApprox(this->coeff(i, j), static_cast<Scalar>(1), prec))
+        if(!internal::isApprox(self.coeff(i, j), static_cast<Scalar>(1), prec))
           return false;
       }
       else
       {
-        if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<RealScalar>(1), prec))
+        if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<RealScalar>(1), prec))
           return false;
       }
     }
@@ -759,7 +756,6 @@ struct setIdentity_impl
 template<typename Derived>
 struct setIdentity_impl<Derived, true>
 {
-  typedef typename Derived::Index Index;
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& run(Derived& m)
   {
@@ -787,8 +783,8 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
 
 /** \brief Resizes to the given size, and writes the identity expression (not necessarily square) into *this.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setIdentity_int_int.cpp
   * Output: \verbinclude Matrix_setIdentity_int_int.out
@@ -796,9 +792,9 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
   * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Identity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index nbRows, Index nbCols)
+EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index cols)
 {
-  derived().resize(nbRows, nbCols);
+  derived().resize(rows, cols);
   return setIdentity();
 }
 

Eigen/src/Core/CwiseUnaryOp.h

@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 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
@@ -44,10 +44,7 @@ struct traits<CwiseUnaryOp<UnaryOp, XprType> >
   typedef typename XprType::Nested XprTypeNested;
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   enum {
-    Flags = _XprTypeNested::Flags & (
-      HereditaryBits | LinearAccessBit | AlignedBit
-      | (functor_traits<UnaryOp>::PacketAccess ? PacketAccessBit : 0)),
-    CoeffReadCost = _XprTypeNested::CoeffReadCost + functor_traits<UnaryOp>::Cost
+    Flags = _XprTypeNested::Flags & RowMajorBit 
   };
 };
 }
@@ -56,16 +53,16 @@ template<typename UnaryOp, typename XprType, typename StorageKind>
 class CwiseUnaryOpImpl;
 
 template<typename UnaryOp, typename XprType>
-class CwiseUnaryOp : internal::no_assignment_operator,
-  public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>
+class CwiseUnaryOp : public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>, internal::no_assignment_operator
 {
   public:
 
     typedef typename CwiseUnaryOpImpl<UnaryOp, XprType,typename internal::traits<XprType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryOp)
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
 
     EIGEN_DEVICE_FUNC
-    inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
+    explicit inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
       : m_xpr(xpr), m_functor(func) {}
 
     EIGEN_DEVICE_FUNC
@@ -92,42 +89,13 @@ class CwiseUnaryOp : internal::no_assignment_operator,
     const UnaryOp m_functor;
 };
 
-// This is the generic implementation for dense storage.
-// It can be used for any expression types implementing the dense concept.
-template<typename UnaryOp, typename XprType>
-class CwiseUnaryOpImpl<UnaryOp,XprType,Dense>
-  : public internal::dense_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type
+// Generic API dispatcher
+template<typename UnaryOp, typename XprType, typename StorageKind>
+class CwiseUnaryOpImpl
+  : public internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type
 {
-  public:
-
-    typedef CwiseUnaryOp<UnaryOp, XprType> Derived;
-    typedef typename internal::dense_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
-
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(rowId, colId));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().functor().packetOp(derived().nestedExpression().template packet<LoadMode>(rowId, colId));
-    }
-
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(index));
-    }
-
-    template<int LoadMode>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return derived().functor().packetOp(derived().nestedExpression().template packet<LoadMode>(index));
-    }
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base;
 };
 
 } // end namespace Eigen

Eigen/src/Core/CwiseUnaryView.h

@@ -37,8 +37,8 @@ struct traits<CwiseUnaryView<ViewOp, MatrixType> >
   typedef typename MatrixType::Nested MatrixTypeNested;
   typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
-    Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)),
-    CoeffReadCost = traits<_MatrixTypeNested>::CoeffReadCost + functor_traits<ViewOp>::Cost,
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions
     MatrixTypeInnerStride =  inner_stride_at_compile_time<MatrixType>::ret,
     // need to cast the sizeof's from size_t to int explicitly, otherwise:
     // "error: no integral type can represent all of the enumerator values
@@ -62,8 +62,9 @@ class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename in
 
     typedef typename CwiseUnaryViewImpl<ViewOp, MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryView)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
-    inline CwiseUnaryView(const MatrixType& mat, const ViewOp& func = ViewOp())
+    explicit inline CwiseUnaryView(MatrixType& mat, const ViewOp& func = ViewOp())
       : m_matrix(mat), m_functor(func) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryView)
@@ -83,11 +84,19 @@ class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename in
     nestedExpression() { return m_matrix.const_cast_derived(); }
 
   protected:
-    // FIXME changed from MatrixType::Nested because of a weird compilation error with sun CC
-    typename internal::nested<MatrixType>::type m_matrix;
+    typename internal::ref_selector<MatrixType>::type m_matrix;
     ViewOp m_functor;
 };
 
+// Generic API dispatcher
+template<typename ViewOp, typename XprType, typename StorageKind>
+class CwiseUnaryViewImpl
+  : public internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type Base;
+};
+
 template<typename ViewOp, typename MatrixType>
 class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
   : public internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type
@@ -100,38 +109,18 @@ class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
     EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl)
     
-    inline Scalar* data() { return &coeffRef(0); }
-    inline const Scalar* data() const { return &coeff(0); }
+    EIGEN_DEVICE_FUNC inline Scalar* data() { return &(this->coeffRef(0)); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return &(this->coeff(0)); }
 
-    inline Index innerStride() const
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
     {
       return derived().nestedExpression().innerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
     }
 
-    inline Index outerStride() const
+    EIGEN_DEVICE_FUNC inline Index outerStride() const
     {
       return derived().nestedExpression().outerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
     }
-
-    EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(row, col));
-    }
-
-    EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(index));
-    }
-
-    EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col)
-    {
-      return derived().functor()(const_cast_derived().nestedExpression().coeffRef(row, col));
-    }
-
-    EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
-    {
-      return derived().functor()(const_cast_derived().nestedExpression().coeffRef(index));
-    }
 };
 
 } // end namespace Eigen

Eigen/src/Core/DenseBase.h

@@ -49,22 +49,37 @@ template<typename Derived> class DenseBase
   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/;
 
-    class InnerIterator;
+
+    /** Inner iterator type to iterate over the coefficients of a row or column.
+      * \sa class InnerIterator
+      */
+    typedef Eigen::InnerIterator<Derived> InnerIterator;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
 
-    /** \brief The type of indices 
-      * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
-      * \sa \ref TopicPreprocessorDirectives.
-      */
-    typedef typename internal::traits<Derived>::Index Index; 
+    /**
+      * \brief The type used to store indices
+      * \details This typedef is relevant for types that store multiple indices such as
+      *          PermutationMatrix or Transpositions, otherwise it defaults to Eigen::Index
+      * \sa \ref TopicPreprocessorDirectives, Eigen::Index, SparseMatrixBase.
+     */
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
 
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc. */
     typedef typename internal::traits<Derived>::Scalar Scalar;
-    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc.
+      *
+      * It is an alias for the Scalar type */
+    typedef Scalar value_type;
+    
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
-    typedef 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::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -74,16 +89,6 @@ template<typename Derived> class DenseBase
     using Base::colIndexByOuterInner;
     using Base::coeff;
     using Base::coeffByOuterInner;
-    using Base::packet;
-    using Base::packetByOuterInner;
-    using Base::writePacket;
-    using Base::writePacketByOuterInner;
-    using Base::coeffRef;
-    using Base::coeffRefByOuterInner;
-    using Base::copyCoeff;
-    using Base::copyCoeffByOuterInner;
-    using Base::copyPacket;
-    using Base::copyPacketByOuterInner;
     using Base::operator();
     using Base::operator[];
     using Base::x;
@@ -169,16 +174,42 @@ template<typename Derived> class DenseBase
       InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime)
                              : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
 
-      CoeffReadCost = internal::traits<Derived>::CoeffReadCost,
-        /**< This is a rough measure of how expensive it is to read one coefficient from
-          * this expression.
-          */
-
       InnerStrideAtCompileTime = internal::inner_stride_at_compile_time<Derived>::ret,
       OuterStrideAtCompileTime = internal::outer_stride_at_compile_time<Derived>::ret
     };
+    
+    typedef typename internal::find_best_packet<Scalar,SizeAtCompileTime>::type PacketScalar;
 
-    enum { ThisConstantIsPrivateInPlainObjectBase };
+    enum { IsPlainObjectBase = 0 };
+    
+    /** The plain matrix type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Matrix<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainMatrix;
+    
+    /** The plain array type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Array<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainArray;
+
+    /** \brief The plain matrix or array type corresponding to this expression.
+      *
+      * This is not necessarily exactly the return type of eval(). In the case of plain matrices,
+      * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed
+      * that the return type of eval() is either PlainObject or const PlainObject&.
+      */
+    typedef typename internal::conditional<internal::is_same<typename internal::traits<Derived>::XprKind,MatrixXpr >::value,
+                                 PlainMatrix, PlainArray>::type PlainObject;
 
     /** \returns the number of nonzero coefficients which is in practice the number
       * of stored coefficients. */
@@ -229,22 +260,21 @@ template<typename Derived> class DenseBase
       * nothing else.
       */
     EIGEN_DEVICE_FUNC
-    void resize(Index nbRows, Index nbCols)
+    void resize(Index rows, Index cols)
     {
-      EIGEN_ONLY_USED_FOR_DEBUG(nbRows);
-      EIGEN_ONLY_USED_FOR_DEBUG(nbCols);
-      eigen_assert(nbRows == this->rows() && nbCols == this->cols()
+      EIGEN_ONLY_USED_FOR_DEBUG(rows);
+      EIGEN_ONLY_USED_FOR_DEBUG(cols);
+      eigen_assert(rows == this->rows() && cols == this->cols()
                 && "DenseBase::resize() does not actually allow to resize.");
     }
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
     /** \internal Represents a vector with linearly spaced coefficients that allows sequential access only. */
-    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,false>,Derived> SequentialLinSpacedReturnType;
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar,false>,PlainObject> SequentialLinSpacedReturnType;
     /** \internal Represents a vector with linearly spaced coefficients that allows random access. */
-    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,true>,Derived> RandomAccessLinSpacedReturnType;
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar,true>,PlainObject> RandomAccessLinSpacedReturnType;
     /** \internal the return type of MatrixBase::eigenvalues() */
     typedef Matrix<typename NumTraits<typename internal::traits<Derived>::Scalar>::Real, internal::traits<Derived>::ColsAtCompileTime, 1> EigenvaluesReturnType;
 
@@ -277,37 +307,34 @@ template<typename Derived> class DenseBase
     EIGEN_DEVICE_FUNC
     Derived& operator=(const ReturnByValue<OtherDerived>& func);
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-    /** Copies \a other into *this without evaluating other. \returns a reference to *this. */
+    /** \ínternal
+      * Copies \a other into *this without evaluating other. \returns a reference to *this.
+      * \deprecated */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     Derived& lazyAssign(const DenseBase<OtherDerived>& other);
-#endif // not EIGEN_PARSED_BY_DOXYGEN
 
     EIGEN_DEVICE_FUNC
     CommaInitializer<Derived> operator<< (const Scalar& s);
 
+    /** \deprecated it now returns \c *this */
     template<unsigned int Added,unsigned int Removed>
-    const Flagged<Derived, Added, Removed> flagged() const;
+    EIGEN_DEPRECATED
+    const Derived& flagged() const
+    { return derived(); }
 
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     CommaInitializer<Derived> operator<< (const DenseBase<OtherDerived>& other);
 
+    typedef Transpose<Derived> TransposeReturnType;
     EIGEN_DEVICE_FUNC
-    Eigen::Transpose<Derived> transpose();
+    TransposeReturnType transpose();
     typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
     EIGEN_DEVICE_FUNC
     ConstTransposeReturnType transpose() const;
     EIGEN_DEVICE_FUNC
     void transposeInPlace();
-#ifndef EIGEN_NO_DEBUG
-  protected:
-    template<typename OtherDerived>
-    void checkTransposeAliasing(const OtherDerived& other) const;
-  public:
-#endif
-
 
     EIGEN_DEVICE_FUNC static const ConstantReturnType
     Constant(Index rows, Index cols, const Scalar& value);
@@ -326,13 +353,13 @@ template<typename Derived> class DenseBase
     LinSpaced(const Scalar& low, const Scalar& high);
 
     template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func);
     template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(Index size, const CustomNullaryOp& func);
     template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(const CustomNullaryOp& func);
 
     EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index rows, Index cols);
@@ -378,6 +405,8 @@ template<typename Derived> class DenseBase
       *
       * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
       * a const reference, in order to avoid a useless copy.
+      * 
+      * \warning Be carefull with eval() and the auto C++ keyword, as detailed in this \link TopicPitfalls_auto_keyword page \endlink.
       */
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE EvalReturnType eval() const
@@ -387,16 +416,17 @@ template<typename Derived> class DenseBase
       // size types on MSVC.
       return typename internal::eval<Derived>::type(derived());
     }
-
+    
     /** swaps *this with the expression \a other.
       *
       */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    void swap(const DenseBase<OtherDerived>& other,
-              int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase)
+    void swap(const DenseBase<OtherDerived>& other)
     {
-      SwapWrapper<Derived>(derived()).lazyAssign(other.derived());
+      EIGEN_STATIC_ASSERT(!OtherDerived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
     }
 
     /** swaps *this with the matrix or array \a other.
@@ -406,10 +436,10 @@ template<typename Derived> class DenseBase
     EIGEN_DEVICE_FUNC
     void swap(PlainObjectBase<OtherDerived>& other)
     {
-      SwapWrapper<Derived>(derived()).lazyAssign(other.derived());
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.derived(), internal::swap_assign_op<Scalar>());
     }
 
-
     EIGEN_DEVICE_FUNC inline const NestByValue<Derived> nestByValue() const;
     EIGEN_DEVICE_FUNC inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
     EIGEN_DEVICE_FUNC inline ForceAlignedAccess<Derived> forceAlignedAccess();
@@ -438,8 +468,7 @@ template<typename Derived> class DenseBase
 
     template<typename BinaryOp>
     EIGEN_DEVICE_FUNC
-    typename internal::result_of<BinaryOp(typename internal::traits<Derived>::Scalar)>::type
-    redux(const BinaryOp& func) const;
+    Scalar redux(const BinaryOp& func) const;
 
     template<typename Visitor>
     EIGEN_DEVICE_FUNC
@@ -465,14 +494,35 @@ template<typename Derived> class DenseBase
     typedef VectorwiseOp<Derived, Vertical> ColwiseReturnType;
     typedef const VectorwiseOp<const Derived, Vertical> ConstColwiseReturnType;
 
-    ConstRowwiseReturnType rowwise() const;
-    RowwiseReturnType rowwise();
-    ConstColwiseReturnType colwise() const;
-    ColwiseReturnType colwise();
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_rowwise.cpp
+    * Output: \verbinclude MatrixBase_rowwise.out
+    *
+    * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC inline ConstRowwiseReturnType rowwise() const {
+      return ConstRowwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC RowwiseReturnType rowwise();
 
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random(Index rows, Index cols);
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random(Index size);
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random();
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_colwise.cpp
+    * Output: \verbinclude MatrixBase_colwise.out
+    *
+    * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    EIGEN_DEVICE_FUNC inline ConstColwiseReturnType colwise() const {
+      return ConstColwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC ColwiseReturnType colwise();
+
+    typedef CwiseNullaryOp<internal::scalar_random_op<Scalar>,PlainObject> RandomReturnType;
+    static const RandomReturnType Random(Index rows, Index cols);
+    static const RandomReturnType Random(Index size);
+    static const RandomReturnType Random();
 
     template<typename ThenDerived,typename ElseDerived>
     const Select<Derived,ThenDerived,ElseDerived>
@@ -490,14 +540,33 @@ template<typename Derived> class DenseBase
     template<int p> RealScalar lpNorm() const;
 
     template<int RowFactor, int ColFactor>
+    EIGEN_DEVICE_FUNC
     const Replicate<Derived,RowFactor,ColFactor> replicate() const;
-    const Replicate<Derived,Dynamic,Dynamic> replicate(Index rowFacor,Index colFactor) const;
+    /**
+    * \return an expression of the replication of \c *this
+    *
+    * Example: \include MatrixBase_replicate_int_int.cpp
+    * Output: \verbinclude MatrixBase_replicate_int_int.out
+    *
+    * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC
+    const Replicate<Derived, Dynamic, Dynamic> replicate(Index rowFactor, Index colFactor) const
+    {
+      return Replicate<Derived, Dynamic, Dynamic>(derived(), rowFactor, colFactor);
+    }
 
     typedef Reverse<Derived, BothDirections> ReverseReturnType;
     typedef const Reverse<const Derived, BothDirections> ConstReverseReturnType;
-    ReverseReturnType reverse();
-    ConstReverseReturnType reverse() const;
-    void reverseInPlace();
+    EIGEN_DEVICE_FUNC ReverseReturnType reverse();
+    /** This is the const version of reverse(). */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC ConstReverseReturnType reverse() const
+    {
+      return ConstReverseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC void reverseInPlace();
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::DenseBase
 #   include "../plugins/BlockMethods.h"

Eigen/src/Core/DenseCoeffsBase.h

@@ -35,7 +35,6 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
 {
   public:
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
 
@@ -97,8 +96,8 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const
     {
       eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      return derived().coeff(row, col);
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
     }
 
     EIGEN_DEVICE_FUNC
@@ -117,7 +116,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     {
       eigen_assert(row >= 0 && row < rows()
           && col >= 0 && col < cols());
-      return derived().coeff(row, col);
+      return coeff(row, col);
     }
 
     /** Short version: don't use this function, use
@@ -140,7 +139,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     coeff(Index index) const
     {
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return internal::evaluator<Derived>(derived()).coeff(index);
     }
 
 
@@ -159,7 +158,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
                           THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
       eigen_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return coeff(index);
     }
 
     /** \returns the coefficient at given index.
@@ -177,7 +176,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     operator()(Index index) const
     {
       eigen_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return coeff(index);
     }
 
     /** equivalent to operator[](0).  */
@@ -217,9 +216,9 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     template<int LoadMode>
     EIGEN_STRONG_INLINE PacketReturnType packet(Index row, Index col) const
     {
-      eigen_internal_assert(row >= 0 && row < rows()
-                      && col >= 0 && col < cols());
-      return derived().template packet<LoadMode>(row,col);
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
+      eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(row,col);
     }
 
 
@@ -244,8 +243,9 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     template<int LoadMode>
     EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
     {
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().template packet<LoadMode>(index);
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(index);
     }
 
   protected:
@@ -288,7 +288,6 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -325,8 +324,8 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col)
     {
       eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      return derived().coeffRef(row, col);
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeffRef(row,col);
     }
 
     EIGEN_DEVICE_FUNC
@@ -348,7 +347,7 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     {
       eigen_assert(row >= 0 && row < rows()
           && col >= 0 && col < cols());
-      return derived().coeffRef(row, col);
+      return coeffRef(row, col);
     }
 
 
@@ -372,7 +371,7 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     coeffRef(Index index)
     {
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return internal::evaluator<Derived>(derived()).coeffRef(index);
     }
 
     /** \returns a reference to the coefficient at given index.
@@ -389,7 +388,7 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
                           THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
       eigen_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return coeffRef(index);
     }
 
     /** \returns a reference to the coefficient at given index.
@@ -406,7 +405,7 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     operator()(Index index)
     {
       eigen_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return coeffRef(index);
     }
 
     /** equivalent to operator[](0).  */
@@ -432,144 +431,6 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     w() { return (*this)[3]; }
-
-    /** \internal
-      * Stores the given packet of coefficients, at the given row and column of this expression. It is your responsibility
-      * to ensure that a packet really starts there. This method is only available on expressions having the
-      * PacketAccessBit.
-      *
-      * The \a LoadMode parameter may have the value \a #Aligned or \a #Unaligned. Its effect is to select
-      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
-      * starting at an address which is a multiple of the packet size.
-      */
-
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacket
-    (Index row, Index col, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().template writePacket<StoreMode>(row,col,val);
-    }
-
-
-    /** \internal */
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacketByOuterInner
-    (Index outer, Index inner, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      writePacket<StoreMode>(rowIndexByOuterInner(outer, inner),
-                            colIndexByOuterInner(outer, inner),
-                            val);
-    }
-
-    /** \internal
-      * Stores the given packet of coefficients, at the given index in this expression. It is your responsibility
-      * to ensure that a packet really starts there. This method is only available on expressions having the
-      * PacketAccessBit and the LinearAccessBit.
-      *
-      * The \a LoadMode parameter may have the value \a Aligned or \a Unaligned. Its effect is to select
-      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
-      * starting at an address which is a multiple of the packet size.
-      */
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacket
-    (Index index, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().template writePacket<StoreMode>(index,val);
-    }
-
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-
-    /** \internal Copies the coefficient at position (row,col) of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().coeffRef(row, col) = other.derived().coeff(row, col);
-    }
-
-    /** \internal Copies the coefficient at the given index of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().coeffRef(index) = other.derived().coeff(index);
-    }
-
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void copyCoeffByOuterInner(Index outer, Index inner, const DenseBase<OtherDerived>& other)
-    {
-      const Index row = rowIndexByOuterInner(outer,inner);
-      const Index col = colIndexByOuterInner(outer,inner);
-      // derived() is important here: copyCoeff() may be reimplemented in Derived!
-      derived().copyCoeff(row, col, other);
-    }
-
-    /** \internal Copies the packet at position (row,col) of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().template writePacket<StoreMode>(row, col,
-        other.derived().template packet<LoadMode>(row, col));
-    }
-
-    /** \internal Copies the packet at the given index of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().template writePacket<StoreMode>(index,
-        other.derived().template packet<LoadMode>(index));
-    }
-
-    /** \internal */
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacketByOuterInner(Index outer, Index inner, const DenseBase<OtherDerived>& other)
-    {
-      const Index row = rowIndexByOuterInner(outer,inner);
-      const Index col = colIndexByOuterInner(outer,inner);
-      // derived() is important here: copyCoeff() may be reimplemented in Derived!
-      derived().template copyPacket< OtherDerived, StoreMode, LoadMode>(row, col, other);
-    }
-#endif
-
 };
 
 /** \brief Base class providing direct read-only coefficient access to matrices and arrays.
@@ -589,7 +450,6 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
   public:
 
     typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -664,7 +524,6 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
   public:
 
     typedef DenseCoeffsBase<Derived, WriteAccessors> Base;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -723,33 +582,42 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
 
 namespace internal {
 
-template<typename Derived, bool JustReturnZero>
+template<int Alignment, typename Derived, bool JustReturnZero>
 struct first_aligned_impl
 {
-  static inline typename Derived::Index run(const Derived&)
+  static inline Index run(const Derived&)
   { return 0; }
 };
 
-template<typename Derived>
-struct first_aligned_impl<Derived, false>
+template<int Alignment, typename Derived>
+struct first_aligned_impl<Alignment, Derived, false>
 {
-  static inline typename Derived::Index run(const Derived& m)
+  static inline Index run(const Derived& m)
   {
-    return internal::first_aligned(&m.const_cast_derived().coeffRef(0,0), m.size());
+    return internal::first_aligned<Alignment>(&m.const_cast_derived().coeffRef(0,0), m.size());
   }
 };
 
-/** \internal \returns the index of the first element of the array that is well aligned for vectorization.
+/** \internal \returns the index of the first element of the array stored by \a m that is properly aligned with respect to \a Alignment for vectorization.
+  *
+  * \tparam Alignment requested alignment in Bytes.
   *
   * There is also the variant first_aligned(const Scalar*, Integer) defined in Memory.h. See it for more
   * documentation.
   */
-template<typename Derived>
-static inline typename Derived::Index first_aligned(const Derived& m)
+template<int Alignment, typename Derived>
+static inline Index first_aligned(const DenseBase<Derived>& m)
 {
-  return first_aligned_impl
-          <Derived, (Derived::Flags & AlignedBit) || !(Derived::Flags & DirectAccessBit)>
-          ::run(m);
+  enum { ReturnZero = (int(evaluator<Derived>::Alignment) >= Alignment) || !(Derived::Flags & DirectAccessBit) };
+  return first_aligned_impl<Alignment, Derived, ReturnZero>::run(m.derived());
+}
+
+template<typename Derived>
+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);
 }
 
 template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret>

Eigen/src/Core/DenseStorage.h

@@ -40,8 +40,7 @@ void check_static_allocation_size()
   */
 template <typename T, int Size, int MatrixOrArrayOptions,
           int Alignment = (MatrixOrArrayOptions&DontAlign) ? 0
-                        : (((Size*sizeof(T))%EIGEN_ALIGN_BYTES)==0) ? EIGEN_ALIGN_BYTES
-                        : 0 >
+                        : compute_default_alignment<T,Size>::value >
 struct plain_array
 {
   T array[Size];
@@ -81,14 +80,71 @@ struct plain_array
 #endif
 
 template <typename T, int Size, int MatrixOrArrayOptions>
-struct plain_array<T, Size, MatrixOrArrayOptions, EIGEN_ALIGN_BYTES>
+struct plain_array<T, Size, MatrixOrArrayOptions, 8>
 {
-  EIGEN_USER_ALIGN_DEFAULT T array[Size];
+  EIGEN_ALIGN_TO_BOUNDARY(8) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(7);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 16>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(16) T array[Size];
 
   EIGEN_DEVICE_FUNC
   plain_array() 
   { 
-    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(EIGEN_ALIGN_BYTES-1);
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(15);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 32>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(32) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(31);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 64>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(64) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  { 
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(63);
     check_static_allocation_size<T,Size>();
   }
 
@@ -102,7 +158,7 @@ struct plain_array<T, Size, MatrixOrArrayOptions, EIGEN_ALIGN_BYTES>
 template <typename T, int MatrixOrArrayOptions, int Alignment>
 struct plain_array<T, 0, MatrixOrArrayOptions, Alignment>
 {
-  EIGEN_USER_ALIGN_DEFAULT T array[1];
+  T array[1];
   EIGEN_DEVICE_FUNC plain_array() {}
   EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {}
 };
@@ -130,7 +186,7 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
   public:
     EIGEN_DEVICE_FUNC DenseStorage() {}
     EIGEN_DEVICE_FUNC
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()) {}
     EIGEN_DEVICE_FUNC 
     DenseStorage(const DenseStorage& other) : m_data(other.m_data) {}
@@ -140,12 +196,18 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
       if (this != &other) m_data = other.m_data;
       return *this; 
     }
-    EIGEN_DEVICE_FUNC DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols==_Cols);
+      EIGEN_UNUSED_VARIABLE(size);
+      EIGEN_UNUSED_VARIABLE(rows);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
     EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
-    EIGEN_DEVICE_FUNC static DenseIndex rows(void) {return _Rows;}
-    EIGEN_DEVICE_FUNC static DenseIndex cols(void) {return _Cols;}
-    EIGEN_DEVICE_FUNC void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
-    EIGEN_DEVICE_FUNC void resize(DenseIndex,DenseIndex,DenseIndex) {}
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
     EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
     EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
@@ -155,15 +217,15 @@ template<typename T, int _Rows, int _Cols, int _Options> class DenseStorage<T, 0
 {
   public:
     EIGEN_DEVICE_FUNC DenseStorage() {}
-    EIGEN_DEVICE_FUNC DenseStorage(internal::constructor_without_unaligned_array_assert) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) {}
     EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage&) {}
     EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage&) { return *this; }
-    EIGEN_DEVICE_FUNC DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index,Index,Index) {}
     EIGEN_DEVICE_FUNC void swap(DenseStorage& ) {}
-    EIGEN_DEVICE_FUNC static DenseIndex rows(void) {return _Rows;}
-    EIGEN_DEVICE_FUNC static DenseIndex cols(void) {return _Cols;}
-    EIGEN_DEVICE_FUNC void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
-    EIGEN_DEVICE_FUNC void resize(DenseIndex,DenseIndex,DenseIndex) {}
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
     EIGEN_DEVICE_FUNC const T *data() const { return 0; }
     EIGEN_DEVICE_FUNC T *data() { return 0; }
 };
@@ -182,14 +244,14 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, 0, Dynamic,
 template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic, Dynamic, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_rows;
-    DenseIndex m_cols;
+    Index m_rows;
+    Index m_cols;
   public:
     EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0), m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0), m_cols(0) {}
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {}
-    DenseStorage& operator=(const DenseStorage& other) 
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
     { 
       if (this != &other)
       {
@@ -199,13 +261,13 @@ template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic
       }
       return *this; 
     }
-    DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) : m_rows(nbRows), m_cols(nbCols) {}
-    void swap(DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index cols) : m_rows(rows), m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    EIGEN_DEVICE_FUNC DenseIndex rows() const {return m_rows;}
-    EIGEN_DEVICE_FUNC DenseIndex cols() const {return m_cols;}
-    void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
-    void resize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
+    EIGEN_DEVICE_FUNC Index rows() const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols() const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
     EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
     EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
@@ -214,13 +276,13 @@ template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic
 template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Size, Dynamic, _Cols, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_rows;
+    Index m_rows;
   public:
     EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {}
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {}
-    DenseStorage& operator=(const DenseStorage& other) 
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
     {
       if (this != &other)
       {
@@ -229,12 +291,12 @@ template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Si
       }
       return *this; 
     }
-    DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex) : m_rows(nbRows) {}
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
-    EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return m_rows;}
-    EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return _Cols;}
-    void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
-    void resize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index) : m_rows(rows) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index) { m_rows = rows; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index) { m_rows = rows; }
     EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
     EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
@@ -243,13 +305,13 @@ template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Si
 template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Size, _Rows, Dynamic, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_cols;
+    Index m_cols;
   public:
     EIGEN_DEVICE_FUNC DenseStorage() : m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {}
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {}
-    DenseStorage& operator=(const DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
       if (this != &other)
       {
@@ -258,12 +320,12 @@ template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Si
       }
       return *this;
     }
-    DenseStorage(DenseIndex, DenseIndex, DenseIndex nbCols) : m_cols(nbCols) {}
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
-    EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return _Rows;}
-    EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return m_cols;}
-    void conservativeResize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
-    void resize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index, Index cols) : m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index, Index, Index cols) { m_cols = cols; }
+    void resize(Index, Index, Index cols) { m_cols = cols; }
     EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
     EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
@@ -272,23 +334,26 @@ template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Si
 template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynamic, _Options>
 {
     T *m_data;
-    DenseIndex m_rows;
-    DenseIndex m_cols;
+    Index m_rows;
+    Index m_cols;
   public:
     EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
        : m_data(0), m_rows(0), m_cols(0) {}
-    DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
-      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows), m_cols(nbCols)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    DenseStorage(const DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols >=0);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
       : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*other.m_cols))
       , m_rows(other.m_rows)
       , m_cols(other.m_cols)
     {
       internal::smart_copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data);
     }
-    DenseStorage& operator=(const DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
       if (this != &other)
       {
@@ -298,13 +363,17 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
       return *this;
     }
 #ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
     DenseStorage(DenseStorage&& other)
       : m_data(std::move(other.m_data))
       , m_rows(std::move(other.m_rows))
       , m_cols(std::move(other.m_cols))
     {
       other.m_data = nullptr;
+      other.m_rows = 0;
+      other.m_cols = 0;
     }
+    EIGEN_DEVICE_FUNC
     DenseStorage& operator=(DenseStorage&& other)
     {
       using std::swap;
@@ -314,18 +383,18 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
       return *this;
     }
 #endif
-    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
-    void swap(DenseStorage& other)
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return m_rows;}
-    EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return m_cols;}
-    void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index size, Index rows, Index cols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*m_cols);
-      m_rows = nbRows;
-      m_cols = nbCols;
+      m_rows = rows;
+      m_cols = cols;
     }
-    void resize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC void resize(Index size, Index rows, Index cols)
     {
       if(size != m_rows*m_cols)
       {
@@ -336,8 +405,8 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
           m_data = 0;
         EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
       }
-      m_rows = nbRows;
-      m_cols = nbCols;
+      m_rows = rows;
+      m_cols = cols;
     }
     EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
     EIGEN_DEVICE_FUNC T *data() { return m_data; }
@@ -347,19 +416,23 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
 template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Rows, Dynamic, _Options>
 {
     T *m_data;
-    DenseIndex m_cols;
+    Index m_cols;
   public:
     EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_cols(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
-    DenseStorage(DenseIndex size, DenseIndex, DenseIndex nbCols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(nbCols)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    DenseStorage(const DenseStorage& other)
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols >=0);
+      EIGEN_UNUSED_VARIABLE(rows);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
       : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(_Rows*other.m_cols))
       , m_cols(other.m_cols)
     {
       internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data);
     }
-    DenseStorage& operator=(const DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
       if (this != &other)
       {
@@ -369,12 +442,15 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
       return *this;
     }    
 #ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
     DenseStorage(DenseStorage&& other)
       : m_data(std::move(other.m_data))
       , m_cols(std::move(other.m_cols))
     {
       other.m_data = nullptr;
+      other.m_cols = 0;
     }
+    EIGEN_DEVICE_FUNC
     DenseStorage& operator=(DenseStorage&& other)
     {
       using std::swap;
@@ -383,16 +459,16 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
       return *this;
     }
 #endif
-    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
-    EIGEN_DEVICE_FUNC static DenseIndex rows(void) {return _Rows;}
-    EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return m_cols;}
-    void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index size, Index, Index cols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, _Rows*m_cols);
-      m_cols = nbCols;
+      m_cols = cols;
     }
-    EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index, Index cols)
     {
       if(size != _Rows*m_cols)
       {
@@ -403,7 +479,7 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
           m_data = 0;
         EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
       }
-      m_cols = nbCols;
+      m_cols = cols;
     }
     EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
     EIGEN_DEVICE_FUNC T *data() { return m_data; }
@@ -413,19 +489,23 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
 template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dynamic, _Cols, _Options>
 {
     T *m_data;
-    DenseIndex m_rows;
+    Index m_rows;
   public:
     EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0) {}
-    DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
-    DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    DenseStorage(const DenseStorage& other)
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols == _Cols);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
       : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*_Cols))
       , m_rows(other.m_rows)
     {
       internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data);
     }
-    DenseStorage& operator=(const DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
     {
       if (this != &other)
       {
@@ -435,12 +515,15 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
       return *this;
     }    
 #ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
     DenseStorage(DenseStorage&& other)
       : m_data(std::move(other.m_data))
       , m_rows(std::move(other.m_rows))
     {
       other.m_data = nullptr;
+      other.m_rows = 0;
     }
+    EIGEN_DEVICE_FUNC
     DenseStorage& operator=(DenseStorage&& other)
     {
       using std::swap;
@@ -449,16 +532,16 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
       return *this;
     }
 #endif
-    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
-    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
-    EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return m_rows;}
-    EIGEN_DEVICE_FUNC static DenseIndex cols(void) {return _Cols;}
-    void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex)
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    void conservativeResize(Index size, Index rows, Index)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*_Cols);
-      m_rows = nbRows;
+      m_rows = rows;
     }
-    EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex nbRows, DenseIndex)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index rows, Index)
     {
       if(size != m_rows*_Cols)
       {
@@ -469,7 +552,7 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
           m_data = 0;
         EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
       }
-      m_rows = nbRows;
+      m_rows = rows;
     }
     EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
     EIGEN_DEVICE_FUNC T *data() { return m_data; }

Eigen/src/Core/Diagonal.h

@@ -37,7 +37,7 @@ template<typename MatrixType, int DiagIndex>
 struct traits<Diagonal<MatrixType,DiagIndex> >
  : traits<MatrixType>
 {
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   typedef typename MatrixType::StorageKind StorageKind;
   enum {
@@ -52,8 +52,7 @@ struct traits<Diagonal<MatrixType,DiagIndex> >
                                                  MatrixType::MaxColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))),
     MaxColsAtCompileTime = 1,
     MaskLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
-    Flags = (unsigned int)_MatrixTypeNested::Flags & (HereditaryBits | LinearAccessBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit,
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost,
+    Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions
     MatrixTypeOuterStride = outer_stride_at_compile_time<MatrixType>::ret,
     InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1,
     OuterStrideAtCompileTime = 0
@@ -71,17 +70,15 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
     EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal)
 
     EIGEN_DEVICE_FUNC
-    inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
+    explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal)
 
     EIGEN_DEVICE_FUNC
     inline Index rows() const
     {
-      EIGEN_USING_STD_MATH(min);
-      return m_index.value()<0 ? (min)(Index(m_matrix.cols()),Index(m_matrix.rows()+m_index.value()))
-                               : (min)(Index(m_matrix.rows()),Index(m_matrix.cols()-m_index.value()));
-      
+      return m_index.value()<0 ? numext::mini<Index>(m_matrix.cols(),m_matrix.rows()+m_index.value())
+                               : numext::mini<Index>(m_matrix.rows(),m_matrix.cols()-m_index.value());
     }
 
     EIGEN_DEVICE_FUNC
@@ -149,14 +146,14 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
     }
 
     EIGEN_DEVICE_FUNC
-    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    inline const typename internal::remove_all<typename MatrixType::Nested>::type& 
     nestedExpression() const 
     {
       return m_matrix;
     }
 
     EIGEN_DEVICE_FUNC
-    int index() const
+    inline Index index() const
     {
       return m_index.value();
     }
@@ -173,7 +170,7 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
     EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); }
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; }
-    // triger a compile time error is someone try to call packet
+    // trigger a compile-time error if someone try to call packet
     template<int LoadMode> typename MatrixType::PacketReturnType packet(Index) const;
     template<int LoadMode> typename MatrixType::PacketReturnType packet(Index,Index) const;
 };
@@ -190,7 +187,7 @@ template<typename Derived>
 inline typename MatrixBase<Derived>::DiagonalReturnType
 MatrixBase<Derived>::diagonal()
 {
-  return derived();
+  return DiagonalReturnType(derived());
 }
 
 /** This is the const version of diagonal(). */
@@ -213,18 +210,18 @@ MatrixBase<Derived>::diagonal() const
   *
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
-inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<DynamicIndex>::Type
+inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index)
 {
-  return typename DiagonalIndexReturnType<DynamicIndex>::Type(derived(), index);
+  return DiagonalDynamicIndexReturnType(derived(), index);
 }
 
 /** This is the const version of diagonal(Index). */
 template<typename Derived>
-inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<DynamicIndex>::Type
+inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index) const
 {
-  return typename ConstDiagonalIndexReturnType<DynamicIndex>::Type(derived(), index);
+  return ConstDiagonalDynamicIndexReturnType(derived(), index);
 }
 
 /** \returns an expression of the \a DiagIndex-th sub or super diagonal of the matrix \c *this
@@ -239,20 +236,20 @@ MatrixBase<Derived>::diagonal(Index index) const
   *
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
-template<int Index>
-inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index>::Type
+template<int Index_>
+inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal()
 {
-  return derived();
+  return typename DiagonalIndexReturnType<Index_>::Type(derived());
 }
 
 /** This is the const version of diagonal<int>(). */
 template<typename Derived>
-template<int Index>
-inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index>::Type
+template<int Index_>
+inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal() const
 {
-  return derived();
+  return typename ConstDiagonalIndexReturnType<Index_>::Type(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/DiagonalMatrix.h

@@ -22,7 +22,7 @@ class DiagonalBase : public EigenBase<Derived>
     typedef typename DiagonalVectorType::Scalar Scalar;
     typedef typename DiagonalVectorType::RealScalar RealScalar;
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
 
     enum {
       RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
@@ -30,7 +30,7 @@ class DiagonalBase : public EigenBase<Derived>
       MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
       MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
       IsVectorAtCompileTime = 0,
-      Flags = 0
+      Flags = NoPreferredStorageOrderBit
     };
 
     typedef Matrix<Scalar, RowsAtCompileTime, ColsAtCompileTime, 0, MaxRowsAtCompileTime, MaxColsAtCompileTime> DenseMatrixType;
@@ -44,18 +44,7 @@ class DiagonalBase : public EigenBase<Derived>
 
     EIGEN_DEVICE_FUNC
     DenseMatrixType toDenseMatrix() const { return derived(); }
-    template<typename DenseDerived>
-    EIGEN_DEVICE_FUNC
-    void evalTo(MatrixBase<DenseDerived> &other) const;
-    template<typename DenseDerived>
-    EIGEN_DEVICE_FUNC
-    void addTo(MatrixBase<DenseDerived> &other) const
-    { other.diagonal() += diagonal(); }
-    template<typename DenseDerived>
-    EIGEN_DEVICE_FUNC
-    void subTo(MatrixBase<DenseDerived> &other) const
-    { other.diagonal() -= diagonal(); }
-
+    
     EIGEN_DEVICE_FUNC
     inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); }
     EIGEN_DEVICE_FUNC
@@ -66,44 +55,37 @@ class DiagonalBase : public EigenBase<Derived>
     EIGEN_DEVICE_FUNC
     inline Index cols() const { return diagonal().size(); }
 
-    /** \returns the diagonal matrix product of \c *this by the matrix \a matrix.
-      */
     template<typename MatrixDerived>
     EIGEN_DEVICE_FUNC
-    const DiagonalProduct<MatrixDerived, Derived, OnTheLeft>
+    const Product<Derived,MatrixDerived,LazyProduct>
     operator*(const MatrixBase<MatrixDerived> &matrix) const
     {
-      return DiagonalProduct<MatrixDerived, Derived, OnTheLeft>(matrix.derived(), derived());
+      return Product<Derived, MatrixDerived, LazyProduct>(derived(),matrix.derived());
     }
 
+    typedef DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> > InverseReturnType;
     EIGEN_DEVICE_FUNC
-    inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> >
+    inline const InverseReturnType
     inverse() const
     {
-      return diagonal().cwiseInverse();
+      return InverseReturnType(diagonal().cwiseInverse());
     }
     
+    typedef DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> > ScalarMultipleReturnType;
     EIGEN_DEVICE_FUNC
-    inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> >
+    inline const ScalarMultipleReturnType
     operator*(const Scalar& scalar) const
     {
-      return diagonal() * scalar;
+      return ScalarMultipleReturnType(diagonal() * scalar);
     }
     EIGEN_DEVICE_FUNC
-    friend inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> >
+    friend inline const ScalarMultipleReturnType
     operator*(const Scalar& scalar, const DiagonalBase& other)
     {
-      return other.diagonal() * scalar;
+      return ScalarMultipleReturnType(other.diagonal() * scalar);
     }
 };
 
-template<typename Derived>
-template<typename DenseDerived>
-void DiagonalBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
-{
-  other.setZero();
-  other.diagonal() = diagonal();
-}
 #endif
 
 /** \class DiagonalMatrix
@@ -125,10 +107,9 @@ struct traits<DiagonalMatrix<_Scalar,SizeAtCompileTime,MaxSizeAtCompileTime> >
  : traits<Matrix<_Scalar,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
   typedef Matrix<_Scalar,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1> DiagonalVectorType;
-  typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef DiagonalShape StorageKind;
   enum {
-    Flags = LvalueBit
+    Flags = LvalueBit | NoPreferredStorageOrderBit
   };
 };
 }
@@ -142,7 +123,7 @@ class DiagonalMatrix
     typedef const DiagonalMatrix& Nested;
     typedef _Scalar Scalar;
     typedef typename internal::traits<DiagonalMatrix>::StorageKind StorageKind;
-    typedef typename internal::traits<DiagonalMatrix>::Index Index;
+    typedef typename internal::traits<DiagonalMatrix>::StorageIndex StorageIndex;
     #endif
 
   protected:
@@ -164,7 +145,7 @@ class DiagonalMatrix
 
     /** Constructs a diagonal matrix with given dimension  */
     EIGEN_DEVICE_FUNC
-    inline DiagonalMatrix(Index dim) : m_diagonal(dim) {}
+    explicit inline DiagonalMatrix(Index dim) : m_diagonal(dim) {}
 
     /** 2D constructor. */
     EIGEN_DEVICE_FUNC
@@ -248,14 +229,15 @@ struct traits<DiagonalWrapper<_DiagonalVectorType> >
 {
   typedef _DiagonalVectorType DiagonalVectorType;
   typedef typename DiagonalVectorType::Scalar Scalar;
-  typedef typename DiagonalVectorType::Index Index;
-  typedef typename DiagonalVectorType::StorageKind StorageKind;
+  typedef typename DiagonalVectorType::StorageIndex StorageIndex;
+  typedef DiagonalShape StorageKind;
+  typedef typename traits<DiagonalVectorType>::XprKind XprKind;
   enum {
     RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
     ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    MaxRowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    MaxColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    Flags =  traits<DiagonalVectorType>::Flags & LvalueBit
+    MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    Flags =  (traits<DiagonalVectorType>::Flags & LvalueBit) | NoPreferredStorageOrderBit
   };
 };
 }
@@ -272,7 +254,7 @@ class DiagonalWrapper
 
     /** Constructor from expression of diagonal coefficients to wrap. */
     EIGEN_DEVICE_FUNC
-    inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {}
+    explicit inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {}
 
     /** \returns a const reference to the wrapped expression of diagonal coefficients. */
     EIGEN_DEVICE_FUNC
@@ -295,7 +277,7 @@ template<typename Derived>
 inline const DiagonalWrapper<const Derived>
 MatrixBase<Derived>::asDiagonal() const
 {
-  return derived();
+  return DiagonalWrapper<const Derived>(derived());
 }
 
 /** \returns true if *this is approximately equal to a diagonal matrix,
@@ -326,6 +308,33 @@ bool MatrixBase<Derived>::isDiagonal(const RealScalar& prec) const
   return true;
 }
 
+namespace internal {
+
+template<> struct storage_kind_to_shape<DiagonalShape> { typedef DiagonalShape Shape; };
+
+struct Diagonal2Dense {};
+
+template<> struct AssignmentKind<DenseShape,DiagonalShape> { typedef Diagonal2Dense Kind; };
+
+// Diagonal matrix to Dense assignment
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar>
+struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Dense, Scalar>
+{
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar> &/*func*/)
+  {
+    dst.setZero();
+    dst.diagonal() = src.diagonal();
+  }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar> &/*func*/)
+  { dst.diagonal() += src.diagonal(); }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar> &/*func*/)
+  { dst.diagonal() -= src.diagonal(); }
+};
+
+} // namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_DIAGONALMATRIX_H

Eigen/src/Core/DiagonalProduct.h

@@ -13,116 +13,14 @@
 
 namespace Eigen { 
 
-namespace internal {
-template<typename MatrixType, typename DiagonalType, int ProductOrder>
-struct traits<DiagonalProduct<MatrixType, DiagonalType, ProductOrder> >
- : traits<MatrixType>
-{
-  typedef typename scalar_product_traits<typename MatrixType::Scalar, typename DiagonalType::Scalar>::ReturnType Scalar;
-  enum {
-    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
-    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
-    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
-
-    _StorageOrder = MatrixType::Flags & RowMajorBit ? RowMajor : ColMajor,
-    _ScalarAccessOnDiag =  !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft)
-                          ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)),
-    _SameTypes = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value,
-    // FIXME currently we need same types, but in the future the next rule should be the one
-    //_Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))),
-    _Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))),
-    _LinearAccessMask = (RowsAtCompileTime==1 || ColsAtCompileTime==1) ? LinearAccessBit : 0,
-
-    Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit,//(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit),
-    CoeffReadCost = NumTraits<Scalar>::MulCost + MatrixType::CoeffReadCost + DiagonalType::DiagonalVectorType::CoeffReadCost
-  };
-};
-}
-
-template<typename MatrixType, typename DiagonalType, int ProductOrder>
-class DiagonalProduct : internal::no_assignment_operator,
-                        public MatrixBase<DiagonalProduct<MatrixType, DiagonalType, ProductOrder> >
-{
-  public:
-
-    typedef MatrixBase<DiagonalProduct> Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(DiagonalProduct)
-
-    inline DiagonalProduct(const MatrixType& matrix, const DiagonalType& diagonal)
-      : m_matrix(matrix), m_diagonal(diagonal)
-    {
-      eigen_assert(diagonal.diagonal().size() == (ProductOrder == OnTheLeft ? matrix.rows() : matrix.cols()));
-    }
-
-    EIGEN_STRONG_INLINE Index rows() const { return m_matrix.rows(); }
-    EIGEN_STRONG_INLINE Index cols() const { return m_matrix.cols(); }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
-    {
-      return m_diagonal.diagonal().coeff(ProductOrder == OnTheLeft ? row : col) * m_matrix.coeff(row, col);
-    }
-    
-    EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const
-    {
-      enum {
-        StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor
-      };
-      return coeff(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const
-    {
-      enum {
-        StorageOrder = Flags & RowMajorBit ? RowMajor : ColMajor
-      };
-      const Index indexInDiagonalVector = ProductOrder == OnTheLeft ? row : col;
-      return packet_impl<LoadMode>(row,col,indexInDiagonalVector,typename internal::conditional<
-        ((int(StorageOrder) == RowMajor && int(ProductOrder) == OnTheLeft)
-       ||(int(StorageOrder) == ColMajor && int(ProductOrder) == OnTheRight)), internal::true_type, internal::false_type>::type());
-    }
-    
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const
-    {
-      enum {
-        StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor
-      };
-      return packet<LoadMode>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
-    }
-
-  protected:
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::true_type) const
-    {
-      return internal::pmul(m_matrix.template packet<LoadMode>(row, col),
-                     internal::pset1<PacketScalar>(m_diagonal.diagonal().coeff(id)));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::false_type) const
-    {
-      enum {
-        InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime,
-        DiagonalVectorPacketLoadMode = (LoadMode == Aligned && (((InnerSize%16) == 0) || (int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit)==AlignedBit) ? Aligned : Unaligned)
-      };
-      return internal::pmul(m_matrix.template packet<LoadMode>(row, col),
-                     m_diagonal.diagonal().template packet<DiagonalVectorPacketLoadMode>(id));
-    }
-
-    typename MatrixType::Nested m_matrix;
-    typename DiagonalType::Nested m_diagonal;
-};
-
 /** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal.
   */
 template<typename Derived>
 template<typename DiagonalDerived>
-inline const DiagonalProduct<Derived, DiagonalDerived, OnTheRight>
+inline const Product<Derived, DiagonalDerived, LazyProduct>
 MatrixBase<Derived>::operator*(const DiagonalBase<DiagonalDerived> &a_diagonal) const
 {
-  return DiagonalProduct<Derived, DiagonalDerived, OnTheRight>(derived(), a_diagonal.derived());
+  return Product<Derived, DiagonalDerived, LazyProduct>(derived(),a_diagonal.derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/Dot.h

@@ -99,7 +99,7 @@ EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scala
 template<typename Derived>
 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
 {
-  using std::sqrt;
+  EIGEN_USING_STD_MATH(sqrt)
   return sqrt(squaredNorm());
 }
 
@@ -113,8 +113,7 @@ template<typename Derived>
 inline const typename MatrixBase<Derived>::PlainObject
 MatrixBase<Derived>::normalized() const
 {
-  typedef typename internal::nested<Derived>::type Nested;
-  typedef typename internal::remove_reference<Nested>::type _Nested;
+  typedef typename internal::nested_eval<Derived,2>::type _Nested;
   _Nested n(derived());
   return n / n.norm();
 }
@@ -142,7 +141,7 @@ struct lpNorm_selector
   EIGEN_DEVICE_FUNC
   static inline RealScalar run(const MatrixBase<Derived>& m)
   {
-    using std::pow;
+    EIGEN_USING_STD_MATH(pow)
     return pow(m.cwiseAbs().array().pow(p).sum(), RealScalar(1)/p);
   }
 };
@@ -206,8 +205,8 @@ template<typename OtherDerived>
 bool MatrixBase<Derived>::isOrthogonal
 (const MatrixBase<OtherDerived>& other, const RealScalar& prec) const
 {
-  typename internal::nested<Derived,2>::type nested(derived());
-  typename internal::nested<OtherDerived,2>::type otherNested(other.derived());
+  typename internal::nested_eval<Derived,2>::type nested(derived());
+  typename internal::nested_eval<OtherDerived,2>::type otherNested(other.derived());
   return numext::abs2(nested.dot(otherNested)) <= prec * prec * nested.squaredNorm() * otherNested.squaredNorm();
 }
 
@@ -225,13 +224,13 @@ bool MatrixBase<Derived>::isOrthogonal
 template<typename Derived>
 bool MatrixBase<Derived>::isUnitary(const RealScalar& prec) const
 {
-  typename Derived::Nested nested(derived());
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index i = 0; i < cols(); ++i)
   {
-    if(!internal::isApprox(nested.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
+    if(!internal::isApprox(self.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
       return false;
     for(Index j = 0; j < i; ++j)
-      if(!internal::isMuchSmallerThan(nested.col(i).dot(nested.col(j)), static_cast<Scalar>(1), prec))
+      if(!internal::isMuchSmallerThan(self.col(i).dot(self.col(j)), static_cast<Scalar>(1), prec))
         return false;
   }
   return true;

Eigen/src/Core/EigenBase.h

@@ -13,7 +13,9 @@
 
 namespace Eigen {
 
-/** Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
+/** \class EigenBase
+  * 
+  * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
   *
   * In other words, an EigenBase object is an object that can be copied into a MatrixBase.
   *
@@ -26,9 +28,16 @@ namespace Eigen {
 template<typename Derived> struct EigenBase
 {
 //   typedef typename internal::plain_matrix_type<Derived>::type PlainObject;
+  
+  /** \brief The interface type of indices
+    * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
+    * \deprecated Since Eigen 3.3, its usage is deprecated. Use Eigen::Index instead.
+    * \sa StorageIndex, \ref TopicPreprocessorDirectives.
+    */
+  typedef Eigen::Index Index;
 
+  // FIXME is it needed?
   typedef typename internal::traits<Derived>::StorageKind StorageKind;
-  typedef typename internal::traits<Derived>::Index Index;
 
   /** \returns a reference to the derived object */
   EIGEN_DEVICE_FUNC
@@ -121,7 +130,7 @@ template<typename Derived>
 template<typename OtherDerived>
 Derived& DenseBase<Derived>::operator=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().evalTo(derived());
+  call_assignment(derived(), other.derived());
   return derived();
 }
 
@@ -129,7 +138,7 @@ template<typename Derived>
 template<typename OtherDerived>
 Derived& DenseBase<Derived>::operator+=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().addTo(derived());
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar>());
   return derived();
 }
 
@@ -137,7 +146,7 @@ template<typename Derived>
 template<typename OtherDerived>
 Derived& DenseBase<Derived>::operator-=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().subTo(derived());
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar>());
   return derived();
 }
 

Eigen/src/Core/Flagged.h

@@ -1,140 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@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_FLAGGED_H
-#define EIGEN_FLAGGED_H
-
-namespace Eigen { 
-
-/** \class Flagged
-  * \ingroup Core_Module
-  *
-  * \brief Expression with modified flags
-  *
-  * \param ExpressionType the type of the object of which we are modifying the flags
-  * \param Added the flags added to the expression
-  * \param Removed the flags removed from the expression (has priority over Added).
-  *
-  * This class represents an expression whose flags have been modified.
-  * It is the return type of MatrixBase::flagged()
-  * and most of the time this is the only way it is used.
-  *
-  * \sa MatrixBase::flagged()
-  */
-
-namespace internal {
-template<typename ExpressionType, unsigned int Added, unsigned int Removed>
-struct traits<Flagged<ExpressionType, Added, Removed> > : traits<ExpressionType>
-{
-  enum { Flags = (ExpressionType::Flags | Added) & ~Removed };
-};
-}
-
-template<typename ExpressionType, unsigned int Added, unsigned int Removed> class Flagged
-  : public MatrixBase<Flagged<ExpressionType, Added, Removed> >
-{
-  public:
-
-    typedef MatrixBase<Flagged> Base;
-    
-    EIGEN_DENSE_PUBLIC_INTERFACE(Flagged)
-    typedef typename internal::conditional<internal::must_nest_by_value<ExpressionType>::ret,
-        ExpressionType, const ExpressionType&>::type ExpressionTypeNested;
-    typedef typename ExpressionType::InnerIterator InnerIterator;
-
-    inline Flagged(const ExpressionType& matrix) : m_matrix(matrix) {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-    inline Index outerStride() const { return m_matrix.outerStride(); }
-    inline Index innerStride() const { return m_matrix.innerStride(); }
-
-    inline CoeffReturnType coeff(Index row, Index col) const
-    {
-      return m_matrix.coeff(row, col);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_matrix.coeff(index);
-    }
-    
-    inline const Scalar& coeffRef(Index row, Index col) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-
-    inline const Scalar& coeffRef(Index index) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
-    {
-      return m_matrix.template packet<LoadMode>(row, col);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(row, col, x);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return m_matrix.template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(index, x);
-    }
-
-    const ExpressionType& _expression() const { return m_matrix; }
-
-    template<typename OtherDerived>
-    typename ExpressionType::PlainObject solveTriangular(const MatrixBase<OtherDerived>& other) const;
-
-    template<typename OtherDerived>
-    void solveTriangularInPlace(const MatrixBase<OtherDerived>& other) const;
-
-  protected:
-    ExpressionTypeNested m_matrix;
-};
-
-/** \returns an expression of *this with added and removed flags
-  *
-  * This is mostly for internal use.
-  *
-  * \sa class Flagged
-  */
-template<typename Derived>
-template<unsigned int Added,unsigned int Removed>
-inline const Flagged<Derived, Added, Removed>
-DenseBase<Derived>::flagged() const
-{
-  return derived();
-}
-
-} // end namespace Eigen
-
-#endif // EIGEN_FLAGGED_H

Eigen/src/Core/ForceAlignedAccess.h

@@ -39,29 +39,29 @@ template<typename ExpressionType> class ForceAlignedAccess
     typedef typename internal::dense_xpr_base<ForceAlignedAccess>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(ForceAlignedAccess)
 
-    inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {}
+    EIGEN_DEVICE_FUNC explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {}
 
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
 
-    inline Scalar& coeffRef(Index row, Index col)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
     {
       return m_expression.const_cast_derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
-    inline Scalar& coeffRef(Index index)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
@@ -90,7 +90,7 @@ template<typename ExpressionType> class ForceAlignedAccess
       m_expression.const_cast_derived().template writePacket<Aligned>(index, x);
     }
 
-    operator const ExpressionType&() const { return m_expression; }
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
 
   protected:
     const ExpressionType& m_expression;
@@ -127,7 +127,7 @@ template<bool Enable>
 inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type
 MatrixBase<Derived>::forceAlignedAccessIf() const
 {
-  return derived();
+  return derived();  // FIXME This should not work but apparently is never used
 }
 
 /** \returns an expression of *this with forced aligned access if \a Enable is true.
@@ -138,7 +138,7 @@ template<bool Enable>
 inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type
 MatrixBase<Derived>::forceAlignedAccessIf()
 {
-  return derived();
+  return derived();  // FIXME This should not work but apparently is never used
 }
 
 } // end namespace Eigen

Eigen/src/Core/Fuzzy.h

@@ -22,10 +22,9 @@ struct isApprox_selector
   EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec)
   {
-    EIGEN_USING_STD_MATH(min);
-    typename internal::nested<Derived,2>::type nested(x);
-    typename internal::nested<OtherDerived,2>::type otherNested(y);
-    return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
+    typename internal::nested_eval<Derived,2>::type nested(x);
+    typename internal::nested_eval<OtherDerived,2>::type otherNested(y);
+    return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * numext::mini(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
   }
 };
 

Eigen/src/Core/GeneralProduct.h

@@ -11,29 +11,7 @@
 #ifndef EIGEN_GENERAL_PRODUCT_H
 #define EIGEN_GENERAL_PRODUCT_H
 
-namespace Eigen { 
-
-/** \class GeneralProduct
-  * \ingroup Core_Module
-  *
-  * \brief Expression of the product of two general matrices or vectors
-  *
-  * \param LhsNested the type used to store the left-hand side
-  * \param RhsNested the type used to store the right-hand side
-  * \param ProductMode the type of the product
-  *
-  * This class represents an expression of the product of two general matrices.
-  * We call a general matrix, a dense matrix with full storage. For instance,
-  * This excludes triangular, selfadjoint, and sparse matrices.
-  * It is the return type of the operator* between general matrices. Its template
-  * arguments are determined automatically by ProductReturnType. Therefore,
-  * GeneralProduct should never be used direclty. To determine the result type of a
-  * function which involves a matrix product, use ProductReturnType::Type.
-  *
-  * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
-  */
-template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value>
-class GeneralProduct;
+namespace Eigen {
 
 enum {
   Large = 2,
@@ -59,14 +37,14 @@ template<typename Lhs, typename Rhs> struct product_type
   typedef typename remove_all<Lhs>::type _Lhs;
   typedef typename remove_all<Rhs>::type _Rhs;
   enum {
-    MaxRows  = _Lhs::MaxRowsAtCompileTime,
-    Rows  = _Lhs::RowsAtCompileTime,
-    MaxCols  = _Rhs::MaxColsAtCompileTime,
-    Cols  = _Rhs::ColsAtCompileTime,
-    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime,
-                                           _Rhs::MaxRowsAtCompileTime),
-    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime,
-                                        _Rhs::RowsAtCompileTime)
+    MaxRows = traits<_Lhs>::MaxRowsAtCompileTime,
+    Rows    = traits<_Lhs>::RowsAtCompileTime,
+    MaxCols = traits<_Rhs>::MaxColsAtCompileTime,
+    Cols    = traits<_Rhs>::ColsAtCompileTime,
+    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::MaxColsAtCompileTime,
+                                           traits<_Rhs>::MaxRowsAtCompileTime),
+    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::ColsAtCompileTime,
+                                        traits<_Rhs>::RowsAtCompileTime)
   };
 
   // the splitting into different lines of code here, introducing the _select enums and the typedef below,
@@ -81,7 +59,8 @@ private:
 
 public:
   enum {
-    value = selector::ret
+    value = selector::ret,
+    ret = selector::ret
   };
 #ifdef EIGEN_DEBUG_PRODUCT
   static void debug()
@@ -97,6 +76,31 @@ public:
 #endif
 };
 
+// template<typename Lhs, typename Rhs> struct product_tag
+// {
+// private:
+//   
+//   typedef typename remove_all<Lhs>::type _Lhs;
+//   typedef typename remove_all<Rhs>::type _Rhs;
+//   enum {
+//     Rows  = _Lhs::RowsAtCompileTime,
+//     Cols  = _Rhs::ColsAtCompileTime,
+//     Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, _Rhs::RowsAtCompileTime)
+//   };
+// 
+//   enum {
+//     rows_select = Rows==1 ? int(Rows) : int(Large),
+//     cols_select = Cols==1 ? int(Cols) : int(Large),
+//     depth_select = Depth==1 ? int(Depth) : int(Large)
+//   };
+//   typedef product_type_selector<rows_select, cols_select, depth_select> selector;
+// 
+// public:
+//   enum {
+//     ret = selector::ret
+//   };
+// 
+// };
 
 /* The following allows to select the kind of product at compile time
  * based on the three dimensions of the product.
@@ -127,54 +131,6 @@ template<>              struct product_type_selector<Large,Large,Small>  { enum
 
 } // end namespace internal
 
-/** \class ProductReturnType
-  * \ingroup Core_Module
-  *
-  * \brief Helper class to get the correct and optimized returned type of operator*
-  *
-  * \param Lhs the type of the left-hand side
-  * \param Rhs the type of the right-hand side
-  * \param ProductMode the type of the product (determined automatically by internal::product_mode)
-  *
-  * This class defines the typename Type representing the optimized product expression
-  * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type
-  * is the recommended way to define the result type of a function returning an expression
-  * which involve a matrix product. The class Product should never be
-  * used directly.
-  *
-  * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
-  */
-template<typename Lhs, typename Rhs, int ProductType>
-struct ProductReturnType
-{
-  // TODO use the nested type to reduce instanciations ????
-//   typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
-//   typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
-
-  typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type;
-};
-
-template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type;
-};
-
-template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type;
-};
-
-// this is a workaround for sun CC
-template<typename Lhs, typename Rhs>
-struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{};
-
 /***********************************************************************
 *  Implementation of Inner Vector Vector Product
 ***********************************************************************/
@@ -186,119 +142,10 @@ struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedPr
 // product ends up to a row-vector times col-vector product... To tackle this use
 // case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x);
 
-namespace internal {
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> >
- : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, InnerProduct>
-  : internal::no_assignment_operator,
-    public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1>
-{
-    typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base;
-  public:
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs)
-    {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-      Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
-    }
-
-    /** Convertion to scalar */
-    operator const typename Base::Scalar() const {
-      return Base::coeff(0,0);
-    }
-};
-
 /***********************************************************************
 *  Implementation of Outer Vector Vector Product
 ***********************************************************************/
 
-namespace internal {
-
-// Column major
-template<typename ProductType, typename Dest, typename Func>
-EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const false_type&)
-{
-  typedef typename Dest::Index Index;
-  // FIXME make sure lhs is sequentially stored
-  // FIXME not very good if rhs is real and lhs complex while alpha is real too
-  const Index cols = dest.cols();
-  for (Index j=0; j<cols; ++j)
-    func(dest.col(j), prod.rhs().coeff(0,j) * prod.lhs());
-}
-
-// Row major
-template<typename ProductType, typename Dest, typename Func>
-EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const true_type&) {
-  typedef typename Dest::Index Index;
-  // FIXME make sure rhs is sequentially stored
-  // FIXME not very good if lhs is real and rhs complex while alpha is real too
-  const Index rows = dest.rows();
-  for (Index i=0; i<rows; ++i)
-    func(dest.row(i), prod.lhs().coeff(i,0) * prod.rhs());
-}
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, OuterProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs>
-{
-    template<typename T> struct IsRowMajor : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {};
-    
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
-    {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-    }
-    
-    struct set  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
-    struct add  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } };
-    struct sub  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } };
-    struct adds {
-      Scalar m_scale;
-      adds(const Scalar& s) : m_scale(s) {}
-      template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const {
-        dst.const_cast_derived() += m_scale * src;
-      }
-    };
-    
-    template<typename Dest>
-    inline void evalTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, set(), IsRowMajor<Dest>());
-    }
-    
-    template<typename Dest>
-    inline void addTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, add(), IsRowMajor<Dest>());
-    }
-
-    template<typename Dest>
-    inline void subTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, sub(), IsRowMajor<Dest>());
-    }
-
-    template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const
-    {
-      internal::outer_product_selector_run(*this, dest, adds(alpha), IsRowMajor<Dest>());
-    }
-};
-
 /***********************************************************************
 *  Implementation of General Matrix Vector Product
 ***********************************************************************/
@@ -312,60 +159,13 @@ class GeneralProduct<Lhs, Rhs, OuterProduct>
  */
 namespace internal {
 
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> >
-{};
-
 template<int Side, int StorageOrder, bool BlasCompatible>
-struct gemv_selector;
+struct gemv_dense_sense_selector;
 
 } // end namespace internal
 
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, GemvProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs>
-{
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    typedef typename Lhs::Scalar LhsScalar;
-    typedef typename Rhs::Scalar RhsScalar;
-
-    GeneralProduct(const Lhs& a_lhs, const Rhs& a_rhs) : Base(a_lhs,a_rhs)
-    {
-//       EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value),
-//         YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-    }
-
-    enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
-    typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType;
-
-    template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
-    {
-      eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols());
-      internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
-                       bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha);
-    }
-};
-
 namespace internal {
 
-// The vector is on the left => transposition
-template<int StorageOrder, bool BlasCompatible>
-struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible>
-{
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
-  {
-    Transpose<Dest> destT(dest);
-    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
-    gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
-      ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct>
-        (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha);
-  }
-};
-
 template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if;
 
 template<typename Scalar,int Size,int MaxSize>
@@ -383,7 +183,7 @@ struct gemv_static_vector_if<Scalar,Size,Dynamic,true>
 template<typename Scalar,int Size,int MaxSize>
 struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
 {
-  #if EIGEN_ALIGN_STATICALLY
+  #if EIGEN_MAX_STATIC_ALIGN_BYTES!=0
   internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0> m_data;
   EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; }
   #else
@@ -396,33 +196,48 @@ struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
   internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?PacketSize:0),0> m_data;
   EIGEN_STRONG_INLINE Scalar* data() {
     return ForceAlignment
-            ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(EIGEN_ALIGN_BYTES-1))) + EIGEN_ALIGN_BYTES)
+            ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(EIGEN_MAX_ALIGN_BYTES-1))) + EIGEN_MAX_ALIGN_BYTES)
             : m_data.array;
   }
   #endif
 };
 
-template<> struct gemv_selector<OnTheRight,ColMajor,true>
+// The vector is on the left => transposition
+template<int StorageOrder, bool BlasCompatible>
+struct gemv_dense_sense_selector<OnTheLeft,StorageOrder,BlasCompatible>
 {
-  template<typename ProductType, typename Dest>
-  static inline void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::LhsScalar   LhsScalar;
-    typedef typename ProductType::RhsScalar   RhsScalar;
-    typedef typename ProductType::Scalar      ResScalar;
-    typedef typename ProductType::RealScalar  RealScalar;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
+    Transpose<Dest> destT(dest);
+    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
+    gemv_dense_sense_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
+      ::run(rhs.transpose(), lhs.transpose(), destT, alpha);
+  }
+};
+
+template<> struct gemv_dense_sense_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)
+  {
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    typedef typename Dest::RealScalar  RealScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  
     typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
 
-    ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs());
-    ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs());
+    ActualLhsType actualLhs = LhsBlasTraits::extract(lhs);
+    ActualRhsType actualRhs = RhsBlasTraits::extract(rhs);
 
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
@@ -434,14 +249,14 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
 
     gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
 
-    bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
-    bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
-    
+    const bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
+    const bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+
     RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
 
     ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
                                                   evalToDest ? dest.data() : static_dest.data());
-    
+
     if(!evalToDest)
     {
       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
@@ -457,11 +272,13 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
         MappedDest(actualDestPtr, dest.size()) = dest;
     }
 
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
     general_matrix_vector_product
-      <Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
+        <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
         actualLhs.rows(), actualLhs.cols(),
-        actualLhs.data(), actualLhs.outerStride(),
-        actualRhs.data(), actualRhs.innerStride(),
+        LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+        RhsMapper(actualRhs.data(), actualRhs.innerStride()),
         actualDestPtr, 1,
         compatibleAlpha);
 
@@ -475,34 +292,34 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
   }
 };
 
-template<> struct gemv_selector<OnTheRight,RowMajor,true>
+template<> struct gemv_dense_sense_selector<OnTheRight,RowMajor,true>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename ProductType::LhsScalar LhsScalar;
-    typedef typename ProductType::RhsScalar RhsScalar;
-    typedef typename ProductType::Scalar    ResScalar;
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::_ActualRhsType _ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
 
-    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
-    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
 
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
       // on, the other hand it is good for the cache to pack the vector anyways...
-      DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1
+      DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
     };
 
-    gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+    gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
 
     ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
         DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
@@ -513,42 +330,42 @@ template<> struct gemv_selector<OnTheRight,RowMajor,true>
       Index size = actualRhs.size();
       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
       #endif
-      Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
     }
 
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
     general_matrix_vector_product
-      <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
+        <Index,LhsScalar,LhsMapper,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
         actualLhs.rows(), actualLhs.cols(),
-        actualLhs.data(), actualLhs.outerStride(),
-        actualRhsPtr, 1,
+        LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+        RhsMapper(actualRhsPtr, 1),
         dest.data(), dest.innerStride(),
         actualAlpha);
   }
 };
 
-template<> struct gemv_selector<OnTheRight,ColMajor,false>
+template<> struct gemv_dense_sense_selector<OnTheRight,ColMajor,false>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename Dest::Index Index;
     // TODO makes sure dest is sequentially stored in memory, otherwise use a temp
-    const Index size = prod.rhs().rows();
+    const Index size = rhs.rows();
     for(Index k=0; k<size; ++k)
-      dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k);
+      dest += (alpha*rhs.coeff(k)) * lhs.col(k);
   }
 };
 
-template<> struct gemv_selector<OnTheRight,RowMajor,false>
+template<> struct gemv_dense_sense_selector<OnTheRight,RowMajor,false>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename Dest::Index Index;
     // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp
-    const Index rows = prod.rows();
+    const Index rows = dest.rows();
     for(Index i=0; i<rows; ++i)
-      dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum();
+      dest.coeffRef(i) += alpha * (lhs.row(i).cwiseProduct(rhs.transpose())).sum();
   }
 };
 
@@ -566,7 +383,6 @@ template<> struct gemv_selector<OnTheRight,RowMajor,false>
   */
 #ifndef __CUDACC__
 
-#ifdef EIGEN_TEST_EVALUATORS
 template<typename Derived>
 template<typename OtherDerived>
 inline const Product<Derived, OtherDerived>
@@ -594,42 +410,12 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 #ifdef EIGEN_DEBUG_PRODUCT
   internal::product_type<Derived,OtherDerived>::debug();
 #endif
-  
+
   return Product<Derived, OtherDerived>(derived(), other.derived());
 }
-#else
-template<typename Derived>
-template<typename OtherDerived>
-inline const typename ProductReturnType<Derived, OtherDerived>::Type
-MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
-{
-  // A note regarding the function declaration: In MSVC, this function will sometimes
-  // not be inlined since DenseStorage is an unwindable object for dynamic
-  // matrices and product types are holding a member to store the result.
-  // Thus it does not help tagging this function with EIGEN_STRONG_INLINE.
-  enum {
-    ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
-                   || OtherDerived::RowsAtCompileTime==Dynamic
-                   || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
-    AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
-    SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
-  };
-  // note to the lost user:
-  //    * for a dot product use: v1.dot(v2)
-  //    * for a coeff-wise product use: v1.cwiseProduct(v2)
-  EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
-    INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
-  EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
-    INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
-  EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
-#ifdef EIGEN_DEBUG_PRODUCT
-  internal::product_type<Derived,OtherDerived>::debug();
-#endif
-  return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
-}
-#endif
 
-#endif
+#endif // __CUDACC__
+
 /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
   *
   * The returned product will behave like any other expressions: the coefficients of the product will be
@@ -643,7 +429,7 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
   */
 template<typename Derived>
 template<typename OtherDerived>
-const typename LazyProductReturnType<Derived,OtherDerived>::Type
+const Product<Derived,OtherDerived,LazyProduct>
 MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
 {
   enum {
@@ -662,7 +448,7 @@ MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
     INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
   EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
 
-  return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
+  return Product<Derived,OtherDerived,LazyProduct>(derived(), other.derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/GenericPacketMath.h

@@ -49,16 +49,20 @@ struct default_packet_traits
     HasMul    = 1,
     HasNegate = 1,
     HasAbs    = 1,
+    HasArg    = 0,
     HasAbs2   = 1,
     HasMin    = 1,
     HasMax    = 1,
     HasConj   = 1,
     HasSetLinear = 1,
+    HasBlend  = 0,
 
     HasDiv    = 0,
     HasSqrt   = 0,
+    HasRsqrt  = 0,
     HasExp    = 0,
     HasLog    = 0,
+    HasLog10    = 0,
     HasPow    = 0,
 
     HasSin    = 0,
@@ -66,7 +70,14 @@ struct default_packet_traits
     HasTan    = 0,
     HasASin   = 0,
     HasACos   = 0,
-    HasATan   = 0
+    HasATan   = 0,
+    HasSinh    = 0,
+    HasCosh    = 0,
+    HasTanh    = 0,
+
+    HasRound  = 0,
+    HasFloor  = 0,
+    HasCeil   = 0
   };
 };
 
@@ -94,6 +105,30 @@ template<typename T> struct packet_traits : default_packet_traits
   };
 };
 
+template<typename T> struct packet_traits<const T> : packet_traits<T> { };
+
+template <typename Src, typename Tgt> struct type_casting_traits {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+/** \internal \returns static_cast<TgtType>(a) (coeff-wise) */
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a) {
+  return static_cast<TgtPacket>(a);
+}
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a, const SrcPacket& /*b*/) {
+  return static_cast<TgtPacket>(a);
+}
+
+
 /** \internal \returns a + b (coeff-wise) */
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 padd(const Packet& a,
@@ -126,17 +161,21 @@ pdiv(const Packet& a,
 /** \internal \returns the min of \a a and \a b  (coeff-wise) */
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmin(const Packet& a,
-        const Packet& b) { EIGEN_USING_STD_MATH(min); return (min)(a, b); }
+        const Packet& b) { return numext::mini(a, b); }
 
 /** \internal \returns the max of \a a and \a b  (coeff-wise) */
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmax(const Packet& a,
-        const Packet& b) { EIGEN_USING_STD_MATH(max); return (max)(a, b); }
+        const Packet& b) { return numext::maxi(a, b); }
 
 /** \internal \returns the absolute value of \a a */
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pabs(const Packet& a) { using std::abs; return abs(a); }
 
+/** \internal \returns the phase angle of \a a */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+parg(const Packet& a) { using numext::arg; return arg(a); }
+
 /** \internal \returns the bitwise and of \a a and \a b */
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pand(const Packet& a, const Packet& b) { return a & b; }
@@ -222,8 +261,8 @@ inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
 }
 
 /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */
-template<typename Scalar> inline typename packet_traits<Scalar>::type
-plset(const Scalar& a) { return a; }
+template<typename Packet> inline Packet
+plset(const typename unpacket_traits<Packet>::type& a) { return a; }
 
 /** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */
 template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from)
@@ -233,17 +272,25 @@ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(
 template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from)
 {  (*to) = from; }
 
- template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, DenseIndex /*stride*/)
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, Index /*stride*/)
  { return ploadu<Packet>(from); }
 
- template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, DenseIndex /*stride*/)
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, Index /*stride*/)
  { pstore(to, from); }
 
 /** \internal tries to do cache prefetching of \a addr */
 template<typename Scalar> inline void prefetch(const Scalar* addr)
 {
-#if !defined(_MSC_VER)
-__builtin_prefetch(addr);
+#ifdef __CUDA_ARCH__
+#if defined(__LP64__)
+  // 64-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr));
+#else
+  // 32-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=r"(addr) : "r"(addr));
+#endif
+#elif !EIGEN_COMP_MSVC
+  __builtin_prefetch(addr);
 #endif
 }
 
@@ -284,6 +331,21 @@ template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Pack
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a)
 { return a; }
 
+template<size_t offset, typename Packet>
+struct protate_impl
+{
+  // Empty so attempts to use this unimplemented path will fail to compile.
+  // Only specializations of this template should be used.
+};
+
+/** \internal \returns a packet with the coefficients rotated to the right in little-endian convention,
+  * by the given offset, e.g. for offset == 1:
+  *     (packet[3], packet[2], packet[1], packet[0]) becomes (packet[0], packet[3], packet[2], packet[1])
+  */
+template<size_t offset, typename Packet> EIGEN_DEVICE_FUNC inline Packet protate(const Packet& a)
+{
+  return offset ? protate_impl<offset, Packet>::run(a) : a;
+}
 
 /** \internal \returns \a a with real and imaginary part flipped (for complex type only) */
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a)
@@ -318,10 +380,22 @@ Packet pasin(const Packet& a) { using std::asin; return asin(a); }
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet pacos(const Packet& a) { using std::acos; return acos(a); }
 
-/** \internal \returns the atan of \a a (coeff-wise) */
+/** \internal \returns the arc tangent of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet patan(const Packet& a) { using std::atan; return atan(a); }
 
+/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet psinh(const Packet& a) { using std::sinh; return sinh(a); }
+
+/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pcosh(const Packet& a) { using std::cosh; return cosh(a); }
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet ptanh(const Packet& a) { using std::tanh; return tanh(a); }
+
 /** \internal \returns the exp of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet pexp(const Packet& a) { using std::exp; return exp(a); }
@@ -330,10 +404,32 @@ Packet pexp(const Packet& a) { using std::exp; return exp(a); }
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet plog(const Packet& a) { using std::log; return log(a); }
 
+/** \internal \returns the log10 of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plog10(const Packet& a) { using std::log10; return log10(a); }
+
 /** \internal \returns the square-root of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet psqrt(const Packet& a) { using std::sqrt; return sqrt(a); }
 
+/** \internal \returns the reciprocal square-root of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet prsqrt(const Packet& a) {
+  return pdiv(pset1<Packet>(1), psqrt(a));
+}
+
+/** \internal \returns the rounded value of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pround(const Packet& a) { using numext::round; return round(a); }
+
+/** \internal \returns the floor of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pfloor(const Packet& a) { using numext::floor; return floor(a); }
+
+/** \internal \returns the ceil of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); }
+
 /***************************************************************************
 * The following functions might not have to be overwritten for vectorized types
 ***************************************************************************/
@@ -354,27 +450,38 @@ pmadd(const Packet&  a,
 { return padd(pmul(a, b),c); }
 
 /** \internal \returns a packet version of \a *from.
-  * If LoadMode equals #Aligned, \a from must be 16 bytes aligned */
-template<typename Packet, int LoadMode>
-inline Packet ploadt(const typename unpacket_traits<Packet>::type* from)
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt(const typename unpacket_traits<Packet>::type* from)
 {
-  if(LoadMode == Aligned)
+  if(Alignment >= unpacket_traits<Packet>::alignment)
     return pload<Packet>(from);
   else
     return ploadu<Packet>(from);
 }
 
 /** \internal copy the packet \a from to \a *to.
-  * If StoreMode equals #Aligned, \a to must be 16 bytes aligned */
-template<typename Scalar, typename Packet, int LoadMode>
-inline void pstoret(Scalar* to, const Packet& from)
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Scalar, typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& from)
 {
-  if(LoadMode == Aligned)
+  if(Alignment >= unpacket_traits<Packet>::alignment)
     pstore(to, from);
   else
     pstoreu(to, from);
 }
 
+/** \internal \returns a packet version of \a *from.
+  * Unlike ploadt, ploadt_ro takes advantage of the read-only memory path on the
+  * hardware if available to speedup the loading of data that won't be modified
+  * by the current computation.
+  */
+template<typename Packet, int LoadMode>
+inline Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from)
+{
+  return ploadt<Packet, LoadMode>(from);
+}
+
 /** \internal default implementation of palign() allowing partial specialization */
 template<int Offset,typename PacketType>
 struct palign_impl
@@ -433,6 +540,19 @@ ptranspose(PacketBlock<Packet,1>& /*kernel*/) {
   // Nothing to do in the scalar case, i.e. a 1x1 matrix.
 }
 
+/***************************************************************************
+ * Selector, i.e. vector of N boolean values used to select (i.e. blend)
+ * words from 2 packets.
+***************************************************************************/
+template <size_t N> struct Selector {
+  bool select[N];
+};
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pblend(const Selector<unpacket_traits<Packet>::size>& ifPacket, const Packet& thenPacket, const Packet& elsePacket) {
+  return ifPacket.select[0] ? thenPacket : elsePacket;
+}
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/GlobalFunctions.h

@@ -14,8 +14,8 @@
 #define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR) \
   template<typename Derived> \
   inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \
-  NAME(const Eigen::ArrayBase<Derived>& x) { \
-    return x.derived(); \
+  (NAME)(const Eigen::ArrayBase<Derived>& x) { \
+    return Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived>(x.derived()); \
   }
 
 #define EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(NAME,FUNCTOR) \
@@ -30,26 +30,40 @@
   { \
     static inline typename NAME##_retval<ArrayBase<Derived> >::type run(const Eigen::ArrayBase<Derived>& x) \
     { \
-      return x.derived(); \
+      return typename NAME##_retval<ArrayBase<Derived> >::type(x.derived()); \
     } \
   };
 
-
 namespace Eigen
 {
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(real,scalar_real_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(imag,scalar_imag_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(conj,scalar_conjugate_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(inverse,scalar_inverse_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sin,scalar_sin_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cos,scalar_cos_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(atan,scalar_atan_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op)
+  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(exp,scalar_exp_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs2,scalar_abs2_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(arg,scalar_arg_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sqrt,scalar_sqrt_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(square,scalar_square_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cube,scalar_cube_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(round,scalar_round_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(floor,scalar_floor_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(ceil,scalar_ceil_op)
+  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)
   
   template<typename Derived>
   inline const Eigen::CwiseUnaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar>, const Derived>
@@ -57,16 +71,46 @@ namespace Eigen
     return x.derived().pow(exponent);
   }
 
-  template<typename Derived>
-  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const Derived, const Derived>
-  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<Derived>& exponents) 
+  /** \returns an expression of the coefficient-wise power of \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power.
+    *
+    * Example: \include Cwise_array_power_array.cpp
+    * Output: \verbinclude Cwise_array_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    */
+  template<typename Derived,typename ExponentDerived>
+  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>
+  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<ExponentDerived>& exponents) 
   {
-    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const Derived, const Derived>(
+    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>(
       x.derived(),
       exponents.derived()
     );
   }
   
+  /** \returns an expression of the coefficient-wise power of the scalar \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power between a scalar and an array of exponents.
+    * Beaware that the scalar type of the input scalar \a x and the exponents \a exponents must be the same.
+    *
+    * Example: \include Cwise_scalar_power_array.cpp
+    * Output: \verbinclude Cwise_scalar_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    */
+  template<typename Derived>
+  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const typename Derived::ConstantReturnType, const Derived>
+  pow(const typename Derived::Scalar& x, const Eigen::ArrayBase<Derived>& exponents) 
+  {
+    typename Derived::ConstantReturnType constant_x(exponents.rows(), exponents.cols(), x);
+    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const typename Derived::ConstantReturnType, const Derived>(
+      constant_x,
+      exponents.derived()
+    );
+  }
+  
   /**
   * \brief Component-wise division of a scalar by array elements.
   **/

Eigen/src/Core/IO.h

@@ -164,7 +164,6 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat&
   
   typename Derived::Nested m = _m;
   typedef typename Derived::Scalar Scalar;
-  typedef typename Derived::Index Index;
 
   Index width = 0;
 

Eigen/src/Core/Inverse.h

@@ -0,0 +1,126 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 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_INVERSE_H
+#define EIGEN_INVERSE_H
+
+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 {
+
+template<typename XprType>
+struct traits<Inverse<XprType> >
+  : traits<typename XprType::PlainObject>
+{
+  typedef typename XprType::PlainObject PlainObject;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit
+  };
+};
+
+} // end namespace internal
+
+/** \class Inverse
+  *
+  * \brief Expression of the inverse of another expression
+  *
+  * \tparam XprType the type of the expression we are taking the inverse
+  *
+  * This class represents an abstract expression of A.inverse()
+  * and most of the time this is the only way it is used.
+  *
+  */
+template<typename XprType>
+class Inverse : public InverseImpl<XprType,typename internal::traits<XprType>::StorageKind>
+{
+public:
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename XprType::PlainObject                       PlainObject;
+  typedef typename internal::ref_selector<XprType>::type      XprTypeNested;
+  typedef typename internal::remove_all<XprTypeNested>::type  XprTypeNestedCleaned;
+  
+  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(); }
+
+  EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; }
+
+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> >
+{
+  typedef Inverse<XprType> Derived;
+  
+public:
+  
+  typedef MatrixBase<Derived> Base;
+  EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+  typedef typename internal::remove_all<XprType>::type NestedExpression;
+
+private:
+  
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+namespace internal {
+
+/** \internal
+  * \brief Default evaluator for Inverse expression.
+  * 
+  * This default evaluator for Inverse expression simply evaluate the inverse into a temporary
+  * by a call to internal::call_assignment_no_alias.
+  * Therefore, inverse implementers only have to specialize Assignment<Dst,Inverse<...>, ...> for
+  * there own nested expression.
+  *
+  * \sa class Inverse
+  */
+template<typename ArgType>
+struct unary_evaluator<Inverse<ArgType> >
+  : public evaluator<typename Inverse<ArgType>::PlainObject>
+{
+  typedef Inverse<ArgType> InverseType;
+  typedef typename InverseType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  enum { Flags = Base::Flags | EvalBeforeNestingBit };
+
+  unary_evaluator(const InverseType& inv_xpr)
+    : m_result(inv_xpr.rows(), inv_xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    internal::call_assignment_no_alias(m_result, inv_xpr);
+  }
+  
+protected:
+  PlainObject m_result;
+};
+  
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_INVERSE_H

Eigen/src/Core/Map.h

@@ -19,7 +19,7 @@ namespace Eigen {
   * \brief A matrix or vector expression mapping an existing array of data.
   *
   * \tparam PlainObjectType the equivalent matrix type of the mapped data
-  * \tparam MapOptions specifies whether the pointer is \c #Aligned, or \c #Unaligned.
+  * \tparam MapOptions specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
   *                The default is \c #Unaligned.
   * \tparam StrideType optionally specifies strides. By default, Map assumes the memory layout
   *                   of an ordinary, contiguous array. This can be overridden by specifying strides.
@@ -70,8 +70,6 @@ struct traits<Map<PlainObjectType, MapOptions, StrideType> >
   : public traits<PlainObjectType>
 {
   typedef traits<PlainObjectType> TraitsBase;
-  typedef typename PlainObjectType::Index Index;
-  typedef typename PlainObjectType::Scalar Scalar;
   enum {
     InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
                              ? int(PlainObjectType::InnerStrideAtCompileTime)
@@ -79,22 +77,9 @@ struct traits<Map<PlainObjectType, MapOptions, StrideType> >
     OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
                              ? int(PlainObjectType::OuterStrideAtCompileTime)
                              : int(StrideType::OuterStrideAtCompileTime),
-    HasNoInnerStride = InnerStrideAtCompileTime == 1,
-    HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0,
-    HasNoStride = HasNoInnerStride && HasNoOuterStride,
-    IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned),
-    IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic,
-    KeepsPacketAccess = bool(HasNoInnerStride)
-                        && ( bool(IsDynamicSize)
-                           || HasNoOuterStride
-                           || ( OuterStrideAtCompileTime!=Dynamic
-                           && ((static_cast<int>(sizeof(Scalar))*OuterStrideAtCompileTime)%EIGEN_ALIGN_BYTES)==0 ) ),
+    Alignment = int(MapOptions)&int(AlignedMask),
     Flags0 = TraitsBase::Flags & (~NestByRefBit),
-    Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit),
-    Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime))
-           ? int(Flags1) : int(Flags1 & ~LinearAccessBit),
-    Flags3 = is_lvalue<PlainObjectType>::value ? int(Flags2) : (int(Flags2) & ~LvalueBit),
-    Flags = KeepsPacketAccess ? int(Flags3) : (int(Flags3) & ~PacketAccessBit)
+    Flags = is_lvalue<PlainObjectType>::value ? int(Flags0) : (int(Flags0) & ~LvalueBit)
   };
 private:
   enum { Options }; // Expressions don't have Options
@@ -132,11 +117,11 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the fixed-size case.
       *
       * \param dataPtr pointer to the array to map
-      * \param a_stride optional Stride object, passing the strides.
+      * \param stride optional Stride object, passing the strides.
       */
     EIGEN_DEVICE_FUNC
-    inline Map(PointerArgType dataPtr, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr)), m_stride(a_stride)
+    explicit inline Map(PointerArgType dataPtr, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr)), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -144,12 +129,12 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the dynamic-size vector case.
       *
       * \param dataPtr pointer to the array to map
-      * \param a_size the size of the vector expression
-      * \param a_stride optional Stride object, passing the strides.
+      * \param size the size of the vector expression
+      * \param stride optional Stride object, passing the strides.
       */
     EIGEN_DEVICE_FUNC
-    inline Map(PointerArgType dataPtr, Index a_size, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr), a_size), m_stride(a_stride)
+    inline Map(PointerArgType dataPtr, Index size, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), size), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -157,13 +142,13 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the dynamic-size matrix case.
       *
       * \param dataPtr pointer to the array to map
-      * \param nbRows the number of rows of the matrix expression
-      * \param nbCols the number of columns of the matrix expression
-      * \param a_stride optional Stride object, passing the strides.
+      * \param rows the number of rows of the matrix expression
+      * \param cols the number of columns of the matrix expression
+      * \param stride optional Stride object, passing the strides.
       */
     EIGEN_DEVICE_FUNC
-    inline Map(PointerArgType dataPtr, Index nbRows, Index nbCols, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr), nbRows, nbCols), m_stride(a_stride)
+    inline Map(PointerArgType dataPtr, Index rows, Index cols, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), rows, cols), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }

Eigen/src/Core/MapBase.h

@@ -12,7 +12,7 @@
 #define EIGEN_MAPBASE_H
 
 #define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \
-      EIGEN_STATIC_ASSERT((int(internal::traits<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \
+      EIGEN_STATIC_ASSERT((int(internal::evaluator<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \
                           YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT)
 
 namespace Eigen { 
@@ -37,7 +37,6 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     };
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -85,7 +84,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       *
       * \sa innerStride(), outerStride()
       */
-    inline const Scalar* data() const { return m_data; }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_data; }
 
     EIGEN_DEVICE_FUNC
     inline const Scalar& coeff(Index rowId, Index colId) const
@@ -128,7 +127,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     }
 
     EIGEN_DEVICE_FUNC
-    inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
+    explicit inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
     {
       EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
       checkSanity();
@@ -147,12 +146,12 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     }
 
     EIGEN_DEVICE_FUNC
-    inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols)
-            : m_data(dataPtr), m_rows(nbRows), m_cols(nbCols)
+    inline MapBase(PointerType dataPtr, Index rows, Index cols)
+            : m_data(dataPtr), m_rows(rows), m_cols(cols)
     {
       eigen_assert( (dataPtr == 0)
-              || (   nbRows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == nbRows)
-                  && nbCols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == nbCols)));
+              || (   rows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+                  && cols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols)));
       checkSanity();
     }
 
@@ -161,11 +160,9 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     EIGEN_DEVICE_FUNC
     void checkSanity() const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(internal::traits<Derived>::Flags&PacketAccessBit,
-                                        internal::inner_stride_at_compile_time<Derived>::ret==1),
-                          PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
-      eigen_assert(EIGEN_IMPLIES(internal::traits<Derived>::Flags&AlignedBit, (size_t(m_data) % EIGEN_ALIGN_BYTES) == 0)
-                   && "data is not aligned");
+#if EIGEN_MAX_ALIGN_BYTES>0
+      eigen_assert(((size_t(m_data) % EIGEN_PLAIN_ENUM_MAX(1,internal::traits<Derived>::Alignment)) == 0) && "data is not aligned");
+#endif
     }
 
     PointerType m_data;
@@ -176,13 +173,14 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
 template<typename Derived> class MapBase<Derived, WriteAccessors>
   : public MapBase<Derived, ReadOnlyAccessors>
 {
+    typedef MapBase<Derived, ReadOnlyAccessors> ReadOnlyMapBase;
   public:
 
     typedef MapBase<Derived, ReadOnlyAccessors> Base;
 
     typedef typename Base::Scalar Scalar;
     typedef typename Base::PacketScalar PacketScalar;
-    typedef typename Base::Index Index;
+    typedef typename Base::StorageIndex StorageIndex;
     typedef typename Base::PointerType PointerType;
 
     using Base::derived;
@@ -238,16 +236,18 @@ template<typename Derived> class MapBase<Derived, WriteAccessors>
 
     EIGEN_DEVICE_FUNC explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {}
     EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {}
-    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols) : Base(dataPtr, nbRows, nbCols) {}
+    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index rows, Index cols) : Base(dataPtr, rows, cols) {}
 
     EIGEN_DEVICE_FUNC
     Derived& operator=(const MapBase& other)
     {
-      Base::Base::operator=(other);
+      ReadOnlyMapBase::Base::operator=(other);
       return derived();
     }
 
-    using Base::Base::operator=;
+    // In theory we could simply refer to Base:Base::operator=, but MSVC does not like Base::Base,
+    // see bugs 821 and 920.
+    using ReadOnlyMapBase::Base::operator=;
 };
 
 #undef EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS

Eigen/src/Core/MathFunctions.h

@@ -10,11 +10,14 @@
 #ifndef EIGEN_MATHFUNCTIONS_H
 #define EIGEN_MATHFUNCTIONS_H
 
+// source: http://www.geom.uiuc.edu/~huberty/math5337/groupe/digits.html
+#define EIGEN_PI 3.141592653589793238462643383279502884197169399375105820974944592307816406
+
 namespace Eigen {
 
 // On WINCE, std::abs is defined for int only, so let's defined our own overloads:
 // This issue has been confirmed with MSVC 2008 only, but the issue might exist for more recent versions too.
-#if defined(_WIN32_WCE) && defined(_MSC_VER) && _MSC_VER<=1500
+#if EIGEN_OS_WINCE && EIGEN_COMP_MSVC && EIGEN_COMP_MSVC<=1500
 long        abs(long        x) { return (labs(x));  }
 double      abs(double      x) { return (fabs(x));  }
 float       abs(float       x) { return (fabsf(x)); }
@@ -276,7 +279,7 @@ struct norm1_default_impl
   EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(real(x)) + abs(imag(x));
   }
 };
@@ -287,7 +290,7 @@ struct norm1_default_impl<Scalar, false>
   EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(x);
   }
 };
@@ -313,8 +316,8 @@ struct hypot_impl
   {
     EIGEN_USING_STD_MATH(max);
     EIGEN_USING_STD_MATH(min);
-    using std::abs;
-    using std::sqrt;
+    EIGEN_USING_STD_MATH(abs);
+    EIGEN_USING_STD_MATH(sqrt);
     RealScalar _x = abs(x);
     RealScalar _y = abs(y);
     Scalar p, qp;
@@ -328,6 +331,7 @@ struct hypot_impl
       p = _y;
       qp = _x / p;
     }
+    if(p==RealScalar(0)) return RealScalar(0);
     return p * sqrt(RealScalar(1) + qp*qp);
   }
 };
@@ -345,6 +349,7 @@ struct hypot_retval
 template<typename OldType, typename NewType>
 struct cast_impl
 {
+  EIGEN_DEVICE_FUNC
   static inline NewType run(const OldType& x)
   {
     return static_cast<NewType>(x);
@@ -354,56 +359,121 @@ struct cast_impl
 // here, for once, we're plainly returning NewType: we don't want cast to do weird things.
 
 template<typename OldType, typename NewType>
+EIGEN_DEVICE_FUNC
 inline NewType cast(const OldType& x)
 {
   return cast_impl<OldType, NewType>::run(x);
 }
 
 /****************************************************************************
-* Implementation of atanh2                                                *
+* Implementation of round                                                   *
 ****************************************************************************/
 
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct round_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      using std::round;
+      return round(x);
+    }
+  };
+#else
+  template<typename Scalar>
+  struct round_impl
+  {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      EIGEN_USING_STD_MATH(floor);
+      EIGEN_USING_STD_MATH(ceil);
+      return (x > Scalar(0)) ? floor(x + Scalar(0.5)) : ceil(x - Scalar(0.5));
+    }
+  };
+#endif
+
 template<typename Scalar>
-struct atanh2_impl
+struct round_retval
 {
-  static inline Scalar run(const Scalar& x, const Scalar& r)
+  typedef Scalar type;
+};
+
+/****************************************************************************
+* Implementation of arg                                                     *
+****************************************************************************/
+
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct arg_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+#else
+  template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
+  struct arg_default_impl
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      return (x < Scalar(0)) ? Scalar(EIGEN_PI) : Scalar(0); }
+  };
+
+  template<typename Scalar>
+  struct arg_default_impl<Scalar,true>
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+
+  template<typename Scalar> struct arg_impl : arg_default_impl<Scalar> {};
+#endif
+
+template<typename Scalar>
+struct arg_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of log1p                                                   *
+****************************************************************************/
+template<typename Scalar, bool isComplex = NumTraits<Scalar>::IsComplex >
+struct log1p_impl
+{
+  static inline Scalar run(const Scalar& x)
   {
     EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
-    #if (__cplusplus >= 201103L) && !defined(__CYGWIN__)
-      using std::log1p;
-      return log1p(2 * x / (r - x)) / 2;
-    #else
-      using std::abs;
-      using std::log;
-      using std::sqrt;
-      Scalar z = x / r;
-      if (r == 0 || abs(z) > sqrt(NumTraits<Scalar>::epsilon()))
-        return log((r + x) / (r - x)) / 2;
-      else
-        return z + z*z*z / 3;
-    #endif
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_USING_STD_MATH(log);
+    Scalar x1p = RealScalar(1) + x;
+    return ( x1p == Scalar(1) ) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) );
   }
 };
 
-template<typename RealScalar>
-struct atanh2_impl<std::complex<RealScalar> >
-{
-  typedef std::complex<RealScalar> Scalar;
-  static inline Scalar run(const Scalar& x, const Scalar& r)
+#if EIGEN_HAS_CXX11_MATH
+template<typename Scalar>
+struct log1p_impl<Scalar, false> {
+  static inline Scalar run(const Scalar& x)
   {
-    using std::log;
-    using std::norm;
-    using std::sqrt;
-    Scalar z = x / r;
-    if (r == Scalar(0) || norm(z) > NumTraits<RealScalar>::epsilon())
-      return RealScalar(0.5) * log((r + x) / (r - x));
-    else
-      return z + z*z*z / RealScalar(3);
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    using std::log1p;
+    return log1p(x);
   }
 };
+#endif
 
 template<typename Scalar>
-struct atanh2_retval
+struct log1p_retval
 {
   typedef Scalar type;
 };
@@ -418,7 +488,7 @@ struct pow_default_impl
   typedef Scalar retval;
   static inline Scalar run(const Scalar& x, const Scalar& y)
   {
-    using std::pow;
+    EIGEN_USING_STD_MATH(pow);
     return pow(x, y);
   }
 };
@@ -486,48 +556,48 @@ struct random_default_impl<Scalar, false, false>
 };
 
 enum {
-  floor_log2_terminate,
-  floor_log2_move_up,
-  floor_log2_move_down,
-  floor_log2_bogus
+  meta_floor_log2_terminate,
+  meta_floor_log2_move_up,
+  meta_floor_log2_move_down,
+  meta_floor_log2_bogus
 };
 
-template<unsigned int n, int lower, int upper> struct floor_log2_selector
+template<unsigned int n, int lower, int upper> struct meta_floor_log2_selector
 {
   enum { middle = (lower + upper) / 2,
-         value = (upper <= lower + 1) ? int(floor_log2_terminate)
-               : (n < (1 << middle)) ? int(floor_log2_move_down)
-               : (n==0) ? int(floor_log2_bogus)
-               : int(floor_log2_move_up)
+         value = (upper <= lower + 1) ? int(meta_floor_log2_terminate)
+               : (n < (1 << middle)) ? int(meta_floor_log2_move_down)
+               : (n==0) ? int(meta_floor_log2_bogus)
+               : int(meta_floor_log2_move_up)
   };
 };
 
 template<unsigned int n,
          int lower = 0,
          int upper = sizeof(unsigned int) * CHAR_BIT - 1,
-         int selector = floor_log2_selector<n, lower, upper>::value>
-struct floor_log2 {};
+         int selector = meta_floor_log2_selector<n, lower, upper>::value>
+struct meta_floor_log2 {};
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_move_down>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_down>
 {
-  enum { value = floor_log2<n, lower, floor_log2_selector<n, lower, upper>::middle>::value };
+  enum { value = meta_floor_log2<n, lower, meta_floor_log2_selector<n, lower, upper>::middle>::value };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_move_up>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_up>
 {
-  enum { value = floor_log2<n, floor_log2_selector<n, lower, upper>::middle, upper>::value };
+  enum { value = meta_floor_log2<n, meta_floor_log2_selector<n, lower, upper>::middle, upper>::value };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_terminate>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_terminate>
 {
   enum { value = (n >= ((unsigned int)(1) << (lower+1))) ? lower+1 : lower };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_bogus>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_bogus>
 {
   // no value, error at compile time
 };
@@ -535,11 +605,24 @@ struct floor_log2<n, lower, upper, floor_log2_bogus>
 template<typename Scalar>
 struct random_default_impl<Scalar, false, true>
 {
-  typedef typename NumTraits<Scalar>::NonInteger NonInteger;
-
   static inline Scalar run(const Scalar& x, const Scalar& y)
-  {
-    return x + Scalar((NonInteger(y)-x+1) * std::rand() / (RAND_MAX + NonInteger(1)));
+  { 
+    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;
+    std::size_t range = ScalarX(y)-ScalarX(x);
+    std::size_t offset = 0;
+    // rejection sampling
+    std::size_t divisor    = (range+RAND_MAX-1)/(range+1);
+    std::size_t multiplier = (range+RAND_MAX-1)/std::size_t(RAND_MAX);
+
+    do {
+      offset = ( (std::size_t(std::rand()) * multiplier) / divisor );
+    } while (offset > range);
+
+    return Scalar(ScalarX(x) + offset);
   }
 
   static inline Scalar run()
@@ -547,7 +630,7 @@ struct random_default_impl<Scalar, false, true>
 #ifdef EIGEN_MAKING_DOCS
     return run(Scalar(NumTraits<Scalar>::IsSigned ? -10 : 0), Scalar(10));
 #else
-    enum { rand_bits = floor_log2<(unsigned int)(RAND_MAX)+1>::value,
+    enum { rand_bits = meta_floor_log2<(unsigned int)(RAND_MAX)+1>::value,
            scalar_bits = sizeof(Scalar) * CHAR_BIT,
            shift = EIGEN_PLAIN_ENUM_MAX(0, int(rand_bits) - int(scalar_bits)),
            offset = NumTraits<Scalar>::IsSigned ? (1 << (EIGEN_PLAIN_ENUM_MIN(rand_bits,scalar_bits)-1)) : 0
@@ -587,11 +670,55 @@ inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random()
 } // end namespace internal
 
 /****************************************************************************
-* Generic math function                                                    *
+* Generic math functions                                                    *
 ****************************************************************************/
 
 namespace numext {
 
+#ifndef __CUDA_ARCH__
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(min);
+  return min EIGEN_NOT_A_MACRO (x,y);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(max);
+  return max EIGEN_NOT_A_MACRO (x,y);
+}
+#else
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  return y < x ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
+{
+  return fmin(x, y);
+}
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  return x < y ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y)
+{
+  return fmax(x, y);
+}
+#endif
+
+
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(real, Scalar) real(const Scalar& x)
@@ -620,6 +747,13 @@ inline EIGEN_MATHFUNC_RETVAL(imag, Scalar) imag(const Scalar& x)
   return EIGEN_MATHFUNC_IMPL(imag, Scalar)::run(x);
 }
 
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(arg, Scalar) arg(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(arg, Scalar)::run(x);
+}
+
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
 inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) >::type imag_ref(const Scalar& x)
@@ -664,9 +798,9 @@ inline EIGEN_MATHFUNC_RETVAL(hypot, Scalar) hypot(const Scalar& x, const Scalar&
 
 template<typename Scalar>
 EIGEN_DEVICE_FUNC
-inline EIGEN_MATHFUNC_RETVAL(atanh2, Scalar) atanh2(const Scalar& x, const Scalar& y)
+inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x)
 {
-  return EIGEN_MATHFUNC_IMPL(atanh2, Scalar)::run(x, y);
+  return EIGEN_MATHFUNC_IMPL(log1p, Scalar)::run(x);
 }
 
 template<typename Scalar>
@@ -676,22 +810,96 @@ inline EIGEN_MATHFUNC_RETVAL(pow, Scalar) pow(const Scalar& x, const Scalar& y)
   return EIGEN_MATHFUNC_IMPL(pow, Scalar)::run(x, y);
 }
 
-// std::isfinite is non standard, so let's define our own version,
-// even though it is not very efficient.
 template<typename T>
 EIGEN_DEVICE_FUNC
 bool (isfinite)(const T& x)
 {
-  return x<NumTraits<T>::highest() && x>NumTraits<T>::lowest();
+  #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)
 {
-  using std::real;
-  using std::imag;
-  return isfinite(real(x)) && isfinite(imag(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 Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(round, Scalar) round(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(round, Scalar)::run(x);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (floor)(const T& x)
+{
+  EIGEN_USING_STD_MATH(floor);
+  return floor(x);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (ceil)(const T& x)
+{
+  EIGEN_USING_STD_MATH(ceil);
+  return ceil(x);
+}
+
+// Log base 2 for 32 bits positive integers.
+// Conveniently returns 0 for x==0.
+inline int log2(int x)
+{
+  eigen_assert(x>=0);
+  unsigned int v(x);
+  static const int table[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+  v |= v >> 1;
+  v |= v >> 2;
+  v |= v >> 4;
+  v |= v >> 8;
+  v |= v >> 16;
+  return table[(v * 0x07C4ACDDU) >> 27];
 }
 
 } // end namespace numext
@@ -714,14 +922,14 @@ struct scalar_fuzzy_default_impl<Scalar, false, false>
   template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(x) <= abs(y) * prec;
   }
   EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
     EIGEN_USING_STD_MATH(min);
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(x - y) <= (min)(abs(x), abs(y)) * prec;
   }
   EIGEN_DEVICE_FUNC

Eigen/src/Core/Matrix.h

@@ -24,13 +24,13 @@ namespace Eigen {
   * The %Matrix class encompasses \em both fixed-size and dynamic-size objects (\ref fixedsize "note").
   *
   * The first three template parameters are required:
-  * \tparam _Scalar \anchor matrix_tparam_scalar Numeric type, e.g. float, double, int or std::complex<float>.
-  *                 User defined sclar types are supported as well (see \ref user_defined_scalars "here").
+  * \tparam _Scalar Numeric type, e.g. float, double, int or std::complex<float>.
+  *                 User defined scalar types are supported as well (see \ref user_defined_scalars "here").
   * \tparam _Rows Number of rows, or \b Dynamic
   * \tparam _Cols Number of columns, or \b Dynamic
   *
   * The remaining template parameters are optional -- in most cases you don't have to worry about them.
-  * \tparam _Options \anchor matrix_tparam_options A combination of either \b #RowMajor or \b #ColMajor, and of either
+  * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of either
   *                 \b #AutoAlign or \b #DontAlign.
   *                 The former controls \ref TopicStorageOrders "storage order", and defaults to column-major. The latter controls alignment, which is required
   *                 for vectorization. It defaults to aligning matrices except for fixed sizes that aren't a multiple of the packet size.
@@ -97,6 +97,40 @@ namespace Eigen {
   * are the dimensions of the original matrix, while _Rows and _Cols are Dynamic.</dd>
   * </dl>
   *
+  * <i><b>ABI and storage layout</b></i>
+  * 
+  * The table below summarizes the ABI of some possible Matrix instances which is fixed thorough the lifetime of Eigen 3.
+  * <table  class="manual">
+  * <tr><th>Matrix type</th><th>Equivalent C structure</th></tr>
+  * <tr><td>\code Matrix<T,Dynamic,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code
+  * Matrix<T,Dynamic,1>
+  * Matrix<T,1,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index size;
+  *  };
+  * \endcode</td></tr>
+  * <tr><td>\code Matrix<T,Rows,Cols> \endcode</td><td>\code
+  * struct {
+  *   T data[Rows*Cols];        // with (size_t(data)%A(Rows*Cols*sizeof(T)))==0
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code Matrix<T,Dynamic,Dynamic,0,MaxRows,MaxCols> \endcode</td><td>\code
+  * struct {
+  *   T data[MaxRows*MaxCols];  // with (size_t(data)%A(MaxRows*MaxCols*sizeof(T)))==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * </table>
+  * Note that in this table Rows, Cols, MaxRows and MaxCols are all positive integers. A(S) is defined to the largest possible power-of-two
+  * smaller to EIGEN_MAX_STATIC_ALIGN_BYTES.
+  * 
   * \see MatrixBase for the majority of the API methods for matrices, \ref TopicClassHierarchy, 
   * \ref TopicStorageOrders 
   */
@@ -105,9 +139,23 @@ namespace internal {
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
 struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
 {
+private:
+  enum { size = internal::size_at_compile_time<_Rows,_Cols>::ret };
+  typedef typename find_best_packet<_Scalar,size>::type PacketScalar;
+  enum {
+      row_major_bit = _Options&RowMajor ? RowMajorBit : 0,
+      is_dynamic_size_storage = _MaxRows==Dynamic || _MaxCols==Dynamic,
+      max_size = is_dynamic_size_storage ? Dynamic : _MaxRows*_MaxCols,
+      default_alignment = compute_default_alignment<_Scalar,max_size>::value,
+      actual_alignment = ((_Options&DontAlign)==0) ? default_alignment : 0,
+      required_alignment = unpacket_traits<PacketScalar>::alignment,
+      packet_access_bit = packet_traits<_Scalar>::Vectorizable && (actual_alignment>=required_alignment) ? PacketAccessBit : 0
+    };
+    
+public:
   typedef _Scalar Scalar;
   typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef Eigen::Index StorageIndex;
   typedef MatrixXpr XprKind;
   enum {
     RowsAtCompileTime = _Rows,
@@ -115,10 +163,13 @@ struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
     MaxRowsAtCompileTime = _MaxRows,
     MaxColsAtCompileTime = _MaxCols,
     Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret,
-    CoeffReadCost = NumTraits<Scalar>::ReadCost,
     Options = _Options,
     InnerStrideAtCompileTime = 1,
-    OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime
+    OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime,
+    
+    // FIXME, the following flag in only used to define NeedsToAlign in PlainObjectBase
+    EvaluatorFlags = LinearAccessBit | DirectAccessBit | packet_access_bit | row_major_bit,
+    Alignment = actual_alignment
   };
 };
 }
@@ -169,7 +220,7 @@ class Matrix
       */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Matrix& operator=(const MatrixBase<OtherDerived>& other)
+    EIGEN_STRONG_INLINE Matrix& operator=(const DenseBase<OtherDerived>& other)
     {
       return Base::_set(other);
     }
@@ -213,11 +264,12 @@ class Matrix
 
     // FIXME is it still needed
     EIGEN_DEVICE_FUNC
-    Matrix(internal::constructor_without_unaligned_array_assert)
+    explicit Matrix(internal::constructor_without_unaligned_array_assert)
       : Base(internal::constructor_without_unaligned_array_assert())
     { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED }
 
 #ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
     Matrix(Matrix&& other)
       : Base(std::move(other))
     {
@@ -225,6 +277,7 @@ class Matrix
       if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
         Base::_set_noalias(other);
     }
+    EIGEN_DEVICE_FUNC
     Matrix& operator=(Matrix&& other)
     {
       other.swap(*this);
@@ -263,8 +316,8 @@ class Matrix
       * 
       * \warning This constructor is disabled for fixed-size \c 1x1 matrices. For instance,
       * calling Matrix<double,1,1>(1) will call the initialization constructor: Matrix(const Scalar&).
-      * For fixed-size \c 1x1 matrices it is thefore recommended to use the default
-      * constructor Matrix() instead, especilly when using one of the non standard
+      * For fixed-size \c 1x1 matrices it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
       * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
       */
     EIGEN_STRONG_INLINE explicit Matrix(Index dim);
@@ -278,8 +331,8 @@ class Matrix
       * 
       * \warning This constructor is disabled for fixed-size \c 1x2 and \c 2x1 vectors. For instance,
       * calling Matrix2f(2,1) will call the initialization constructor: Matrix(const Scalar& x, const Scalar& y).
-      * For fixed-size \c 1x2 or \c 2x1 vectors it is thefore recommended to use the default
-      * constructor Matrix() instead, especilly when using one of the non standard
+      * For fixed-size \c 1x2 or \c 2x1 vectors it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
       * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
       */
     EIGEN_DEVICE_FUNC
@@ -312,37 +365,10 @@ class Matrix
     }
 
 
-    /** \brief Constructor copying the value of the expression \a other */
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Matrix(const MatrixBase<OtherDerived>& other)
-             : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      // This test resides here, to bring the error messages closer to the user. Normally, these checks
-      // are performed deeply within the library, thus causing long and scary error traces.
-      EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
     /** \brief Copy constructor */
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Matrix(const Matrix& other)
-            : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
-    /** \brief Copy constructor with in-place evaluation */
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE Matrix(const ReturnByValue<OtherDerived>& other)
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      other.evalTo(*this);
-    }
+    EIGEN_STRONG_INLINE Matrix(const Matrix& other) : Base(other)
+    { }
 
     /** \brief Copy constructor for generic expressions.
       * \sa MatrixBase::operator=(const EigenBase<OtherDerived>&)
@@ -350,23 +376,8 @@ class Matrix
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const EigenBase<OtherDerived> &other)
-      : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
-    {
-      Base::_check_template_params();
-      Base::_resize_to_match(other);
-      // FIXME/CHECK: isn't *this = other.derived() more efficient. it allows to
-      //              go for pure _set() implementations, right?
-      *this = other;
-    }
-
-    /** \internal
-      * \brief Override MatrixBase::swap() since for dynamic-sized matrices
-      * of same type it is enough to swap the data pointers.
-      */
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    void swap(MatrixBase<OtherDerived> const & other)
-    { this->_swap(other.derived()); }
+      : Base(other.derived())
+    { }
 
     EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
     EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }

Eigen/src/Core/MatrixBase.h

@@ -52,7 +52,7 @@ template<typename Derived> class MatrixBase
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef MatrixBase StorageBaseType;
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -66,8 +66,7 @@ template<typename Derived> class MatrixBase
     using Base::MaxSizeAtCompileTime;
     using Base::IsVectorAtCompileTime;
     using Base::Flags;
-    using Base::CoeffReadCost;
-
+    
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -82,6 +81,7 @@ template<typename Derived> class MatrixBase
     using Base::operator*=;
     using Base::operator/=;
     using Base::operator*;
+    using Base::operator/;
 
     typedef typename Base::CoeffReturnType CoeffReturnType;
     typedef typename Base::ConstTransposeReturnType ConstTransposeReturnType;
@@ -102,23 +102,11 @@ template<typename Derived> class MatrixBase
     EIGEN_DEVICE_FUNC
     inline Index diagonalSize() const { return (std::min)(rows(),cols()); }
 
-    /** \brief The plain matrix type corresponding to this expression.
-      *
-      * This is not necessarily exactly the return type of eval(). In the case of plain matrices,
-      * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed
-      * that the return type of eval() is either PlainObject or const PlainObject&.
-      */
-    typedef Matrix<typename internal::traits<Derived>::Scalar,
-                internal::traits<Derived>::RowsAtCompileTime,
-                internal::traits<Derived>::ColsAtCompileTime,
-                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
-                internal::traits<Derived>::MaxRowsAtCompileTime,
-                internal::traits<Derived>::MaxColsAtCompileTime
-          > PlainObject;
+    typedef typename Base::PlainObject PlainObject;
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
     /** \internal the return type of MatrixBase::adjoint() */
     typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
                         CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>,
@@ -127,7 +115,7 @@ template<typename Derived> class MatrixBase
     /** \internal Return type of eigenvalues() */
     typedef Matrix<std::complex<RealScalar>, internal::traits<Derived>::ColsAtCompileTime, 1, ColMajor> EigenvaluesReturnType;
     /** \internal the return type of identity */
-    typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,Derived> IdentityReturnType;
+    typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,PlainObject> IdentityReturnType;
     /** \internal the return type of unit vectors */
     typedef Block<const CwiseNullaryOp<internal::scalar_identity_op<Scalar>, SquareMatrixType>,
                   internal::traits<Derived>::RowsAtCompileTime,
@@ -165,12 +153,6 @@ template<typename Derived> class MatrixBase
     EIGEN_DEVICE_FUNC
     Derived& operator=(const ReturnByValue<OtherDerived>& other);
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_DEVICE_FUNC
-    Derived& lazyAssign(const ProductBase<ProductDerived, Lhs,Rhs>& other);
-#endif // not EIGEN_PARSED_BY_DOXYGEN
-
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     Derived& operator+=(const MatrixBase<OtherDerived>& other);
@@ -181,26 +163,20 @@ template<typename Derived> class MatrixBase
 #ifdef __CUDACC__
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    const typename LazyProductReturnType<Derived,OtherDerived>::Type
+    const Product<Derived,OtherDerived,LazyProduct>
     operator*(const MatrixBase<OtherDerived> &other) const
     { return this->lazyProduct(other); }
 #else
 
-#ifdef EIGEN_TEST_EVALUATORS
     template<typename OtherDerived>
     const Product<Derived,OtherDerived>
     operator*(const MatrixBase<OtherDerived> &other) const;
-#else
-    template<typename OtherDerived>
-    const typename ProductReturnType<Derived,OtherDerived>::Type
-    operator*(const MatrixBase<OtherDerived> &other) const;
-#endif
 
 #endif
 
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC 
-    const typename LazyProductReturnType<Derived,OtherDerived>::Type
+    const Product<Derived,OtherDerived,LazyProduct>
     lazyProduct(const MatrixBase<OtherDerived> &other) const;
 
     template<typename OtherDerived>
@@ -214,7 +190,7 @@ template<typename Derived> class MatrixBase
 
     template<typename DiagonalDerived>
     EIGEN_DEVICE_FUNC
-    const DiagonalProduct<Derived, DiagonalDerived, OnTheRight>
+    const Product<Derived, DiagonalDerived, LazyProduct>
     operator*(const DiagonalBase<DiagonalDerived> &diagonal) const;
 
     template<typename OtherDerived>
@@ -251,19 +227,14 @@ template<typename Derived> class MatrixBase
     template<int Index>
     EIGEN_DEVICE_FUNC
     typename ConstDiagonalIndexReturnType<Index>::Type diagonal() const;
-
-    // Note: The "MatrixBase::" prefixes are added to help MSVC9 to match these declarations with the later implementations.
-    // On the other hand they confuse MSVC8...
-    #if (defined _MSC_VER) && (_MSC_VER >= 1500) // 2008 or later
-    typename MatrixBase::template DiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index);
-    typename MatrixBase::template ConstDiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index) const;
-    #else
-    EIGEN_DEVICE_FUNC
-    typename DiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index);
     
+    typedef Diagonal<Derived,DynamicIndex> DiagonalDynamicIndexReturnType;
+    typedef typename internal::add_const<Diagonal<const Derived,DynamicIndex> >::type ConstDiagonalDynamicIndexReturnType;
+
     EIGEN_DEVICE_FUNC
-    typename ConstDiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index) const;
-    #endif
+    DiagonalDynamicIndexReturnType diagonal(Index index);
+    EIGEN_DEVICE_FUNC
+    ConstDiagonalDynamicIndexReturnType diagonal(Index index) const;
 
     template<unsigned int Mode> struct TriangularViewReturnType { typedef TriangularView<Derived, Mode> Type; };
     template<unsigned int Mode> struct ConstTriangularViewReturnType { typedef const TriangularView<const Derived, Mode> Type; };
@@ -334,12 +305,14 @@ template<typename Derived> class MatrixBase
 
     NoAlias<Derived,Eigen::MatrixBase > noalias();
 
-    inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
-    inline ForceAlignedAccess<Derived> forceAlignedAccess();
-    template<bool Enable> inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type forceAlignedAccessIf() const;
-    template<bool Enable> inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
+    // TODO forceAlignedAccess is temporarily disabled
+    // Need to find a nicer workaround.
+    inline const Derived& forceAlignedAccess() const { return derived(); }
+    inline Derived& forceAlignedAccess() { return derived(); }
+    template<bool Enable> inline const Derived& forceAlignedAccessIf() const { return derived(); }
+    template<bool Enable> inline Derived& forceAlignedAccessIf() { return derived(); }
 
-    Scalar trace() const;
+    EIGEN_DEVICE_FUNC Scalar trace() const;
 
     template<int p> EIGEN_DEVICE_FUNC RealScalar lpNorm() const;
 
@@ -348,10 +321,10 @@ template<typename Derived> class MatrixBase
 
     /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix
       * \sa ArrayBase::matrix() */
-    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper<Derived> array() { return derived(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper<Derived> array() { return ArrayWrapper<Derived>(derived()); }
     /** \returns a const \link Eigen::ArrayBase Array \endlink expression of this matrix
       * \sa ArrayBase::matrix() */
-    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper<const Derived> array() const { return derived(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper<const Derived> array() const { return ArrayWrapper<const Derived>(derived()); }
 
 /////////// LU module ///////////
 
@@ -361,7 +334,8 @@ template<typename Derived> class MatrixBase
     const PartialPivLU<PlainObject> lu() const;
 
     EIGEN_DEVICE_FUNC
-    const internal::inverse_impl<Derived> inverse() const;
+    const Inverse<Derived> inverse() const;
+    
     template<typename ResultType>
     void computeInverseAndDetWithCheck(
       ResultType& inverse,
@@ -394,6 +368,7 @@ template<typename Derived> class MatrixBase
 /////////// SVD module ///////////
 
     JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const;
+    BDCSVD<PlainObject>    bdcSvd(unsigned int computationOptions = 0) const;
 
 /////////// Geometry module ///////////
 
@@ -420,7 +395,8 @@ template<typename Derived> class MatrixBase
     
     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?Vertical:Horizontal };
+    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;
     

Eigen/src/Core/NestByValue.h

@@ -40,29 +40,29 @@ template<typename ExpressionType> class NestByValue
     typedef typename internal::dense_xpr_base<NestByValue>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(NestByValue)
 
-    inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {}
+    EIGEN_DEVICE_FUNC explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {}
 
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
 
-    inline Scalar& coeffRef(Index row, Index col)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
     {
       return m_expression.const_cast_derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
-    inline Scalar& coeffRef(Index index)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
@@ -91,7 +91,7 @@ template<typename ExpressionType> class NestByValue
       m_expression.const_cast_derived().template writePacket<LoadMode>(index, x);
     }
 
-    operator const ExpressionType&() const { return m_expression; }
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
 
   protected:
     const ExpressionType m_expression;

Eigen/src/Core/NoAlias.h

@@ -30,68 +30,35 @@ namespace Eigen {
 template<typename ExpressionType, template <typename> class StorageBase>
 class NoAlias
 {
-    typedef typename ExpressionType::Scalar Scalar;
   public:
-    NoAlias(ExpressionType& expression) : m_expression(expression) {}
-
-    /** Behaves like MatrixBase::lazyAssign(other)
-      * \sa MatrixBase::lazyAssign() */
+    typedef typename ExpressionType::Scalar Scalar;
+    
+    explicit NoAlias(ExpressionType& expression) : m_expression(expression) {}
+    
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase<OtherDerived>& other)
-    { return internal::assign_selector<ExpressionType,OtherDerived,false>::run(m_expression,other.derived()); }
-
-    /** \sa MatrixBase::operator+= */
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::assign_op<Scalar>());
+      return m_expression;
+    }
+    
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase<OtherDerived>& other)
     {
-      typedef SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, ExpressionType, OtherDerived> SelfAdder;
-      SelfAdder tmp(m_expression);
-      typedef typename internal::nested<OtherDerived>::type OtherDerivedNested;
-      typedef typename internal::remove_all<OtherDerivedNested>::type _OtherDerivedNested;
-      internal::assign_selector<SelfAdder,_OtherDerivedNested,false>::run(tmp,OtherDerivedNested(other.derived()));
+      call_assignment_no_alias(m_expression, other.derived(), internal::add_assign_op<Scalar>());
       return m_expression;
     }
-
-    /** \sa MatrixBase::operator-= */
+    
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase<OtherDerived>& other)
     {
-      typedef SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, ExpressionType, OtherDerived> SelfAdder;
-      SelfAdder tmp(m_expression);
-      typedef typename internal::nested<OtherDerived>::type OtherDerivedNested;
-      typedef typename internal::remove_all<OtherDerivedNested>::type _OtherDerivedNested;
-      internal::assign_selector<SelfAdder,_OtherDerivedNested,false>::run(tmp,OtherDerivedNested(other.derived()));
+      call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op<Scalar>());
       return m_expression;
     }
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE ExpressionType& operator+=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    { other.derived().addTo(m_expression); return m_expression; }
-
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE ExpressionType& operator-=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    { other.derived().subTo(m_expression); return m_expression; }
-
-    template<typename Lhs, typename Rhs, int NestingFlags>
-    EIGEN_STRONG_INLINE ExpressionType& operator+=(const CoeffBasedProduct<Lhs,Rhs,NestingFlags>& other)
-    { return m_expression.derived() += CoeffBasedProduct<Lhs,Rhs,NestByRefBit>(other.lhs(), other.rhs()); }
-
-    template<typename Lhs, typename Rhs, int NestingFlags>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE ExpressionType& operator-=(const CoeffBasedProduct<Lhs,Rhs,NestingFlags>& other)
-    { return m_expression.derived() -= CoeffBasedProduct<Lhs,Rhs,NestByRefBit>(other.lhs(), other.rhs()); }
-    
-    template<typename OtherDerived>
-    ExpressionType& operator=(const ReturnByValue<OtherDerived>& func)
-    { return m_expression = func; }
-#endif
-
     EIGEN_DEVICE_FUNC
     ExpressionType& expression() const
     {
@@ -133,7 +100,7 @@ class NoAlias
 template<typename Derived>
 NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias()
 {
-  return derived();
+  return NoAlias<Derived, Eigen::MatrixBase >(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/NumTraits.h

@@ -83,8 +83,25 @@ template<typename T> struct GenericNumTraits
     // make sure to override this for floating-point types
     return Real(0);
   }
-  static inline T highest() { return (std::numeric_limits<T>::max)(); }
-  static inline T lowest()  { return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::max)()); }
+
+
+  EIGEN_DEVICE_FUNC
+  static inline T highest() {
+#if defined(__CUDA_ARCH__)
+    return (internal::device::numeric_limits<T>::max)();
+#else
+    return (std::numeric_limits<T>::max)();
+#endif
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T lowest()  {
+#if defined(__CUDA_ARCH__)
+    return IsInteger ? (internal::device::numeric_limits<T>::min)() : (-(internal::device::numeric_limits<T>::max)());
+#else
+    return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::max)());
+#endif
+  }
 };
 
 template<typename T> struct NumTraits : GenericNumTraits<T>

Eigen/src/Core/PermutationMatrix.h

@@ -13,7 +13,8 @@
 
 namespace Eigen { 
 
-template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKind> class PermutedImpl;
+// 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
@@ -41,10 +42,6 @@ template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKi
 
 namespace internal {
 
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed=false>
-struct permut_matrix_product_retval;
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed=false>
-struct permut_sparsematrix_product_retval;
 enum PermPermProduct_t {PermPermProduct};
 
 } // end namespace internal
@@ -60,19 +57,18 @@ class PermutationBase : public EigenBase<Derived>
     typedef typename Traits::IndicesType IndicesType;
     enum {
       Flags = Traits::Flags,
-      CoeffReadCost = Traits::CoeffReadCost,
       RowsAtCompileTime = Traits::RowsAtCompileTime,
       ColsAtCompileTime = Traits::ColsAtCompileTime,
       MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
       MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
     };
-    typedef typename Traits::StorageIndexType StorageIndexType;
-    typedef typename Traits::Index Index;
-    typedef Matrix<StorageIndexType,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime>
+    typedef typename Traits::StorageIndex StorageIndex;
+    typedef Matrix<StorageIndex,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime>
             DenseMatrixType;
-    typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,StorageIndexType>
+    typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,StorageIndex>
             PlainPermutationType;
     using Base::derived;
+    typedef Transpose<PermutationBase> TransposeReturnType;
     #endif
 
     /** Copies the other permutation into *this */
@@ -118,7 +114,7 @@ class PermutationBase : public EigenBase<Derived>
     void evalTo(MatrixBase<DenseDerived>& other) const
     {
       other.setZero();
-      for (int i=0; i<rows();++i)
+      for (Index i=0; i<rows(); ++i)
         other.coeffRef(indices().coeff(i),i) = typename DenseDerived::Scalar(1);
     }
     #endif
@@ -147,7 +143,8 @@ class PermutationBase : public EigenBase<Derived>
     /** Sets *this to be the identity permutation matrix */
     void setIdentity()
     {
-      for(StorageIndexType i = 0; i < size(); ++i)
+      StorageIndex n = StorageIndex(size());
+      for(StorageIndex i = 0; i < n; ++i)
         indices().coeffRef(i) = i;
     }
 
@@ -163,18 +160,18 @@ class PermutationBase : public EigenBase<Derived>
       *
       * \returns a reference to *this.
       *
-      * \warning This is much slower than applyTranspositionOnTheRight(int,int):
+      * \warning This is much slower than applyTranspositionOnTheRight(Index,Index):
       * this has linear complexity and requires a lot of branching.
       *
-      * \sa applyTranspositionOnTheRight(int,int)
+      * \sa applyTranspositionOnTheRight(Index,Index)
       */
     Derived& applyTranspositionOnTheLeft(Index i, Index j)
     {
       eigen_assert(i>=0 && j>=0 && i<size() && j<size());
       for(Index k = 0; k < size(); ++k)
       {
-        if(indices().coeff(k) == i) indices().coeffRef(k) = StorageIndexType(j);
-        else if(indices().coeff(k) == j) indices().coeffRef(k) = StorageIndexType(i);
+        if(indices().coeff(k) == i) indices().coeffRef(k) = StorageIndex(j);
+        else if(indices().coeff(k) == j) indices().coeffRef(k) = StorageIndex(i);
       }
       return derived();
     }
@@ -185,7 +182,7 @@ class PermutationBase : public EigenBase<Derived>
       *
       * This is a fast operation, it only consists in swapping two indices.
       *
-      * \sa applyTranspositionOnTheLeft(int,int)
+      * \sa applyTranspositionOnTheLeft(Index,Index)
       */
     Derived& applyTranspositionOnTheRight(Index i, Index j)
     {
@@ -198,14 +195,14 @@ class PermutationBase : public EigenBase<Derived>
       *
       * \note \note_try_to_help_rvo
       */
-    inline Transpose<PermutationBase> inverse() const
-    { return derived(); }
+    inline TransposeReturnType inverse() const
+    { return TransposeReturnType(derived()); }
     /** \returns the tranpose permutation matrix.
       *
       * \note \note_try_to_help_rvo
       */
-    inline Transpose<PermutationBase> transpose() const
-    { return derived(); }
+    inline TransposeReturnType transpose() const
+    { return TransposeReturnType(derived()); }
 
     /**** multiplication helpers to hopefully get RVO ****/
 
@@ -215,13 +212,13 @@ class PermutationBase : public EigenBase<Derived>
     template<typename OtherDerived>
     void assignTranspose(const PermutationBase<OtherDerived>& other)
     {
-      for (int i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i;
+      for (Index i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i;
     }
     template<typename Lhs,typename Rhs>
     void assignProduct(const Lhs& lhs, const Rhs& rhs)
     {
       eigen_assert(lhs.cols() == rhs.rows());
-      for (int i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i));
+      for (Index i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i));
     }
 #endif
 
@@ -250,6 +247,35 @@ class PermutationBase : public EigenBase<Derived>
     template<typename Other> friend
     inline PlainPermutationType operator*(const Transpose<PermutationBase<Other> >& other, const PermutationBase& perm)
     { return PlainPermutationType(internal::PermPermProduct, other.eval(), perm); }
+    
+    /** \returns the determinant of the permutation matrix, which is either 1 or -1 depending on the parity of the permutation.
+      *
+      * This function is O(\c n) procedure allocating a buffer of \c n booleans.
+      */
+    Index determinant() const
+    {
+      Index res = 1;
+      Index n = size();
+      Matrix<bool,RowsAtCompileTime,1,0,MaxRowsAtCompileTime> mask(n);
+      mask.fill(false);
+      Index r = 0;
+      while(r < n)
+      {
+        // search for the next seed
+        while(r<n && mask[r]) r++;
+        if(r>=n)
+          break;
+        // we got one, let's follow it until we are back to the seed
+        Index k0 = r++;
+        mask.coeffRef(k0) = true;
+        for(Index k=indices().coeff(k0); k!=k0; k=indices().coeff(k))
+        {
+          mask.coeffRef(k) = true;
+          res = -res;
+        }
+      }
+      return res;
+    }
 
   protected:
 
@@ -262,7 +288,7 @@ class PermutationBase : public EigenBase<Derived>
   *
   * \param SizeAtCompileTime the number of rows/cols, or Dynamic
   * \param MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
-  * \param StorageIndexType the integer type of the indices
+  * \param StorageIndex the integer type of the indices
   *
   * This class represents a permutation matrix, internally stored as a vector of integers.
   *
@@ -270,27 +296,28 @@ class PermutationBase : public EigenBase<Derived>
   */
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType>
-struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndexType> >
- : traits<Matrix<_StorageIndexType,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
-  typedef Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
-  typedef typename IndicesType::Index Index;
-  typedef _StorageIndexType StorageIndexType;
+  typedef PermutationStorage StorageKind;
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef _StorageIndex StorageIndex;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType>
-class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
 {
     typedef PermutationBase<PermutationMatrix> Base;
     typedef internal::traits<PermutationMatrix> Traits;
   public:
 
+    typedef const PermutationMatrix& Nested;
+
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename Traits::StorageIndexType StorageIndexType;
-    typedef typename Traits::Index Index;
+    typedef typename Traits::StorageIndex StorageIndex;
     #endif
 
     inline PermutationMatrix()
@@ -298,8 +325,10 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
     /** Constructs an uninitialized permutation matrix of given size.
       */
-    inline PermutationMatrix(Index size) : m_indices(size)
-    {}
+    explicit inline PermutationMatrix(Index size) : m_indices(size)
+    {
+      eigen_internal_assert(size <= NumTraits<StorageIndex>::highest());
+    }
 
     /** Copy constructor. */
     template<typename OtherDerived>
@@ -320,7 +349,7 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
       * array's size.
       */
     template<typename Other>
-    explicit inline PermutationMatrix(const MatrixBase<Other>& a_indices) : m_indices(a_indices)
+    explicit inline PermutationMatrix(const MatrixBase<Other>& indices) : m_indices(indices)
     {}
 
     /** Convert the Transpositions \a tr to a permutation matrix */
@@ -368,9 +397,12 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Other>
     PermutationMatrix(const Transpose<PermutationBase<Other> >& other)
-      : m_indices(other.nestedPermutation().size())
+      : m_indices(other.nestedExpression().size())
     {
-      for (int i=0; i<m_indices.size();++i) m_indices.coeffRef(other.nestedPermutation().indices().coeff(i)) = i;
+      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;
     }
     template<typename Lhs,typename Rhs>
     PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs)
@@ -387,19 +419,19 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType, int _PacketAccess>
-struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndexType>,_PacketAccess> >
- : traits<Matrix<_StorageIndexType,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
-  typedef Map<const Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
-  typedef typename IndicesType::Index Index;
-  typedef _StorageIndexType StorageIndexType;
+  typedef PermutationStorage StorageKind;
+  typedef Map<const Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType, int _PacketAccess>
-class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndexType>,_PacketAccess>
-  : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndexType>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess>
+  : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
 {
     typedef PermutationBase<Map> Base;
     typedef internal::traits<Map> Traits;
@@ -407,15 +439,14 @@ class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageInd
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar StorageIndexType;
-    typedef typename IndicesType::Index  Index;
+    typedef typename IndicesType::Scalar StorageIndex;
     #endif
 
-    inline Map(const StorageIndexType* indicesPtr)
+    inline Map(const StorageIndex* indicesPtr)
       : m_indices(indicesPtr)
     {}
 
-    inline Map(const StorageIndexType* indicesPtr, Index size)
+    inline Map(const StorageIndex* indicesPtr, Index size)
       : m_indices(indicesPtr,size)
     {}
 
@@ -462,8 +493,6 @@ class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageInd
   * \sa class PermutationBase, class PermutationMatrix
   */
 
-struct PermutationStorage {};
-
 template<typename _IndicesType> class TranspositionsWrapper;
 namespace internal {
 template<typename _IndicesType>
@@ -471,16 +500,14 @@ struct traits<PermutationWrapper<_IndicesType> >
 {
   typedef PermutationStorage StorageKind;
   typedef typename _IndicesType::Scalar Scalar;
-  typedef typename _IndicesType::Scalar StorageIndexType;
-  typedef typename _IndicesType::Index Index;
+  typedef typename _IndicesType::Scalar StorageIndex;
   typedef _IndicesType IndicesType;
   enum {
     RowsAtCompileTime = _IndicesType::SizeAtCompileTime,
     ColsAtCompileTime = _IndicesType::SizeAtCompileTime,
-    MaxRowsAtCompileTime = IndicesType::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = IndicesType::MaxColsAtCompileTime,
-    Flags = 0,
-    CoeffReadCost = _IndicesType::CoeffReadCost
+    MaxRowsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    Flags = 0
   };
 };
 }
@@ -496,8 +523,8 @@ class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesTyp
     typedef typename Traits::IndicesType IndicesType;
     #endif
 
-    inline PermutationWrapper(const IndicesType& a_indices)
-      : m_indices(a_indices)
+    inline PermutationWrapper(const IndicesType& indices)
+      : m_indices(indices)
     {}
 
     /** const version of indices(). */
@@ -509,105 +536,33 @@ class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesTyp
     typename IndicesType::Nested m_indices;
 };
 
+
 /** \returns the matrix with the permutation applied to the columns.
   */
-template<typename Derived, typename PermutationDerived>
-inline const internal::permut_matrix_product_retval<PermutationDerived, Derived, OnTheRight>
-operator*(const MatrixBase<Derived>& matrix,
-          const PermutationBase<PermutationDerived> &permutation)
+template<typename MatrixDerived, typename PermutationDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const PermutationBase<PermutationDerived>& permutation)
 {
-  return internal::permut_matrix_product_retval
-           <PermutationDerived, Derived, OnTheRight>
-           (permutation.derived(), matrix.derived());
+  return Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+            (matrix.derived(), permutation.derived());
 }
 
 /** \returns the matrix with the permutation applied to the rows.
   */
-template<typename Derived, typename PermutationDerived>
-inline const internal::permut_matrix_product_retval
-               <PermutationDerived, Derived, OnTheLeft>
+template<typename PermutationDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
 operator*(const PermutationBase<PermutationDerived> &permutation,
-          const MatrixBase<Derived>& matrix)
+          const MatrixBase<MatrixDerived>& matrix)
 {
-  return internal::permut_matrix_product_retval
-           <PermutationDerived, Derived, OnTheLeft>
-           (permutation.derived(), matrix.derived());
+  return Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
+            (permutation.derived(), matrix.derived());
 }
 
 namespace internal {
 
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed>
-struct traits<permut_matrix_product_retval<PermutationType, MatrixType, Side, Transposed> >
-{
-  typedef typename MatrixType::PlainObject ReturnType;
-};
-
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed>
-struct permut_matrix_product_retval
- : public ReturnByValue<permut_matrix_product_retval<PermutationType, MatrixType, Side, Transposed> >
-{
-    typedef typename remove_all<typename MatrixType::Nested>::type MatrixTypeNestedCleaned;
-    typedef typename MatrixType::Index Index;
-
-    permut_matrix_product_retval(const PermutationType& perm, const MatrixType& matrix)
-      : m_permutation(perm), m_matrix(matrix)
-    {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-
-    template<typename Dest> inline void evalTo(Dest& dst) const
-    {
-      const Index n = Side==OnTheLeft ? rows() : cols();
-      // FIXME we need an is_same for expression that is not sensitive to constness. For instance
-      // is_same_xpr<Block<const Matrix>, Block<Matrix> >::value should be true.
-      if(is_same<MatrixTypeNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_matrix))
-      {
-        // apply the permutation inplace
-        Matrix<bool,PermutationType::RowsAtCompileTime,1,0,PermutationType::MaxRowsAtCompileTime> mask(m_permutation.size());
-        mask.fill(false);
-        Index r = 0;
-        while(r < m_permutation.size())
-        {
-          // search for the next seed
-          while(r<m_permutation.size() && mask[r]) r++;
-          if(r>=m_permutation.size())
-            break;
-          // we got one, let's follow it until we are back to the seed
-          Index k0 = r++;
-          Index kPrev = k0;
-          mask.coeffRef(k0) = true;
-          for(Index k=m_permutation.indices().coeff(k0); k!=k0; k=m_permutation.indices().coeff(k))
-          {
-                  Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>(dst, k)
-            .swap(Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
-                       (dst,((Side==OnTheLeft) ^ Transposed) ? k0 : kPrev));
-
-            mask.coeffRef(k) = true;
-            kPrev = k;
-          }
-        }
-      }
-      else
-      {
-        for(int i = 0; i < n; ++i)
-        {
-          Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
-               (dst, ((Side==OnTheLeft) ^ Transposed) ? m_permutation.indices().coeff(i) : i)
-
-          =
-
-          Block<const MatrixTypeNestedCleaned,Side==OnTheLeft ? 1 : MatrixType::RowsAtCompileTime,Side==OnTheRight ? 1 : MatrixType::ColsAtCompileTime>
-               (m_matrix, ((Side==OnTheRight) ^ Transposed) ? m_permutation.indices().coeff(i) : i);
-        }
-      }
-    }
-
-  protected:
-    const PermutationType& m_permutation;
-    typename MatrixType::Nested m_matrix;
-};
-
 /* Template partial specialization for transposed/inverse permutations */
 
 template<typename Derived>
@@ -617,6 +572,8 @@ struct traits<Transpose<PermutationBase<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> > >
@@ -631,26 +588,26 @@ class Transpose<PermutationBase<Derived> >
     typedef typename Derived::DenseMatrixType DenseMatrixType;
     enum {
       Flags = Traits::Flags,
-      CoeffReadCost = Traits::CoeffReadCost,
       RowsAtCompileTime = Traits::RowsAtCompileTime,
       ColsAtCompileTime = Traits::ColsAtCompileTime,
       MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
       MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
     };
     typedef typename Traits::Scalar Scalar;
+    typedef typename Traits::StorageIndex StorageIndex;
     #endif
 
     Transpose(const PermutationType& p) : m_permutation(p) {}
 
-    inline int rows() const { return m_permutation.rows(); }
-    inline int cols() const { return m_permutation.cols(); }
+    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 (int i=0; i<rows();++i)
+      for (Index i=0; i<rows();++i)
         other.coeffRef(i, m_permutation.indices().coeff(i)) = typename DenseDerived::Scalar(1);
     }
     #endif
@@ -663,22 +620,22 @@ class Transpose<PermutationBase<Derived> >
     /** \returns the matrix with the inverse permutation applied to the columns.
       */
     template<typename OtherDerived> friend
-    inline const internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheRight, true>
+    const Product<OtherDerived, Transpose, AliasFreeProduct>
     operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trPerm)
     {
-      return internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheRight, true>(trPerm.m_permutation, matrix.derived());
+      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trPerm.derived());
     }
 
     /** \returns the matrix with the inverse permutation applied to the rows.
       */
     template<typename OtherDerived>
-    inline const internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheLeft, true>
+    const Product<Transpose, OtherDerived, AliasFreeProduct>
     operator*(const MatrixBase<OtherDerived>& matrix) const
     {
-      return internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheLeft, true>(m_permutation, matrix.derived());
+      return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived());
     }
 
-    const PermutationType& nestedPermutation() const { return m_permutation; }
+    const PermutationType& nestedExpression() const { return m_permutation; }
 
   protected:
     const PermutationType& m_permutation;
@@ -690,6 +647,12 @@ const PermutationWrapper<const Derived> MatrixBase<Derived>::asPermutation() con
   return derived();
 }
 
+namespace internal {
+
+template<> struct AssignmentKind<DenseShape,PermutationShape> { typedef EigenBase2EigenBase Kind; };
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_PERMUTATIONMATRIX_H

Eigen/src/Core/PlainObjectBase.h

@@ -69,8 +69,9 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct m
 #ifdef EIGEN_PARSED_BY_DOXYGEN
 namespace internal {
 
-// this is a warkaround to doxygen not being able to understand the inheritence logic
+// this is a workaround to doxygen not being able to understand the inheritance logic
 // when it is hidden by the dense_xpr_base helper struct.
+/** This class is just a workaround for Doxygen and it does not not actually exist. */
 template<typename Derived> struct dense_xpr_base_dispatcher_for_doxygen;// : public MatrixBase<Derived> {};
 /** This class is just a workaround for Doxygen and it does not not actually exist. */
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
@@ -95,8 +96,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     typedef typename internal::dense_xpr_base<Derived>::type Base;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
+    
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
     typedef Derived DenseType;
@@ -115,20 +116,23 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     typedef Eigen::Map<Derived, Unaligned>  MapType;
     friend  class Eigen::Map<const Derived, Unaligned>;
     typedef const Eigen::Map<const Derived, Unaligned> ConstMapType;
-    friend  class Eigen::Map<Derived, Aligned>;
-    typedef Eigen::Map<Derived, Aligned> AlignedMapType;
-    friend  class Eigen::Map<const Derived, Aligned>;
-    typedef const Eigen::Map<const Derived, Aligned> ConstAlignedMapType;
+#if EIGEN_MAX_ALIGN_BYTES>0
+    // for EIGEN_MAX_ALIGN_BYTES==0, AlignedMax==Unaligned, and many compilers generate warnings for friend-ing a class twice.
+    friend  class Eigen::Map<Derived, AlignedMax>;
+    friend  class Eigen::Map<const Derived, AlignedMax>;
+#endif
+    typedef Eigen::Map<Derived, AlignedMax> AlignedMapType;
+    typedef const Eigen::Map<const Derived, AlignedMax> ConstAlignedMapType;
     template<typename StrideType> struct StridedMapType { typedef Eigen::Map<Derived, Unaligned, StrideType> type; };
     template<typename StrideType> struct StridedConstMapType { typedef Eigen::Map<const Derived, Unaligned, StrideType> type; };
-    template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, Aligned, StrideType> type; };
-    template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, Aligned, StrideType> type; };
+    template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, AlignedMax, StrideType> type; };
+    template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, AlignedMax, StrideType> type; };
 
   protected:
     DenseStorage<Scalar, Base::MaxSizeAtCompileTime, Base::RowsAtCompileTime, Base::ColsAtCompileTime, Options> m_storage;
 
   public:
-    enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits<Derived>::Flags & AlignedBit) != 0 };
+    enum { NeedsToAlign = (SizeAtCompileTime != Dynamic) && (internal::traits<Derived>::Alignment>0) };
     EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
 
     EIGEN_DEVICE_FUNC
@@ -221,11 +225,11 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 
     /** \returns a const pointer to the data array of this matrix */
-    EIGEN_STRONG_INLINE const Scalar *data() const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar *data() const
     { return m_storage.data(); }
 
     /** \returns a pointer to the data array of this matrix */
-    EIGEN_STRONG_INLINE Scalar *data()
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar *data()
     { return m_storage.data(); }
 
     /** Resizes \c *this to a \a rows x \a cols matrix.
@@ -245,22 +249,22 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * \sa resize(Index) for vectors, resize(NoChange_t, Index), resize(Index, NoChange_t)
       */
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void resize(Index nbRows, Index nbCols)
+    EIGEN_STRONG_INLINE void resize(Index rows, Index cols)
     {
-      eigen_assert(   EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,nbRows==RowsAtCompileTime)
-                   && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,nbCols==ColsAtCompileTime)
-                   && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,nbRows<=MaxRowsAtCompileTime)
-                   && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,nbCols<=MaxColsAtCompileTime)
-                   && nbRows>=0 && nbCols>=0 && "Invalid sizes when resizing a matrix or array.");
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(nbRows, nbCols);
+      eigen_assert(   EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,rows==RowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,cols==ColsAtCompileTime)
+                   && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,rows<=MaxRowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,cols<=MaxColsAtCompileTime)
+                   && rows>=0 && cols>=0 && "Invalid sizes when resizing a matrix or array.");
+      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols);
       #ifdef EIGEN_INITIALIZE_COEFFS
-        Index size = nbRows*nbCols;
+        Index size = rows*cols;
         bool size_changed = size != this->size();
-        m_storage.resize(size, nbRows, nbCols);
+        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(nbRows, nbCols);
-        m_storage.resize(nbRows*nbCols, nbRows, nbCols);
+        internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols);
+        m_storage.resize(rows*cols, rows, cols);
       #endif
     }
 
@@ -301,9 +305,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * \sa resize(Index,Index)
       */
     EIGEN_DEVICE_FUNC
-    inline void resize(NoChange_t, Index nbCols)
+    inline void resize(NoChange_t, Index cols)
     {
-      resize(rows(), nbCols);
+      resize(rows(), cols);
     }
 
     /** Resizes the matrix, changing only the number of rows. For the parameter of type NoChange_t, just pass the special value \c NoChange
@@ -315,9 +319,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * \sa resize(Index,Index)
       */
     EIGEN_DEVICE_FUNC
-    inline void resize(Index nbRows, NoChange_t)
+    inline void resize(Index rows, NoChange_t)
     {
-      resize(nbRows, cols());
+      resize(rows, cols());
     }
 
     /** Resizes \c *this to have the same dimensions as \a other.
@@ -357,9 +361,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * appended to the matrix they will be uninitialized.
       */
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, Index nbCols)
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, Index cols)
     {
-      internal::conservative_resize_like_impl<Derived>::run(*this, nbRows, nbCols);
+      internal::conservative_resize_like_impl<Derived>::run(*this, rows, cols);
     }
 
     /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
@@ -370,10 +374,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * In case the matrix is growing, new rows will be uninitialized.
       */
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, NoChange_t)
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, NoChange_t)
     {
       // Note: see the comment in conservativeResize(Index,Index)
-      conservativeResize(nbRows, cols());
+      conservativeResize(rows, cols());
     }
 
     /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
@@ -384,10 +388,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * In case the matrix is growing, new columns will be uninitialized.
       */
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index nbCols)
+    EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index cols)
     {
       // Note: see the comment in conservativeResize(Index,Index)
-      conservativeResize(rows(), nbCols);
+      conservativeResize(rows(), cols);
     }
 
     /** Resizes the vector to \a size while retaining old values.
@@ -457,7 +461,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     // FIXME is it still needed ?
     /** \internal */
     EIGEN_DEVICE_FUNC
-    PlainObjectBase(internal::constructor_without_unaligned_array_assert)
+    explicit PlainObjectBase(internal::constructor_without_unaligned_array_assert)
       : m_storage(internal::constructor_without_unaligned_array_assert())
     {
 //       _check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
@@ -480,9 +484,13 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 #endif
 
+    /** Copy constructor */
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE PlainObjectBase(Index a_size, Index nbRows, Index nbCols)
-      : m_storage(a_size, nbRows, nbCols)
+    EIGEN_STRONG_INLINE PlainObjectBase(const PlainObjectBase& other)
+      : Base(), m_storage(other.m_storage) { }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(Index size, Index rows, Index cols)
+      : m_storage(size, rows, cols)
     {
 //       _check_template_params();
 //       EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
@@ -499,15 +507,36 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       return this->derived();
     }
 
-    /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const DenseBase<OtherDerived> &other)
+      : m_storage()
+    {
+      _check_template_params();
+      resizeLike(other);
+      _set_noalias(other);
+    }
+
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
-      : m_storage(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
+      : m_storage()
     {
       _check_template_params();
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.derived().rows(), other.derived().cols());
-      Base::operator=(other.derived());
+      resizeLike(other);
+      *this = other.derived();
+    }
+    /** \brief Copy constructor with in-place evaluation */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const ReturnByValue<OtherDerived>& other)
+    {
+      _check_template_params();
+      // FIXME this does not automatically transpose vectors if necessary
+      resize(other.rows(), other.cols());
+      other.evalTo(this->derived());
     }
 
     /** \name Map
@@ -639,22 +668,16 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \internal
       */
+    // aliasing is dealt once in internall::call_assignment
+    // so at this stage we have to assume aliasing... and resising has to be done later.
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE Derived& _set(const DenseBase<OtherDerived>& other)
     {
-      _set_selector(other.derived(), typename internal::conditional<static_cast<bool>(int(OtherDerived::Flags) & EvalBeforeAssigningBit), internal::true_type, internal::false_type>::type());
+      internal::call_assignment(this->derived(), other.derived());
       return this->derived();
     }
 
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC 
-    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::true_type&) { _set_noalias(other.eval()); }
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC 
-    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::false_type&) { _set_noalias(other); }
-
     /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which
       * is the case when creating a new matrix) so one can enforce lazy evaluation.
       *
@@ -669,17 +692,18 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       //_resize_to_match(other);
       // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because
       // it wouldn't allow to copy a row-vector into a column-vector.
-      return internal::assign_selector<Derived,OtherDerived,false>::run(this->derived(), other.derived());
+      internal::call_assignment_no_alias(this->derived(), other.derived(), internal::assign_op<Scalar>());
+      return this->derived();
     }
 
     template<typename T0, typename T1>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void _init2(Index nbRows, Index nbCols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
+    EIGEN_STRONG_INLINE void _init2(Index rows, Index cols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
     {
       EIGEN_STATIC_ASSERT(bool(NumTraits<T0>::IsInteger) &&
                           bool(NumTraits<T1>::IsInteger),
                           FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
-      resize(nbRows,nbCols);
+      resize(rows,cols);
     }
     
     template<typename T0, typename T1>
@@ -704,9 +728,12 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       m_storage.data()[1] = Scalar(val1);
     }
 
+    // The argument is convertible to the Index type and we either have a non 1x1 Matrix, or a dynamic-sized Array,
+    // then the argument is meant to be the size of the object.
     template<typename T>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if<Base::SizeAtCompileTime!=1 || !internal::is_convertible<T, Scalar>::value,T>::type* = 0)
+    EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if<    (Base::SizeAtCompileTime!=1 || !internal::is_convertible<T, Scalar>::value)
+                                                                              && ((!internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value || Base::SizeAtCompileTime==Dynamic)),T>::type* = 0)
     {
       // NOTE MSVC 2008 complains if we directly put bool(NumTraits<T>::IsInteger) as the EIGEN_STATIC_ASSERT argument.
       const bool is_integer = NumTraits<T>::IsInteger;
@@ -714,6 +741,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
                           FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
       resize(size);
     }
+    
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitely converted)
     template<typename T>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if<Base::SizeAtCompileTime==1 && internal::is_convertible<T, Scalar>::value,T>::type* = 0)
@@ -722,6 +751,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       m_storage.data()[0] = val0;
     }
     
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type match the index type)
     template<typename T>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const Index& val0,
@@ -734,18 +764,21 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       m_storage.data()[0] = Scalar(val0);
     }
 
+    // Initialize a fixed size matrix from a pointer to raw data
     template<typename T>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const Scalar* data){
       this->_set_noalias(ConstMapType(data));
     }
 
+    // Initialize an arbitrary matrix from a dense expression
     template<typename T, typename OtherDerived>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const DenseBase<OtherDerived>& other){
       this->_set_noalias(other);
     }
 
+    // Initialize an arbitrary matrix from a generic Eigen expression
     template<typename T, typename OtherDerived>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const EigenBase<OtherDerived>& other){
@@ -766,23 +799,58 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     {
       this->derived() = r;
     }
-
+    
+    // For fixed -size arrays:
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar& val0,
+                                    typename internal::enable_if<    Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Index& val0,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<Index,T>::value)
+                                                                  && Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T*>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
     template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
     friend struct internal::matrix_swap_impl;
 
-    /** \internal generic implementation of swap for dense storage since for dynamic-sized matrices of same type it is enough to swap the
-      * data pointers.
+  public:
+    
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal
+      * \brief Override DenseBase::swap() since for dynamic-sized matrices
+      * of same type it is enough to swap the data pointers.
       */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    void _swap(DenseBase<OtherDerived> const & other)
+    void swap(DenseBase<OtherDerived> & other)
     {
       enum { SwapPointers = internal::is_same<Derived, OtherDerived>::value && Base::SizeAtCompileTime==Dynamic };
-      internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.const_cast_derived());
+      internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.derived());
     }
-
-  public:
-#ifndef EIGEN_PARSED_BY_DOXYGEN
+    
+    /** \internal
+      * \brief const version forwarded to DenseBase::swap
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(DenseBase<OtherDerived> const & other)
+    { Base::swap(other.derived()); }
+    
     EIGEN_DEVICE_FUNC 
     static EIGEN_STRONG_INLINE void _check_template_params()
     {
@@ -797,10 +865,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
                         && (Options & (DontAlign|RowMajor)) == Options),
         INVALID_MATRIX_TEMPLATE_PARAMETERS)
     }
-#endif
 
-private:
-    enum { ThisConstantIsPrivateInPlainObjectBase };
+    enum { IsPlainObjectBase = 1 };
+#endif
 };
 
 namespace internal {
@@ -808,7 +875,6 @@ namespace internal {
 template <typename Derived, typename OtherDerived, bool IsVector>
 struct conservative_resize_like_impl
 {
-  typedef typename Derived::Index Index;
   static void run(DenseBase<Derived>& _this, Index rows, Index cols)
   {
     if (_this.rows() == rows && _this.cols() == cols) return;
@@ -874,7 +940,6 @@ struct conservative_resize_like_impl<Derived,OtherDerived,true>
 {
   using conservative_resize_like_impl<Derived,OtherDerived,false>::run;
   
-  typedef typename Derived::Index Index;
   static void run(DenseBase<Derived>& _this, Index size)
   {
     const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : size;

Eigen/src/Core/Product.h

@@ -12,8 +12,7 @@
 
 namespace Eigen {
 
-template<typename Lhs, typename Rhs> class Product;
-template<typename Lhs, typename Rhs, typename StorageKind> class ProductImpl;
+template<typename Lhs, typename Rhs, int Option, typename StorageKind> class ProductImpl;
 
 /** \class Product
   * \ingroup Core_Module
@@ -24,53 +23,121 @@ template<typename Lhs, typename Rhs, typename StorageKind> class ProductImpl;
   * \param Rhs the type of the right-hand side expression
   *
   * This class represents an expression of the product of two arbitrary matrices.
+  * 
+  * The other template parameters are:
+  * \tparam Option     can be DefaultProduct, AliasFreeProduct, or LazyProduct
   *
   */
 
-// Use ProductReturnType to get correct traits, in particular vectorization flags
+
 namespace internal {
-template<typename Lhs, typename Rhs>
-struct traits<Product<Lhs, Rhs> >
-  : traits<typename ProductReturnType<Lhs, Rhs>::Type>
-{ 
-  // We want A+B*C to be of type Product<Matrix, Sum> and not Product<Matrix, Matrix>
-  // TODO: This flag should eventually go in a separate evaluator traits class
+
+// Determine the scalar of Product<Lhs, Rhs>. This is normally the same as Lhs::Scalar times
+// Rhs::Scalar, but product with permutation matrices inherit the scalar of the other factor.
+template<typename Lhs, typename Rhs, typename LhsShape = typename evaluator_traits<Lhs>::Shape, 
+         typename RhsShape = typename evaluator_traits<Rhs>::Shape >
+struct product_result_scalar
+{
+  typedef typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename RhsShape>
+struct product_result_scalar<Lhs, Rhs, PermutationShape, RhsShape>
+{
+  typedef typename Rhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename LhsShape>
+  struct product_result_scalar<Lhs, Rhs, LhsShape, PermutationShape>
+{
+  typedef typename Lhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename RhsShape>
+struct product_result_scalar<Lhs, Rhs, TranspositionsShape, RhsShape>
+{
+  typedef typename Rhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename LhsShape>
+  struct product_result_scalar<Lhs, Rhs, LhsShape, TranspositionsShape>
+{
+  typedef typename Lhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, int Option>
+struct traits<Product<Lhs, Rhs, Option> >
+{
+  typedef typename remove_all<Lhs>::type LhsCleaned;
+  typedef typename remove_all<Rhs>::type RhsCleaned;
+  typedef traits<LhsCleaned> LhsTraits;
+  typedef traits<RhsCleaned> RhsTraits;
+  
+  typedef MatrixXpr XprKind;
+  
+  typedef typename product_result_scalar<LhsCleaned,RhsCleaned>::Scalar Scalar;
+  typedef typename product_promote_storage_type<typename LhsTraits::StorageKind,
+                                                typename RhsTraits::StorageKind,
+                                                internal::product_type<Lhs,Rhs>::ret>::ret StorageKind;
+  typedef typename promote_index_type<typename LhsTraits::StorageIndex,
+                                      typename RhsTraits::StorageIndex>::type StorageIndex;
+  
   enum {
-    Flags = traits<typename ProductReturnType<Lhs, Rhs>::Type>::Flags & ~(EvalBeforeNestingBit | DirectAccessBit)
+    RowsAtCompileTime    = LhsTraits::RowsAtCompileTime,
+    ColsAtCompileTime    = RhsTraits::ColsAtCompileTime,
+    MaxRowsAtCompileTime = LhsTraits::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = RhsTraits::MaxColsAtCompileTime,
+    
+    // FIXME: only needed by GeneralMatrixMatrixTriangular
+    InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsTraits::ColsAtCompileTime, RhsTraits::RowsAtCompileTime),
+    
+    // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator.
+    Flags = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? RowMajorBit
+          : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0
+          : (   ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit))
+             || ((RhsTraits::Flags&NoPreferredStorageOrderBit) && (LhsTraits::Flags&RowMajorBit)) ) ? RowMajorBit
+          : NoPreferredStorageOrderBit
   };
 };
+
 } // end namespace internal
 
 
-template<typename Lhs, typename Rhs>
-class Product : public ProductImpl<Lhs,Rhs,typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                                                            typename internal::traits<Rhs>::StorageKind>::ret>
+template<typename _Lhs, typename _Rhs, int Option>
+class Product : public ProductImpl<_Lhs,_Rhs,Option,
+                                   typename internal::product_promote_storage_type<typename internal::traits<_Lhs>::StorageKind,
+                                                                                   typename internal::traits<_Rhs>::StorageKind,
+                                                                                   internal::product_type<_Lhs,_Rhs>::ret>::ret>
 {
   public:
     
+    typedef _Lhs Lhs;
+    typedef _Rhs Rhs;
+    
     typedef typename ProductImpl<
-        Lhs, Rhs,
-        typename internal::promote_storage_type<typename Lhs::StorageKind,
-                                                typename Rhs::StorageKind>::ret>::Base Base;
+        Lhs, Rhs, Option,
+        typename internal::product_promote_storage_type<typename internal::traits<Lhs>::StorageKind,
+                                                        typename internal::traits<Rhs>::StorageKind,
+                                                        internal::product_type<Lhs,Rhs>::ret>::ret>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
 
-    typedef typename Lhs::Nested LhsNested;
-    typedef typename Rhs::Nested RhsNested;
+    typedef typename internal::ref_selector<Lhs>::type LhsNested;
+    typedef typename internal::ref_selector<Rhs>::type RhsNested;
     typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
     typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
 
-    Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs)
+    EIGEN_DEVICE_FUNC Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs)
     {
       eigen_assert(lhs.cols() == rhs.rows()
         && "invalid matrix product"
         && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
     }
 
-    inline Index rows() const { return m_lhs.rows(); }
-    inline Index cols() const { return m_rhs.cols(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); }
 
-    const LhsNestedCleaned& lhs() const { return m_lhs; }
-    const RhsNestedCleaned& rhs() const { return m_rhs; }
+    EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; }
+    EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; }
 
   protected:
 
@@ -78,14 +145,76 @@ class Product : public ProductImpl<Lhs,Rhs,typename internal::promote_storage_ty
     RhsNested m_rhs;
 };
 
-template<typename Lhs, typename Rhs>
-class ProductImpl<Lhs,Rhs,Dense> : public internal::dense_xpr_base<Product<Lhs,Rhs> >::type
-{
-    typedef Product<Lhs, Rhs> Derived;
-  public:
+namespace internal {
+  
+template<typename Lhs, typename Rhs, int Option, int ProductTag = internal::product_type<Lhs,Rhs>::ret>
+class dense_product_base
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{};
 
-    typedef typename internal::dense_xpr_base<Product<Lhs, Rhs> >::type Base;
+/** Convertion to scalar for inner-products */
+template<typename Lhs, typename Rhs, int Option>
+class dense_product_base<Lhs, Rhs, Option, InnerProduct>
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{
+  typedef Product<Lhs,Rhs,Option> ProductXpr;
+  typedef typename internal::dense_xpr_base<ProductXpr>::type Base;
+public:
+  using Base::derived;
+  typedef typename Base::Scalar Scalar;
+  
+  operator const Scalar() const
+  {
+    return internal::evaluator<ProductXpr>(derived()).coeff(0,0);
+  }
+};
+
+} // namespace internal
+
+// Generic API dispatcher
+template<typename Lhs, typename Rhs, int Option, typename StorageKind>
+class ProductImpl : public internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type Base;
+};
+
+template<typename Lhs, typename Rhs, int Option>
+class ProductImpl<Lhs,Rhs,Option,Dense>
+  : public internal::dense_product_base<Lhs,Rhs,Option>
+{
+    typedef Product<Lhs, Rhs, Option> Derived;
+    
+  public:
+    
+    typedef typename internal::dense_product_base<Lhs, Rhs, Option> Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+  protected:
+    enum {
+      IsOneByOne = (RowsAtCompileTime == 1 || RowsAtCompileTime == Dynamic) && 
+                   (ColsAtCompileTime == 1 || ColsAtCompileTime == Dynamic),
+      EnableCoeff = IsOneByOne || Option==LazyProduct
+    };
+    
+  public:
+  
+    EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
+    }
+
+    EIGEN_DEVICE_FUNC Scalar coeff(Index i) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(i);
+    }
+    
+  
 };
 
 /***************************************************************************
@@ -102,6 +231,15 @@ 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/ProductBase.h

@@ -1,274 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2009-2010 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_PRODUCTBASE_H
-#define EIGEN_PRODUCTBASE_H
-
-namespace Eigen { 
-
-/** \class ProductBase
-  * \ingroup Core_Module
-  *
-  */
-
-namespace internal {
-template<typename Derived, typename _Lhs, typename _Rhs>
-struct traits<ProductBase<Derived,_Lhs,_Rhs> >
-{
-  typedef MatrixXpr XprKind;
-  typedef typename remove_all<_Lhs>::type Lhs;
-  typedef typename remove_all<_Rhs>::type Rhs;
-  typedef typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType Scalar;
-  typedef typename promote_storage_type<typename traits<Lhs>::StorageKind,
-                                           typename traits<Rhs>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<Lhs>::Index,
-                                         typename traits<Rhs>::Index>::type Index;
-  enum {
-    RowsAtCompileTime = traits<Lhs>::RowsAtCompileTime,
-    ColsAtCompileTime = traits<Rhs>::ColsAtCompileTime,
-    MaxRowsAtCompileTime = traits<Lhs>::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = traits<Rhs>::MaxColsAtCompileTime,
-    Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0)
-          | EvalBeforeNestingBit | EvalBeforeAssigningBit | NestByRefBit,
-                  // Note that EvalBeforeNestingBit and NestByRefBit
-                  // are not used in practice because nested is overloaded for products
-    CoeffReadCost = 0 // FIXME why is it needed ?
-  };
-};
-}
-
-#define EIGEN_PRODUCT_PUBLIC_INTERFACE(Derived) \
-  typedef ProductBase<Derived, Lhs, Rhs > Base; \
-  EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
-  typedef typename Base::LhsNested LhsNested; \
-  typedef typename Base::_LhsNested _LhsNested; \
-  typedef typename Base::LhsBlasTraits LhsBlasTraits; \
-  typedef typename Base::ActualLhsType ActualLhsType; \
-  typedef typename Base::_ActualLhsType _ActualLhsType; \
-  typedef typename Base::RhsNested RhsNested; \
-  typedef typename Base::_RhsNested _RhsNested; \
-  typedef typename Base::RhsBlasTraits RhsBlasTraits; \
-  typedef typename Base::ActualRhsType ActualRhsType; \
-  typedef typename Base::_ActualRhsType _ActualRhsType; \
-  using Base::m_lhs; \
-  using Base::m_rhs;
-
-template<typename Derived, typename Lhs, typename Rhs>
-class ProductBase : public MatrixBase<Derived>
-{
-  public:
-    typedef MatrixBase<Derived> Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(ProductBase)
-    
-    typedef typename Lhs::Nested LhsNested;
-    typedef typename internal::remove_all<LhsNested>::type _LhsNested;
-    typedef internal::blas_traits<_LhsNested> LhsBlasTraits;
-    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
-    typedef typename internal::remove_all<ActualLhsType>::type _ActualLhsType;
-    typedef typename internal::traits<Lhs>::Scalar LhsScalar;
-
-    typedef typename Rhs::Nested RhsNested;
-    typedef typename internal::remove_all<RhsNested>::type _RhsNested;
-    typedef internal::blas_traits<_RhsNested> RhsBlasTraits;
-    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
-    typedef typename internal::remove_all<ActualRhsType>::type _ActualRhsType;
-    typedef typename internal::traits<Rhs>::Scalar RhsScalar;
-
-    // Diagonal of a product: no need to evaluate the arguments because they are going to be evaluated only once
-    typedef CoeffBasedProduct<LhsNested, RhsNested, 0> FullyLazyCoeffBaseProductType;
-
-  public:
-
-    typedef typename Base::PlainObject PlainObject;
-
-    ProductBase(const Lhs& a_lhs, const Rhs& a_rhs)
-      : m_lhs(a_lhs), m_rhs(a_rhs)
-    {
-      eigen_assert(a_lhs.cols() == a_rhs.rows()
-        && "invalid matrix product"
-        && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
-    }
-
-    inline Index rows() const { return m_lhs.rows(); }
-    inline Index cols() const { return m_rhs.cols(); }
-
-    template<typename Dest>
-    inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst,Scalar(1)); }
-
-    template<typename Dest>
-    inline void addTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(1)); }
-
-    template<typename Dest>
-    inline void subTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(-1)); }
-
-    template<typename Dest>
-    inline void scaleAndAddTo(Dest& dst, const Scalar& alpha) const { derived().scaleAndAddTo(dst,alpha); }
-
-    const _LhsNested& lhs() const { return m_lhs; }
-    const _RhsNested& rhs() const { return m_rhs; }
-
-    // Implicit conversion to the nested type (trigger the evaluation of the product)
-    operator const PlainObject& () const
-    {
-      m_result.resize(m_lhs.rows(), m_rhs.cols());
-      derived().evalTo(m_result);
-      return m_result;
-    }
-
-    const Diagonal<const FullyLazyCoeffBaseProductType,0> diagonal() const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); }
-
-    template<int Index>
-    const Diagonal<FullyLazyCoeffBaseProductType,Index> diagonal() const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); }
-
-    const Diagonal<FullyLazyCoeffBaseProductType,Dynamic> diagonal(Index index) const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs).diagonal(index); }
-
-    // restrict coeff accessors to 1x1 expressions. No need to care about mutators here since this isn't an Lvalue expression
-    typename Base::CoeffReturnType coeff(Index row, Index col) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      Matrix<Scalar,1,1> result = *this;
-      return result.coeff(row,col);
-    }
-
-    typename Base::CoeffReturnType coeff(Index i) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      Matrix<Scalar,1,1> result = *this;
-      return result.coeff(i);
-    }
-
-    const Scalar& coeffRef(Index row, Index col) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      return derived().coeffRef(row,col);
-    }
-
-    const Scalar& coeffRef(Index i) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      return derived().coeffRef(i);
-    }
-
-  protected:
-
-    LhsNested m_lhs;
-    RhsNested m_rhs;
-
-    mutable PlainObject m_result;
-};
-
-// here we need to overload the nested rule for products
-// such that the nested type is a const reference to a plain matrix
-namespace internal {
-template<typename Lhs, typename Rhs, int Mode, int N, typename PlainObject>
-struct nested<GeneralProduct<Lhs,Rhs,Mode>, N, PlainObject>
-{
-  typedef PlainObject const& type;
-};
-}
-
-template<typename NestedProduct>
-class ScaledProduct;
-
-// Note that these two operator* functions are not defined as member
-// functions of ProductBase, because, otherwise we would have to
-// define all overloads defined in MatrixBase. Furthermore, Using
-// "using Base::operator*" would not work with MSVC.
-//
-// Also note that here we accept any compatible scalar types
-template<typename Derived,typename Lhs,typename Rhs>
-const ScaledProduct<Derived>
-operator*(const ProductBase<Derived,Lhs,Rhs>& prod, const typename Derived::Scalar& x)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-template<typename Derived,typename Lhs,typename Rhs>
-typename internal::enable_if<!internal::is_same<typename Derived::Scalar,typename Derived::RealScalar>::value,
-                      const ScaledProduct<Derived> >::type
-operator*(const ProductBase<Derived,Lhs,Rhs>& prod, const typename Derived::RealScalar& x)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-
-template<typename Derived,typename Lhs,typename Rhs>
-const ScaledProduct<Derived>
-operator*(const typename Derived::Scalar& x,const ProductBase<Derived,Lhs,Rhs>& prod)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-template<typename Derived,typename Lhs,typename Rhs>
-typename internal::enable_if<!internal::is_same<typename Derived::Scalar,typename Derived::RealScalar>::value,
-                      const ScaledProduct<Derived> >::type
-operator*(const typename Derived::RealScalar& x,const ProductBase<Derived,Lhs,Rhs>& prod)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-namespace internal {
-template<typename NestedProduct>
-struct traits<ScaledProduct<NestedProduct> >
- : traits<ProductBase<ScaledProduct<NestedProduct>,
-                         typename NestedProduct::_LhsNested,
-                         typename NestedProduct::_RhsNested> >
-{
-  typedef typename traits<NestedProduct>::StorageKind StorageKind;
-};
-}
-
-template<typename NestedProduct>
-class ScaledProduct
-  : public ProductBase<ScaledProduct<NestedProduct>,
-                       typename NestedProduct::_LhsNested,
-                       typename NestedProduct::_RhsNested>
-{
-  public:
-    typedef ProductBase<ScaledProduct<NestedProduct>,
-                       typename NestedProduct::_LhsNested,
-                       typename NestedProduct::_RhsNested> Base;
-    typedef typename Base::Scalar Scalar;
-    typedef typename Base::PlainObject PlainObject;
-//     EIGEN_PRODUCT_PUBLIC_INTERFACE(ScaledProduct)
-
-    ScaledProduct(const NestedProduct& prod, const Scalar& x)
-    : Base(prod.lhs(),prod.rhs()), m_prod(prod), m_alpha(x) {}
-
-    template<typename Dest>
-    inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst, Scalar(1)); }
-
-    template<typename Dest>
-    inline void addTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(1)); }
-
-    template<typename Dest>
-    inline void subTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(-1)); }
-
-    template<typename Dest>
-    inline void scaleAndAddTo(Dest& dst, const Scalar& a_alpha) const { m_prod.derived().scaleAndAddTo(dst,a_alpha * m_alpha); }
-
-    const Scalar& alpha() const { return m_alpha; }
-    
-  protected:
-    const NestedProduct& m_prod;
-    Scalar m_alpha;
-};
-
-/** \internal
-  * Overloaded to perform an efficient C = (A*B).lazy() */
-template<typename Derived>
-template<typename ProductDerived, typename Lhs, typename Rhs>
-Derived& MatrixBase<Derived>::lazyAssign(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-{
-  other.derived().evalTo(derived());
-  return derived();
-}
-
-} // end namespace Eigen
-
-#endif // EIGEN_PRODUCTBASE_H

Eigen/src/Core/Random.h

@@ -53,7 +53,7 @@ struct functor_traits<scalar_random_op<Scalar> >
   * \sa DenseBase::setRandom(), DenseBase::Random(Index), DenseBase::Random()
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random(Index rows, Index cols)
 {
   return NullaryExpr(rows, cols, internal::scalar_random_op<Scalar>());
@@ -84,7 +84,7 @@ DenseBase<Derived>::Random(Index rows, Index cols)
   * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random()
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random(Index size)
 {
   return NullaryExpr(size, internal::scalar_random_op<Scalar>());
@@ -110,7 +110,7 @@ DenseBase<Derived>::Random(Index size)
   * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random(Index)
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random()
 {
   return NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_random_op<Scalar>());
@@ -162,8 +162,8 @@ PlainObjectBase<Derived>::setRandom(Index newSize)
   *
   * \not_reentrant
   * 
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setRandom_int_int.cpp
   * Output: \verbinclude Matrix_setRandom_int_int.out
@@ -172,9 +172,9 @@ PlainObjectBase<Derived>::setRandom(Index newSize)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setRandom(Index nbRows, Index nbCols)
+PlainObjectBase<Derived>::setRandom(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setRandom();
 }
 

Eigen/src/Core/Redux.h

@@ -65,6 +65,25 @@ public:
               ? CompleteUnrolling
               : NoUnrolling
   };
+  
+#ifdef EIGEN_DEBUG_ASSIGN
+  static void debug()
+  {
+    std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl;
+    std::cerr.setf(std::ios::hex, std::ios::basefield);
+    EIGEN_DEBUG_VAR(Derived::Flags)
+    std::cerr.unsetf(std::ios::hex);
+    EIGEN_DEBUG_VAR(InnerMaxSize)
+    EIGEN_DEBUG_VAR(PacketSize)
+    EIGEN_DEBUG_VAR(MightVectorize)
+    EIGEN_DEBUG_VAR(MayLinearVectorize)
+    EIGEN_DEBUG_VAR(MaySliceVectorize)
+    EIGEN_DEBUG_VAR(Traversal)
+    EIGEN_DEBUG_VAR(UnrollingLimit)
+    EIGEN_DEBUG_VAR(Unrolling)
+    std::cerr << std::endl;
+  }
+#endif
 };
 
 /***************************************************************************
@@ -146,7 +165,7 @@ struct redux_vec_unroller<Func, Derived, Start, 1>
     index = Start * packet_traits<typename Derived::Scalar>::size,
     outer = index / int(Derived::InnerSizeAtCompileTime),
     inner = index % int(Derived::InnerSizeAtCompileTime),
-    alignment = (Derived::Flags & AlignedBit) ? Aligned : Unaligned
+    alignment = Derived::Alignment
   };
 
   typedef typename Derived::Scalar Scalar;
@@ -154,7 +173,7 @@ struct redux_vec_unroller<Func, Derived, Start, 1>
 
   static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&)
   {
-    return mat.template packetByOuterInner<alignment>(outer, inner);
+    return mat.template packetByOuterInner<alignment,PacketScalar>(outer, inner);
   }
 };
 
@@ -172,9 +191,8 @@ template<typename Func, typename Derived>
 struct redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename Derived::Index Index;
   EIGEN_DEVICE_FUNC
-  static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func)
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     Scalar res;
@@ -198,18 +216,18 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
   typedef typename packet_traits<Scalar>::type PacketScalar;
-  typedef typename Derived::Index Index;
 
-  static Scalar run(const Derived& mat, const Func& func)
+  static Scalar run(const Derived &mat, const Func& func)
   {
     const Index size = mat.size();
-    eigen_assert(size && "you are using an empty matrix");
+    
     const Index packetSize = packet_traits<Scalar>::size;
-    const Index alignedStart = internal::first_aligned(mat);
+    const int packetAlignment = unpacket_traits<PacketScalar>::alignment;
     enum {
-      alignment = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) || bool(Derived::Flags & AlignedBit)
-                ? Aligned : Unaligned
+      alignment0 = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned),
+      alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Derived::Alignment)
     };
+    const Index alignedStart = internal::first_default_aligned(mat.nestedExpression());
     const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
     const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
     const Index alignedEnd2 = alignedStart + alignedSize2;
@@ -217,19 +235,19 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
     Scalar res;
     if(alignedSize)
     {
-      PacketScalar packet_res0 = mat.template packet<alignment>(alignedStart);
+      PacketScalar packet_res0 = mat.template packet<alignment,PacketScalar>(alignedStart);
       if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop
       {
-        PacketScalar packet_res1 = mat.template packet<alignment>(alignedStart+packetSize);
+        PacketScalar packet_res1 = mat.template packet<alignment,PacketScalar>(alignedStart+packetSize);
         for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize)
         {
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(index));
-          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment>(index+packetSize));
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(index));
+          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment,PacketScalar>(index+packetSize));
         }
 
         packet_res0 = func.packetOp(packet_res0,packet_res1);
         if(alignedEnd>alignedEnd2)
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(alignedEnd2));
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(alignedEnd2));
       }
       res = func.predux(packet_res0);
 
@@ -255,10 +273,9 @@ template<typename Func, typename Derived>
 struct redux_impl<Func, Derived, SliceVectorizedTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename packet_traits<Scalar>::type PacketScalar;
-  typedef typename Derived::Index Index;
+  typedef typename packet_traits<Scalar>::type PacketType;
 
-  static Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     const Index innerSize = mat.innerSize();
@@ -270,10 +287,10 @@ struct redux_impl<Func, Derived, SliceVectorizedTraversal, NoUnrolling>
     Scalar res;
     if(packetedInnerSize)
     {
-      PacketScalar packet_res = mat.template packet<Unaligned>(0,0);
+      PacketType packet_res = mat.template packet<Unaligned,PacketType>(0,0);
       for(Index j=0; j<outerSize; ++j)
         for(Index i=(j==0?packetSize:0); i<packetedInnerSize; i+=Index(packetSize))
-          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned>(j,i));
+          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned,PacketType>(j,i));
 
       res = func.predux(packet_res);
       for(Index j=0; j<outerSize; ++j)
@@ -300,7 +317,7 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
     Size = Derived::SizeAtCompileTime,
     VectorizedSize = (Size / PacketSize) * PacketSize
   };
-  static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     if (VectorizedSize > 0) {
@@ -315,6 +332,68 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
   }
 };
 
+// evaluator adaptor
+template<typename _XprType>
+class redux_evaluator
+{
+public:
+  typedef _XprType XprType;
+  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketScalar PacketScalar;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  
+  enum {
+    MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = XprType::MaxColsAtCompileTime,
+    // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator
+    Flags = evaluator<XprType>::Flags & ~DirectAccessBit,
+    IsRowMajor = XprType::IsRowMajor,
+    SizeAtCompileTime = XprType::SizeAtCompileTime,
+    InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime,
+    CoeffReadCost = evaluator<XprType>::CoeffReadCost,
+    Alignment = evaluator<XprType>::Alignment
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+  EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); }
+  EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index index) const
+  { return m_evaluator.coeff(index); }
+
+  template<int LoadMode, typename PacketType>
+  PacketReturnType packet(Index row, Index col) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(row, col); }
+
+  template<int LoadMode, typename PacketType>
+  PacketReturnType packet(Index index) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(index); }
+  
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  template<int LoadMode, typename PacketType>
+  PacketReturnType packetByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  const XprType & nestedExpression() const { return m_xpr; }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
+
 } // end namespace internal
 
 /***************************************************************************
@@ -325,18 +404,31 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
 /** \returns the result of a full redux operation on the whole matrix or vector using \a func
   *
   * The template parameter \a BinaryOp is the type of the functor \a func which must be
-  * an associative operator. Both current STL and TR1 functor styles are handled.
+  * an associative operator. Both current C++98 and C++11 functor styles are handled.
   *
   * \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise()
   */
 template<typename Derived>
 template<typename Func>
-EIGEN_STRONG_INLINE typename internal::result_of<Func(typename internal::traits<Derived>::Scalar)>::type
+typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::redux(const Func& func) const
 {
-  typedef typename internal::remove_all<typename Derived::Nested>::type ThisNested;
-  return internal::redux_impl<Func, ThisNested>
-            ::run(derived(), func);
+  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());
+  
+  return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func);
 }
 
 /** \returns the minimum of all coefficients of \c *this.
@@ -346,7 +438,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff() const
 {
-  return this->redux(Eigen::internal::scalar_min_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_min_op<Scalar>());
 }
 
 /** \returns the maximum of all coefficients of \c *this.
@@ -356,7 +448,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff() const
 {
-  return this->redux(Eigen::internal::scalar_max_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_max_op<Scalar>());
 }
 
 /** \returns the sum of all coefficients of *this
@@ -369,7 +461,7 @@ DenseBase<Derived>::sum() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
     return Scalar(0);
-  return this->redux(Eigen::internal::scalar_sum_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_sum_op<Scalar>());
 }
 
 /** \returns the mean of all coefficients of *this
@@ -380,7 +472,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::mean() const
 {
-  return Scalar(this->redux(Eigen::internal::scalar_sum_op<Scalar>())) / Scalar(this->size());
+  return Scalar(derived().redux(Eigen::internal::scalar_sum_op<Scalar>())) / Scalar(this->size());
 }
 
 /** \returns the product of all coefficients of *this
@@ -396,7 +488,7 @@ DenseBase<Derived>::prod() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
     return Scalar(1);
-  return this->redux(Eigen::internal::scalar_product_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_product_op<Scalar>());
 }
 
 /** \returns the trace of \c *this, i.e. the sum of the coefficients on the main diagonal.

Eigen/src/Core/Ref.h

@@ -12,17 +12,13 @@
 
 namespace Eigen { 
 
-template<typename Derived> class RefBase;
-template<typename PlainObjectType, int Options = 0,
-         typename StrideType = typename internal::conditional<PlainObjectType::IsVectorAtCompileTime,InnerStride<1>,OuterStride<> >::type > class Ref;
-
 /** \class Ref
   * \ingroup Core_Module
   *
   * \brief A matrix or vector expression mapping an existing expression
   *
   * \tparam PlainObjectType the equivalent matrix type of the mapped data
-  * \tparam Options specifies whether the pointer is \c #Aligned, or \c #Unaligned.
+  * \tparam MapOptions specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
   *                The default is \c #Unaligned.
   * \tparam StrideType optionally specifies strides. By default, Ref implies a contiguous storage along the inner dimension (inner stride==1),
   *                   but accepts a variable outer stride (leading dimension).
@@ -52,8 +48,9 @@ template<typename PlainObjectType, int Options = 0,
   * VectorXf a;
   * foo1(a.head());             // OK
   * foo1(A.col());              // OK
-  * foo1(A.row());              // compilation error because here innerstride!=1
-  * foo2(A.row());              // The row is copied into a contiguous temporary
+  * foo1(A.row());              // Compilation error because here innerstride!=1
+  * foo2(A.row());              // Compilation error because A.row() is a 1xN object while foo2 is expecting a Nx1 object
+  * foo2(A.row().transpose());  // The row is copied into a contiguous temporary
   * foo2(2*a);                  // The expression is evaluated into a temporary
   * foo2(A.col().segment(2,4)); // No temporary
   * \endcode
@@ -95,7 +92,8 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
   typedef _StrideType StrideType;
   enum {
     Options = _Options,
-    Flags = traits<Map<_PlainObjectType, _Options, _StrideType> >::Flags | NestByRefBit
+    Flags = traits<Map<_PlainObjectType, _Options, _StrideType> >::Flags | NestByRefBit,
+    Alignment = traits<Map<_PlainObjectType, _Options, _StrideType> >::Alignment
   };
 
   template<typename Derived> struct match {
@@ -107,8 +105,9 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
                       || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1),
       OuterStrideMatch = Derived::IsVectorAtCompileTime
                       || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime),
-      AlignmentMatch = (_Options!=Aligned) || ((PlainObjectType::Flags&AlignedBit)==0) || ((traits<Derived>::Flags&AlignedBit)==AlignedBit),
-      MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch
+      AlignmentMatch = (int(traits<PlainObjectType>::Alignment)==int(Unaligned)) || (int(evaluator<Derived>::Alignment) >= int(Alignment)), // FIXME the first condition is not very clear, it should be replaced by the required alignment
+      ScalarTypeMatch = internal::is_same<typename PlainObjectType::Scalar, typename Derived::Scalar>::value,
+      MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch && ScalarTypeMatch
     };
     typedef typename internal::conditional<MatchAtCompileTime,internal::true_type,internal::false_type>::type type;
   };
@@ -131,12 +130,12 @@ public:
   typedef MapBase<Derived> Base;
   EIGEN_DENSE_PUBLIC_INTERFACE(RefBase)
 
-  inline Index innerStride() const
+  EIGEN_DEVICE_FUNC inline Index innerStride() const
   {
     return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1;
   }
 
-  inline Index outerStride() const
+  EIGEN_DEVICE_FUNC inline Index outerStride() const
   {
     return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
          : IsVectorAtCompileTime ? this->size()
@@ -144,7 +143,7 @@ public:
          : this->rows();
   }
 
-  RefBase()
+  EIGEN_DEVICE_FUNC RefBase()
     : Base(0,RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime),
       // Stride<> does not allow default ctor for Dynamic strides, so let' initialize it with dummy values:
       m_stride(StrideType::OuterStrideAtCompileTime==Dynamic?0:StrideType::OuterStrideAtCompileTime,
@@ -158,7 +157,7 @@ protected:
   typedef Stride<StrideType::OuterStrideAtCompileTime,StrideType::InnerStrideAtCompileTime> StrideBase;
 
   template<typename Expression>
-  void construct(Expression& expr)
+  EIGEN_DEVICE_FUNC void construct(Expression& expr)
   {
     if(PlainObjectType::RowsAtCompileTime==1)
     {
@@ -187,7 +186,11 @@ protected:
 template<typename PlainObjectType, int Options, typename StrideType> class Ref
   : public RefBase<Ref<PlainObjectType, Options, StrideType> >
 {
+  private:
     typedef internal::traits<Ref> Traits;
+    template<typename Derived>
+    EIGEN_DEVICE_FUNC inline Ref(const PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0);
   public:
 
     typedef RefBase<Ref> Base;
@@ -196,20 +199,23 @@ template<typename PlainObjectType, int Options, typename StrideType> class Ref
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Derived>
-    inline Ref(PlainObjectBase<Derived>& expr,
-               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
+    EIGEN_DEVICE_FUNC inline Ref(PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     {
-      Base::construct(expr);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      Base::construct(expr.derived());
     }
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr,
-               typename internal::enable_if<bool(internal::is_lvalue<Derived>::value&&bool(Traits::template match<Derived>::MatchAtCompileTime)),Derived>::type* = 0,
-               int = Derived::ThisConstantIsPrivateInPlainObjectBase)
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     #else
     template<typename Derived>
     inline Ref(DenseBase<Derived>& expr)
     #endif
     {
+      EIGEN_STATIC_ASSERT(bool(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      EIGEN_STATIC_ASSERT(!Derived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
       Base::construct(expr.const_cast_derived());
     }
 
@@ -228,7 +234,8 @@ template<typename TPlainObjectType, int Options, typename StrideType> class Ref<
     EIGEN_DENSE_PUBLIC_INTERFACE(Ref)
 
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr)
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::ScalarTypeMatch),Derived>::type* = 0)
     {
 //      std::cout << match_helper<Derived>::HasDirectAccess << "," << match_helper<Derived>::OuterStrideMatch << "," << match_helper<Derived>::InnerStrideMatch << "\n";
 //      std::cout << int(StrideType::OuterStrideAtCompileTime) << " - " << int(Derived::OuterStrideAtCompileTime) << "\n";
@@ -236,18 +243,27 @@ template<typename TPlainObjectType, int Options, typename StrideType> class Ref<
       construct(expr.derived(), typename Traits::template match<Derived>::type());
     }
 
+    EIGEN_DEVICE_FUNC inline Ref(const Ref& other) : Base(other) {
+      // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy
+    }
+
+    template<typename OtherRef>
+    EIGEN_DEVICE_FUNC inline Ref(const RefBase<OtherRef>& other) {
+      construct(other.derived(), typename Traits::template match<OtherRef>::type());
+    }
+
   protected:
 
     template<typename Expression>
-    void construct(const Expression& expr,internal::true_type)
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr,internal::true_type)
     {
       Base::construct(expr);
     }
 
     template<typename Expression>
-    void construct(const Expression& expr, internal::false_type)
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr, internal::false_type)
     {
-      m_object.lazyAssign(expr);
+      internal::call_assignment_no_alias(m_object,expr,internal::assign_op<Scalar>());
       Base::construct(m_object);
     }
 

Eigen/src/Core/Replicate.h

@@ -35,10 +35,7 @@ struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
   typedef typename MatrixType::Scalar Scalar;
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
-  enum {
-    Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor
-  };
-  typedef typename nested<MatrixType,Factor>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = RowFactor==Dynamic || int(MatrixType::RowsAtCompileTime)==Dynamic
@@ -53,8 +50,9 @@ struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
     IsRowMajor = MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1 ? 1
                : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0
                : (MatrixType::Flags & RowMajorBit) ? 1 : 0,
-    Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0),
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost
+    
+    // FIXME enable DirectAccess with negative strides?
+    Flags = IsRowMajor ? RowMajorBit : 0
   };
 };
 }
@@ -68,10 +66,12 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
 
     typedef typename internal::dense_xpr_base<Replicate>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Replicate)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
     template<typename OriginalMatrixType>
-    inline explicit Replicate(const OriginalMatrixType& a_matrix)
-      : m_matrix(a_matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor)
+    EIGEN_DEVICE_FUNC
+    inline explicit Replicate(const OriginalMatrixType& matrix)
+      : m_matrix(matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor)
     {
       EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
                           THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
@@ -79,41 +79,20 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
     }
 
     template<typename OriginalMatrixType>
-    inline Replicate(const OriginalMatrixType& a_matrix, Index rowFactor, Index colFactor)
-      : m_matrix(a_matrix), m_rowFactor(rowFactor), m_colFactor(colFactor)
+    EIGEN_DEVICE_FUNC
+    inline Replicate(const OriginalMatrixType& matrix, Index rowFactor, Index colFactor)
+      : m_matrix(matrix), m_rowFactor(rowFactor), m_colFactor(colFactor)
     {
       EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
                           THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
     }
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows() * m_rowFactor.value(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_matrix.cols() * m_colFactor.value(); }
 
-    inline Scalar coeff(Index rowId, Index colId) const
-    {
-      // try to avoid using modulo; this is a pure optimization strategy
-      const Index actual_row  = internal::traits<MatrixType>::RowsAtCompileTime==1 ? 0
-                            : RowFactor==1 ? rowId
-                            : rowId%m_matrix.rows();
-      const Index actual_col  = internal::traits<MatrixType>::ColsAtCompileTime==1 ? 0
-                            : ColFactor==1 ? colId
-                            : colId%m_matrix.cols();
-
-      return m_matrix.coeff(actual_row, actual_col);
-    }
-    template<int LoadMode>
-    inline PacketScalar packet(Index rowId, Index colId) const
-    {
-      const Index actual_row  = internal::traits<MatrixType>::RowsAtCompileTime==1 ? 0
-                            : RowFactor==1 ? rowId
-                            : rowId%m_matrix.rows();
-      const Index actual_col  = internal::traits<MatrixType>::ColsAtCompileTime==1 ? 0
-                            : ColFactor==1 ? colId
-                            : colId%m_matrix.cols();
-
-      return m_matrix.template packet<LoadMode>(actual_row, actual_col);
-    }
-
+    EIGEN_DEVICE_FUNC
     const _MatrixTypeNested& nestedExpression() const
     { 
       return m_matrix; 
@@ -135,27 +114,12 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
   */
 template<typename Derived>
 template<int RowFactor, int ColFactor>
-inline const Replicate<Derived,RowFactor,ColFactor>
+const Replicate<Derived,RowFactor,ColFactor>
 DenseBase<Derived>::replicate() const
 {
   return Replicate<Derived,RowFactor,ColFactor>(derived());
 }
 
-/**
-  * \return an expression of the replication of \c *this
-  *
-  * Example: \include MatrixBase_replicate_int_int.cpp
-  * Output: \verbinclude MatrixBase_replicate_int_int.out
-  *
-  * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate
-  */
-template<typename Derived>
-inline const Replicate<Derived,Dynamic,Dynamic>
-DenseBase<Derived>::replicate(Index rowFactor,Index colFactor) const
-{
-  return Replicate<Derived,Dynamic,Dynamic>(derived(),rowFactor,colFactor);
-}
-
 /**
   * \return an expression of the replication of each column (or row) of \c *this
   *

Eigen/src/Core/ReturnByValue.h

@@ -38,9 +38,10 @@ struct traits<ReturnByValue<Derived> >
  * So internal::nested always gives the plain return matrix type.
  *
  * FIXME: I don't understand why we need this specialization: isn't this taken care of by the EvalBeforeNestingBit ??
+ * Answer: EvalBeforeNestingBit should be deprecated since we have the evaluators
  */
 template<typename Derived,int n,typename PlainObject>
-struct nested<ReturnByValue<Derived>, n, PlainObject>
+struct nested_eval<ReturnByValue<Derived>, n, PlainObject>
 {
   typedef typename traits<Derived>::ReturnType type;
 };
@@ -48,7 +49,7 @@ struct nested<ReturnByValue<Derived>, n, PlainObject>
 } // end namespace internal
 
 template<typename Derived> class ReturnByValue
-  : internal::no_assignment_operator, public internal::dense_xpr_base< ReturnByValue<Derived> >::type
+  : public internal::dense_xpr_base< ReturnByValue<Derived> >::type, internal::no_assignment_operator
 {
   public:
     typedef typename internal::traits<Derived>::ReturnType ReturnType;
@@ -73,6 +74,7 @@ template<typename Derived> class ReturnByValue
     const Unusable& coeff(Index,Index) const { return *reinterpret_cast<const Unusable*>(this); }
     Unusable& coeffRef(Index) { return *reinterpret_cast<Unusable*>(this); }
     Unusable& coeffRef(Index,Index) { return *reinterpret_cast<Unusable*>(this); }
+#undef Unusable
 #endif
 };
 
@@ -84,6 +86,33 @@ Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
   return derived();
 }
 
+namespace internal {
+
+// Expression is evaluated in a temporary; default implementation of Assignment is bypassed so that
+// when a ReturnByValue expression is assigned, the evaluator is not constructed.
+// TODO: Finalize port to new regime; ReturnByValue should not exist in the expression world
+  
+template<typename Derived>
+struct evaluator<ReturnByValue<Derived> >
+  : public evaluator<typename internal::traits<Derived>::ReturnType>
+{
+  typedef ReturnByValue<Derived> XprType;
+  typedef typename internal::traits<Derived>::ReturnType PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    xpr.evalTo(m_result);
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_RETURNBYVALUE_H

Eigen/src/Core/Reverse.h

@@ -37,32 +37,25 @@ struct traits<Reverse<MatrixType, Direction> >
   typedef typename MatrixType::Scalar Scalar;
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
     ColsAtCompileTime = MatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
-
-    // let's enable LinearAccess only with vectorization because of the product overhead
-    LinearAccess = ( (Direction==BothDirections) && (int(_MatrixTypeNested::Flags)&PacketAccessBit) )
-                 ? LinearAccessBit : 0,
-
-    Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess),
-
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost
+    Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit)
   };
 };
 
-template<typename PacketScalar, bool ReversePacket> struct reverse_packet_cond
+template<typename PacketType, bool ReversePacket> struct reverse_packet_cond
 {
-  static inline PacketScalar run(const PacketScalar& x) { return preverse(x); }
+  static inline PacketType run(const PacketType& x) { return preverse(x); }
 };
 
-template<typename PacketScalar> struct reverse_packet_cond<PacketScalar,false>
+template<typename PacketType> struct reverse_packet_cond<PacketType,false>
 {
-  static inline PacketScalar run(const PacketScalar& x) { return x; }
+  static inline PacketType run(const PacketType& x) { return x; }
 };
 
 } // end namespace internal 
@@ -74,12 +67,9 @@ template<typename MatrixType, int Direction> class Reverse
 
     typedef typename internal::dense_xpr_base<Reverse>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Reverse)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
     using Base::IsRowMajor;
 
-    // next line is necessary because otherwise const version of operator()
-    // is hidden by non-const version defined in this file
-    using Base::operator(); 
-
   protected:
     enum {
       PacketSize = internal::packet_traits<Scalar>::size,
@@ -95,82 +85,19 @@ template<typename MatrixType, int Direction> class Reverse
     typedef internal::reverse_packet_cond<PacketScalar,ReversePacket> reverse_packet;
   public:
 
-    inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
+    EIGEN_DEVICE_FUNC explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse)
 
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); }
 
-    inline Index innerStride() const
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
     {
       return -m_matrix.innerStride();
     }
 
-    inline Scalar& operator()(Index row, Index col)
-    {
-      eigen_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
-      return coeffRef(row, col);
-    }
-
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      return m_matrix.const_cast_derived().coeffRef(ReverseRow ? m_matrix.rows() - row - 1 : row,
-                                                    ReverseCol ? m_matrix.cols() - col - 1 : col);
-    }
-
-    inline CoeffReturnType coeff(Index row, Index col) const
-    {
-      return m_matrix.coeff(ReverseRow ? m_matrix.rows() - row - 1 : row,
-                            ReverseCol ? m_matrix.cols() - col - 1 : col);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_matrix.coeff(m_matrix.size() - index - 1);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_matrix.const_cast_derived().coeffRef(m_matrix.size() - index - 1);
-    }
-
-    inline Scalar& operator()(Index index)
-    {
-      eigen_assert(index >= 0 && index < m_matrix.size());
-      return coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
-    {
-      return reverse_packet::run(m_matrix.template packet<LoadMode>(
-                                    ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
-                                    ReverseCol ? m_matrix.cols() - col - OffsetCol : col));
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(
-                                      ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
-                                      ReverseCol ? m_matrix.cols() - col - OffsetCol : col,
-                                      reverse_packet::run(x));
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return internal::preverse(m_matrix.template packet<LoadMode>( m_matrix.size() - index - PacketSize ));
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(m_matrix.size() - index - PacketSize, internal::preverse(x));
-    }
-
-    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    EIGEN_DEVICE_FUNC const typename internal::remove_all<typename MatrixType::Nested>::type&
     nestedExpression() const 
     {
       return m_matrix;
@@ -190,33 +117,93 @@ template<typename Derived>
 inline typename DenseBase<Derived>::ReverseReturnType
 DenseBase<Derived>::reverse()
 {
-  return derived();
+  return ReverseReturnType(derived());
 }
 
-/** This is the const version of reverse(). */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstReverseReturnType
-DenseBase<Derived>::reverse() const
-{
-  return derived();
-}
+
+//reverse const overload moved DenseBase.h due to a CUDA compiler bug
 
 /** This is the "in place" version of reverse: it reverses \c *this.
   *
   * In most cases it is probably better to simply use the reversed expression
   * of a matrix. However, when reversing the matrix data itself is really needed,
   * then this "in-place" version is probably the right choice because it provides
-  * the following additional features:
+  * the following additional benefits:
   *  - less error prone: doing the same operation with .reverse() requires special care:
   *    \code m = m.reverse().eval(); \endcode
-  *  - this API allows to avoid creating a temporary (the current implementation creates a temporary, but that could be avoided using swap)
+  *  - this API enables reverse operations without the need for a temporary
   *  - it allows future optimizations (cache friendliness, etc.)
   *
-  * \sa reverse() */
+  * \sa VectorwiseOp::reverseInPlace(), reverse() */
 template<typename Derived>
 inline void DenseBase<Derived>::reverseInPlace()
 {
-  derived() = derived().reverse().eval();
+  if(cols()>rows())
+  {
+    Index half = cols()/2;
+    leftCols(half).swap(rightCols(half).reverse());
+    if((cols()%2)==1)
+    {
+      Index half2 = rows()/2;
+      col(half).head(half2).swap(col(half).tail(half2).reverse());
+    }
+  }
+  else
+  {
+    Index half = rows()/2;
+    topRows(half).swap(bottomRows(half).reverse());
+    if((rows()%2)==1)
+    {
+      Index half2 = cols()/2;
+      row(half).head(half2).swap(row(half).tail(half2).reverse());
+    }
+  }
+}
+
+namespace internal {
+  
+template<int Direction>
+struct vectorwise_reverse_inplace_impl;
+
+template<>
+struct vectorwise_reverse_inplace_impl<Vertical>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.rows()/2;
+    xpr.topRows(half).swap(xpr.bottomRows(half).colwise().reverse());
+  }
+};
+
+template<>
+struct vectorwise_reverse_inplace_impl<Horizontal>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.cols()/2;
+    xpr.leftCols(half).swap(xpr.rightCols(half).rowwise().reverse());
+  }
+};
+
+} // end namespace internal
+
+/** This is the "in place" version of VectorwiseOp::reverse: it reverses each column or row of \c *this.
+  *
+  * In most cases it is probably better to simply use the reversed expression
+  * of a matrix. However, when reversing the matrix data itself is really needed,
+  * then this "in-place" version is probably the right choice because it provides
+  * the following additional benefits:
+  *  - less error prone: doing the same operation with .reverse() requires special care:
+  *    \code m = m.reverse().eval(); \endcode
+  *  - this API enables reverse operations without the need for a temporary
+  *
+  * \sa DenseBase::reverseInPlace(), reverse() */
+template<typename ExpressionType, int Direction>
+void VectorwiseOp<ExpressionType,Direction>::reverseInPlace()
+{
+  internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/Select.h

@@ -43,23 +43,21 @@ struct traits<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
     ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime,
-    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits,
-    CoeffReadCost = traits<typename remove_all<ConditionMatrixNested>::type>::CoeffReadCost
-                  + EIGEN_SIZE_MAX(traits<typename remove_all<ThenMatrixNested>::type>::CoeffReadCost,
-                                   traits<typename remove_all<ElseMatrixNested>::type>::CoeffReadCost)
+    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & RowMajorBit
   };
 };
 }
 
 template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
-class Select : internal::no_assignment_operator,
-  public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type
+class Select : public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type,
+               internal::no_assignment_operator
 {
   public:
 
     typedef typename internal::dense_xpr_base<Select>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Select)
 
+    inline EIGEN_DEVICE_FUNC
     Select(const ConditionMatrixType& a_conditionMatrix,
            const ThenMatrixType& a_thenMatrix,
            const ElseMatrixType& a_elseMatrix)
@@ -69,9 +67,10 @@ class Select : internal::no_assignment_operator,
       eigen_assert(m_condition.cols() == m_then.cols() && m_condition.cols() == m_else.cols());
     }
 
-    Index rows() const { return m_condition.rows(); }
-    Index cols() const { return m_condition.cols(); }
+    inline EIGEN_DEVICE_FUNC Index rows() const { return m_condition.rows(); }
+    inline EIGEN_DEVICE_FUNC Index cols() const { return m_condition.cols(); }
 
+    inline EIGEN_DEVICE_FUNC
     const Scalar coeff(Index i, Index j) const
     {
       if (m_condition.coeff(i,j))
@@ -80,6 +79,7 @@ class Select : internal::no_assignment_operator,
         return m_else.coeff(i,j);
     }
 
+    inline EIGEN_DEVICE_FUNC
     const Scalar coeff(Index i) const
     {
       if (m_condition.coeff(i))
@@ -88,17 +88,17 @@ class Select : internal::no_assignment_operator,
         return m_else.coeff(i);
     }
 
-    const ConditionMatrixType& conditionMatrix() const
+    inline EIGEN_DEVICE_FUNC const ConditionMatrixType& conditionMatrix() const
     {
       return m_condition;
     }
 
-    const ThenMatrixType& thenMatrix() const
+    inline EIGEN_DEVICE_FUNC const ThenMatrixType& thenMatrix() const
     {
       return m_then;
     }
 
-    const ElseMatrixType& elseMatrix() const
+    inline EIGEN_DEVICE_FUNC const ElseMatrixType& elseMatrix() const
     {
       return m_else;
     }

Eigen/src/Core/SelfAdjointView.h

@@ -32,45 +32,42 @@ namespace internal {
 template<typename MatrixType, unsigned int UpLo>
 struct traits<SelfAdjointView<MatrixType, UpLo> > : traits<MatrixType>
 {
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
   typedef MatrixType ExpressionType;
-  typedef typename MatrixType::PlainObject DenseMatrixType;
+  typedef typename MatrixType::PlainObject FullMatrixType;
   enum {
     Mode = UpLo | SelfAdjoint,
-    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits)
-           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)), // FIXME these flags should be preserved
-    CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits|FlagsLvalueBit)
+           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved
   };
 };
 }
 
-template <typename Lhs, int LhsMode, bool LhsIsVector,
-          typename Rhs, int RhsMode, bool RhsIsVector>
-struct SelfadjointProductMatrix;
-
 // FIXME could also be called SelfAdjointWrapper to be consistent with DiagonalWrapper ??
-template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
-  : public TriangularBase<SelfAdjointView<MatrixType, UpLo> >
+template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
+  : public TriangularBase<SelfAdjointView<_MatrixType, UpLo> >
 {
   public:
 
+    typedef _MatrixType MatrixType;
     typedef TriangularBase<SelfAdjointView> Base;
     typedef typename internal::traits<SelfAdjointView>::MatrixTypeNested MatrixTypeNested;
     typedef typename internal::traits<SelfAdjointView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned;
 
     /** \brief The type of coefficients in this matrix */
     typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; 
-
-    typedef typename MatrixType::Index Index;
+    typedef typename MatrixType::StorageIndex StorageIndex;
 
     enum {
-      Mode = internal::traits<SelfAdjointView>::Mode
+      Mode = internal::traits<SelfAdjointView>::Mode,
+      Flags = internal::traits<SelfAdjointView>::Flags
     };
     typedef typename MatrixType::PlainObject PlainObject;
 
     EIGEN_DEVICE_FUNC
-    inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
+    explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
     {}
 
     EIGEN_DEVICE_FUNC
@@ -98,6 +95,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index row, Index col)
     {
+      EIGEN_STATIC_ASSERT_LVALUE(SelfAdjointView);
       Base::check_coordinates_internal(row, col);
       return m_matrix.const_cast_derived().coeffRef(row, col);
     }
@@ -111,26 +109,29 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     EIGEN_DEVICE_FUNC
     MatrixTypeNestedCleaned& nestedExpression() { return *const_cast<MatrixTypeNestedCleaned*>(&m_matrix); }
 
-    /** Efficient self-adjoint matrix times vector/matrix product */
+    /** Efficient triangular matrix times vector/matrix product */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC
-    SelfadjointProductMatrix<MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
+    const Product<SelfAdjointView,OtherDerived>
     operator*(const MatrixBase<OtherDerived>& rhs) const
     {
-      return SelfadjointProductMatrix
-              <MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
-              (m_matrix, rhs.derived());
+      return Product<SelfAdjointView,OtherDerived>(*this, rhs.derived());
     }
 
-    /** Efficient vector/matrix times self-adjoint matrix product */
+    /** Efficient vector/matrix times triangular matrix product */
     template<typename OtherDerived> friend
     EIGEN_DEVICE_FUNC
-    SelfadjointProductMatrix<OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
+    const Product<OtherDerived,SelfAdjointView>
     operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView& rhs)
     {
-      return SelfadjointProductMatrix
-              <OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
-              (lhs.derived(),rhs.m_matrix);
+      return Product<OtherDerived,SelfAdjointView>(lhs.derived(),rhs);
+    }
+    
+    friend EIGEN_DEVICE_FUNC
+    const SelfAdjointView<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>,MatrixType>,UpLo>
+    operator*(const Scalar& s, const SelfAdjointView& mat)
+    {
+      return (s*mat.nestedExpression()).template selfadjointView<UpLo>();
     }
 
     /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this:
@@ -194,96 +195,56 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
 
 namespace internal {
 
-template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount, ClearOpposite>
+// TODO currently a selfadjoint expression has the form SelfAdjointView<.,.>
+//      in the future selfadjoint-ness should be defined by the expression traits
+//      such that Transpose<SelfAdjointView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work)
+template<typename MatrixType, unsigned int Mode>
+struct evaluator_traits<SelfAdjointView<MatrixType,Mode> >
 {
-  enum {
-    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
-  };
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef SelfAdjointShape Shape;
+  
+  static const int AssumeAliasing = 0;
+};
 
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &dst, const Derived2 &src)
+template<int UpLo, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version>
+class triangular_dense_assignment_kernel<UpLo,SelfAdjoint,SetOpposite,DstEvaluatorTypeT,SrcEvaluatorTypeT,Functor,Version>
+  : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version>
+{
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base;
+  typedef typename Base::DstXprType DstXprType;
+  typedef typename Base::SrcXprType SrcXprType;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+public:
+  
+  typedef typename Base::DstEvaluatorType DstEvaluatorType;
+  typedef typename Base::SrcEvaluatorType SrcEvaluatorType;
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::AssignmentTraits AssignmentTraits;
+  
+  
+  EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
   {
-    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount-1, ClearOpposite>::run(dst, src);
-
-    if(row == col)
-      dst.coeffRef(row, col) = numext::real(src.coeff(row, col));
-    else if(row < col)
-      dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col));
+    eigen_internal_assert(row!=col);
+    Scalar tmp = m_src.coeff(row,col);
+    m_functor.assignCoeff(m_dst.coeffRef(row,col), tmp);
+    m_functor.assignCoeff(m_dst.coeffRef(col,row), numext::conj(tmp));
   }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, 0, ClearOpposite>
-{
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount, ClearOpposite>
-{
-  enum {
-    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
-  };
-
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &dst, const Derived2 &src)
+  
+  EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id)
   {
-    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount-1, ClearOpposite>::run(dst, src);
-
-    if(row == col)
-      dst.coeffRef(row, col) = numext::real(src.coeff(row, col));
-    else if(row > col)
-      dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col));
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, 0, ClearOpposite>
-{
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      for(Index i = 0; i < j; ++i)
-      {
-        dst.copyCoeff(i, j, src);
-        dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j));
-      }
-      dst.copyCoeff(j, j, src);
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, Dynamic, ClearOpposite>
-{
-  EIGEN_DEVICE_FUNC
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-  typedef typename Derived1::Index Index;
-    for(Index i = 0; i < dst.rows(); ++i)
-    {
-      for(Index j = 0; j < i; ++j)
-      {
-        dst.copyCoeff(i, j, src);
-        dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j));
-      }
-      dst.copyCoeff(i, i, src);
-    }
+    Base::assignCoeff(id,id);
   }
+  
+  EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index, Index)
+  { eigen_internal_assert(false && "should never be called"); }
 };
 
 } // end namespace internal
@@ -297,7 +258,7 @@ template<unsigned int UpLo>
 typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView() const
 {
-  return derived();
+  return typename ConstSelfAdjointViewReturnType<UpLo>::Type(derived());
 }
 
 template<typename Derived>
@@ -305,7 +266,7 @@ template<unsigned int UpLo>
 typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView()
 {
-  return derived();
+  return typename SelfAdjointViewReturnType<UpLo>::Type(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/SelfCwiseBinaryOp.h

@@ -12,179 +12,11 @@
 
 namespace Eigen { 
 
-/** \class SelfCwiseBinaryOp
-  * \ingroup Core_Module
-  *
-  * \internal
-  *
-  * \brief Internal helper class for optimizing operators like +=, -=
-  *
-  * This is a pseudo expression class re-implementing the copyCoeff/copyPacket
-  * method to directly performs a +=/-= operations in an optimal way. In particular,
-  * this allows to make sure that the input/output data are loaded only once using
-  * aligned packet loads.
-  *
-  * \sa class SwapWrapper for a similar trick.
-  */
-
-namespace internal {
-template<typename BinaryOp, typename Lhs, typename Rhs>
-struct traits<SelfCwiseBinaryOp<BinaryOp,Lhs,Rhs> >
-  : traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >
-{
-  enum {
-    // Note that it is still a good idea to preserve the DirectAccessBit
-    // so that assign can correctly align the data.
-    Flags = traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >::Flags | (Lhs::Flags&AlignedBit) | (Lhs::Flags&DirectAccessBit) | (Lhs::Flags&LvalueBit),
-    OuterStrideAtCompileTime = Lhs::OuterStrideAtCompileTime,
-    InnerStrideAtCompileTime = Lhs::InnerStrideAtCompileTime
-  };
-};
-}
-
-template<typename BinaryOp, typename Lhs, typename Rhs> class SelfCwiseBinaryOp
-  : public internal::dense_xpr_base< SelfCwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
-{
-  public:
-
-    typedef typename internal::dense_xpr_base<SelfCwiseBinaryOp>::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(SelfCwiseBinaryOp)
-
-    typedef typename internal::packet_traits<Scalar>::type Packet;
-
-    EIGEN_DEVICE_FUNC
-    inline SelfCwiseBinaryOp(Lhs& xpr, const BinaryOp& func = BinaryOp()) : m_matrix(xpr), m_functor(func) {}
-
-    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); }
-    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); }
-    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_matrix.outerStride(); }
-    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.innerStride(); }
-    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_matrix.data(); }
-
-    // note that this function is needed by assign to correctly align loads/stores
-    // TODO make Assign use .data()
-    EIGEN_DEVICE_FUNC
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-    EIGEN_DEVICE_FUNC
-    inline const Scalar& coeffRef(Index row, Index col) const
-    {
-      return m_matrix.coeffRef(row, col);
-    }
-
-    // note that this function is needed by assign to correctly align loads/stores
-    // TODO make Assign use .data()
-    EIGEN_DEVICE_FUNC
-    inline Scalar& coeffRef(Index index)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-    EIGEN_DEVICE_FUNC
-    inline const Scalar& coeffRef(Index index) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(row >= 0 && row < rows()
-                         && col >= 0 && col < cols());
-      Scalar& tmp = m_matrix.coeffRef(row,col);
-      tmp = m_functor(tmp, _other.coeff(row,col));
-    }
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_matrix.size());
-      Scalar& tmp = m_matrix.coeffRef(index);
-      tmp = m_functor(tmp, _other.coeff(index));
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      m_matrix.template writePacket<StoreMode>(row, col,
-        m_functor.packetOp(m_matrix.template packet<StoreMode>(row, col),_other.template packet<LoadMode>(row, col)) );
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_matrix.size());
-      m_matrix.template writePacket<StoreMode>(index,
-        m_functor.packetOp(m_matrix.template packet<StoreMode>(index),_other.template packet<LoadMode>(index)) );
-    }
-
-    // reimplement lazyAssign to handle complex *= real
-    // see CwiseBinaryOp ctor for details
-    template<typename RhsDerived>
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE SelfCwiseBinaryOp& lazyAssign(const DenseBase<RhsDerived>& rhs)
-    {
-      EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs,RhsDerived)
-      EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename RhsDerived::Scalar);
-      
-    #ifdef EIGEN_DEBUG_ASSIGN
-      internal::assign_traits<SelfCwiseBinaryOp, RhsDerived>::debug();
-    #endif
-      eigen_assert(rows() == rhs.rows() && cols() == rhs.cols());
-      internal::assign_impl<SelfCwiseBinaryOp, RhsDerived>::run(*this,rhs.derived());
-    #ifndef EIGEN_NO_DEBUG
-      this->checkTransposeAliasing(rhs.derived());
-    #endif
-      return *this;
-    }
-    
-    // overloaded to honor evaluation of special matrices
-    // maybe another solution would be to not use SelfCwiseBinaryOp
-    // at first...
-    EIGEN_DEVICE_FUNC
-    SelfCwiseBinaryOp& operator=(const Rhs& _rhs)
-    {
-      typename internal::nested<Rhs>::type rhs(_rhs);
-      return Base::operator=(rhs);
-    }
-
-    EIGEN_DEVICE_FUNC
-    Lhs& expression() const 
-    { 
-      return m_matrix;
-    }
-
-    EIGEN_DEVICE_FUNC
-    const BinaryOp& functor() const 
-    { 
-      return m_functor;
-    }
-
-  protected:
-    Lhs& m_matrix;
-    const BinaryOp& m_functor;
-
-  private:
-    SelfCwiseBinaryOp& operator=(const SelfCwiseBinaryOp&);
-};
-
 template<typename Derived>
 inline Derived& DenseBase<Derived>::operator*=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<internal::scalar_product_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  tmp = PlainObject::Constant(rows(),cols(),other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op<Scalar>());
   return derived();
 }
 
@@ -192,8 +24,7 @@ template<typename Derived>
 inline Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  tmp = PlainObject::Constant(rows(),cols(),other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op<Scalar>());
   return derived();
 }
 
@@ -201,8 +32,7 @@ template<typename Derived>
 inline Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  tmp = PlainObject::Constant(rows(),cols(),other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op<Scalar>());
   return derived();
 }
 
@@ -210,8 +40,7 @@ template<typename Derived>
 inline Derived& DenseBase<Derived>::operator/=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  tmp = PlainObject::Constant(rows(),cols(), other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op<Scalar>());
   return derived();
 }
 

Eigen/src/Core/Solve.h

@@ -0,0 +1,150 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 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_SOLVE_H
+#define EIGEN_SOLVE_H
+
+namespace Eigen {
+
+template<typename Decomposition, typename RhsType, typename StorageKind> class SolveImpl;
+  
+/** \class Solve
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression representing a solving operation
+  *
+  * \tparam Decomposition the type of the matrix or decomposion object
+  * \tparam Rhstype the type of the right-hand side
+  *
+  * This class represents an expression of A.solve(B)
+  * and most of the time this is the only way it is used.
+  *
+  */
+namespace internal {
+
+// this solve_traits class permits to determine the evaluation type with respect to storage kind (Dense vs Sparse)
+template<typename Decomposition, typename RhsType,typename StorageKind> struct solve_traits;
+
+template<typename Decomposition, typename RhsType>
+struct solve_traits<Decomposition,RhsType,Dense>
+{
+  typedef typename Decomposition::MatrixType MatrixType;
+  typedef Matrix<typename RhsType::Scalar,
+                 MatrixType::ColsAtCompileTime,
+                 RhsType::ColsAtCompileTime,
+                 RhsType::PlainObject::Options,
+                 MatrixType::MaxColsAtCompileTime,
+                 RhsType::MaxColsAtCompileTime> PlainObject;  
+};
+
+template<typename Decomposition, typename RhsType>
+struct traits<Solve<Decomposition, RhsType> >
+  : traits<typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject>
+{
+  typedef typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject PlainObject;
+  typedef typename promote_index_type<typename Decomposition::StorageIndex, typename RhsType::StorageIndex>::type StorageIndex;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit,
+    CoeffReadCost = Dynamic
+  };
+};
+
+}
+
+
+template<typename Decomposition, typename RhsType>
+class Solve : public SolveImpl<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>
+{
+public:
+  typedef typename internal::traits<Solve>::PlainObject PlainObject;
+  typedef typename internal::traits<Solve>::StorageIndex StorageIndex;
+  
+  Solve(const Decomposition &dec, const RhsType &rhs)
+    : m_dec(dec), m_rhs(rhs)
+  {}
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); }
+
+  EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; }
+  EIGEN_DEVICE_FUNC const RhsType&       rhs() const { return m_rhs; }
+
+protected:
+  const Decomposition &m_dec;
+  const RhsType       &m_rhs;
+};
+
+
+// Specialization of the Solve expression for dense results
+template<typename Decomposition, typename RhsType>
+class SolveImpl<Decomposition,RhsType,Dense>
+  : public MatrixBase<Solve<Decomposition,RhsType> >
+{
+  typedef Solve<Decomposition,RhsType> Derived;
+  
+public:
+  
+  typedef MatrixBase<Solve<Decomposition,RhsType> > Base;
+  EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+
+private:
+  
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+// Generic API dispatcher
+template<typename Decomposition, typename RhsType, typename StorageKind>
+class SolveImpl : public internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type Base;
+};
+
+namespace internal {
+
+// Evaluator of Solve -> eval into a temporary
+template<typename Decomposition, typename RhsType>
+struct evaluator<Solve<Decomposition,RhsType> >
+  : public evaluator<typename Solve<Decomposition,RhsType>::PlainObject>
+{
+  typedef Solve<Decomposition,RhsType> SolveType;
+  typedef typename SolveType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const SolveType& solve)
+    : m_result(solve.rows(), solve.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    solve.dec()._solve_impl(solve.rhs(), m_result);
+  }
+  
+protected:  
+  PlainObject m_result;
+};
+
+// Specialization for "dst = dec.solve(rhs)"
+// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar>, Dense2Dense, Scalar>
+{
+  typedef Solve<DecType,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar> &)
+  {
+    // FIXME shall we resize dst here?
+    src.dec()._solve_impl(src.rhs(), dst);
+  }
+};
+
+} // end namepsace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVE_H

Eigen/src/Core/SolveTriangular.h

@@ -68,7 +68,7 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,1>
     if(!useRhsDirectly)
       MappedRhs(actualRhs,rhs.size()) = rhs;
 
-    triangular_solve_vector<LhsScalar, RhsScalar, typename Lhs::Index, Side, Mode, LhsProductTraits::NeedToConjugate,
+    triangular_solve_vector<LhsScalar, RhsScalar, Index, Side, Mode, LhsProductTraits::NeedToConjugate,
                             (int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor>
       ::run(actualLhs.cols(), actualLhs.data(), actualLhs.outerStride(), actualRhs);
 
@@ -82,7 +82,6 @@ template<typename Lhs, typename Rhs, int Side, int Mode>
 struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
 {
   typedef typename Rhs::Scalar Scalar;
-  typedef typename Rhs::Index Index;
   typedef blas_traits<Lhs> LhsProductTraits;
   typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType;
 
@@ -96,7 +95,7 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
     typedef internal::gemm_blocking_space<(Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
               Rhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxRowsAtCompileTime,4> BlockingType;
 
-    BlockingType blocking(rhs.rows(), rhs.cols(), size);
+    BlockingType blocking(rhs.rows(), rhs.cols(), size, 1, false);
 
     triangular_solve_matrix<Scalar,Index,Side,Mode,LhsProductTraits::NeedToConjugate,(int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor,
                                (Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor>
@@ -171,10 +170,10 @@ struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
   */
 template<typename MatrixType, unsigned int Mode>
 template<int Side, typename OtherDerived>
-void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
+void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
 {
   OtherDerived& other = _other.const_cast_derived();
-  eigen_assert( cols() == rows() && ((Side==OnTheLeft && cols() == other.rows()) || (Side==OnTheRight && cols() == other.cols())) );
+  eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) );
   eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
 
   enum { copy = internal::traits<OtherDerived>::Flags & RowMajorBit  && OtherDerived::IsVectorAtCompileTime };
@@ -183,7 +182,7 @@ void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived
   OtherCopy otherCopy(other);
 
   internal::triangular_solver_selector<MatrixType, typename internal::remove_reference<OtherCopy>::type,
-    Side, Mode>::run(nestedExpression(), otherCopy);
+    Side, Mode>::run(derived().nestedExpression(), otherCopy);
 
   if (copy)
     other = otherCopy;
@@ -199,8 +198,8 @@ void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived
   * 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 MatrixBase_marked.cpp
-  * Output: \verbinclude MatrixBase_marked.out
+  * 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.
@@ -213,9 +212,9 @@ void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived
 template<typename Derived, unsigned int Mode>
 template<int Side, typename Other>
 const internal::triangular_solve_retval<Side,TriangularView<Derived,Mode>,Other>
-TriangularView<Derived,Mode>::solve(const MatrixBase<Other>& other) const
+TriangularViewImpl<Derived,Mode,Dense>::solve(const MatrixBase<Other>& other) const
 {
-  return internal::triangular_solve_retval<Side,TriangularView,Other>(*this, other.derived());
+  return internal::triangular_solve_retval<Side,TriangularViewType,Other>(derived(), other.derived());
 }
 
 namespace internal {
@@ -232,7 +231,6 @@ template<int Side, typename TriangularType, typename Rhs> struct triangular_solv
 {
   typedef typename remove_all<typename Rhs::Nested>::type RhsNestedCleaned;
   typedef ReturnByValue<triangular_solve_retval> Base;
-  typedef typename Base::Index Index;
 
   triangular_solve_retval(const TriangularType& tri, const Rhs& rhs)
     : m_triangularMatrix(tri), m_rhs(rhs)

Eigen/src/Core/StableNorm.h

@@ -17,7 +17,6 @@ namespace internal {
 template<typename ExpressionType, typename Scalar>
 inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
 {
-  using std::max;
   Scalar maxCoeff = bl.cwiseAbs().maxCoeff();
   
   if(maxCoeff>scale)
@@ -56,10 +55,7 @@ inline typename NumTraits<typename traits<Derived>::Scalar>::Real
 blueNorm_impl(const EigenBase<Derived>& _vec)
 {
   typedef typename Derived::RealScalar RealScalar;  
-  typedef typename Derived::Index Index;
   using std::pow;
-  EIGEN_USING_STD_MATH(min);
-  EIGEN_USING_STD_MATH(max);
   using std::sqrt;
   using std::abs;
   const Derived& vec(_vec.derived());
@@ -136,8 +132,8 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
   }
   else
     return sqrt(amed);
-  asml = (min)(abig, amed);
-  abig = (max)(abig, amed);
+  asml = numext::mini(abig, amed);
+  abig = numext::maxi(abig, amed);
   if(asml <= abig*relerr)
     return abig;
   else
@@ -160,21 +156,33 @@ template<typename Derived>
 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
 MatrixBase<Derived>::stableNorm() const
 {
-  EIGEN_USING_STD_MATH(min);
   using std::sqrt;
+  using std::abs;
   const Index blockSize = 4096;
   RealScalar scale(0);
   RealScalar invScale(1);
   RealScalar ssq(0); // sum of square
+  
+  typedef typename internal::nested_eval<Derived,2>::type DerivedCopy;
+  typedef typename internal::remove_all<DerivedCopy>::type DerivedCopyClean;
+  DerivedCopy copy(derived());
+  
   enum {
-    Alignment = (int(Flags)&DirectAccessBit) || (int(Flags)&AlignedBit) ? 1 : 0
+    CanAlign = (int(Flags)&DirectAccessBit) || (int(internal::evaluator<DerivedCopyClean>::Alignment)>0) // FIXME
   };
+  typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<DerivedCopyClean>::Alignment>,
+                                                   typename DerivedCopyClean
+                                                   ::ConstSegmentReturnType>::type SegmentWrapper;
   Index n = size();
-  Index bi = internal::first_aligned(derived());
+  
+  if(n==1)
+    return abs(this->coeff(0));
+  
+  Index bi = internal::first_default_aligned(copy);
   if (bi>0)
-    internal::stable_norm_kernel(this->head(bi), ssq, scale, invScale);
+    internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale);
   for (; bi<n; bi+=blockSize)
-    internal::stable_norm_kernel(this->segment(bi,(min)(blockSize, n - bi)).template forceAlignedAccessIf<Alignment>(), ssq, scale, invScale);
+    internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale);
   return scale * sqrt(ssq);
 }
 

Eigen/src/Core/Stride.h

@@ -44,7 +44,7 @@ template<int _OuterStrideAtCompileTime, int _InnerStrideAtCompileTime>
 class Stride
 {
   public:
-    typedef DenseIndex Index;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
     enum {
       InnerStrideAtCompileTime = _InnerStrideAtCompileTime,
       OuterStrideAtCompileTime = _OuterStrideAtCompileTime
@@ -86,26 +86,24 @@ class Stride
 
 /** \brief Convenience specialization of Stride to specify only an inner stride
   * See class Map for some examples */
-template<int Value = Dynamic>
+template<int Value>
 class InnerStride : public Stride<0, Value>
 {
     typedef Stride<0, Value> Base;
   public:
-    typedef DenseIndex Index;
     EIGEN_DEVICE_FUNC InnerStride() : Base() {}
-    EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {}
+    EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} // FIXME making this explicit could break valid code
 };
 
 /** \brief Convenience specialization of Stride to specify only an outer stride
   * See class Map for some examples */
-template<int Value = Dynamic>
+template<int Value>
 class OuterStride : public Stride<Value, 0>
 {
     typedef Stride<Value, 0> Base;
   public:
-    typedef DenseIndex Index;
     EIGEN_DEVICE_FUNC OuterStride() : Base() {}
-    EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {}
+    EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} // FIXME making this explicit could break valid code
 };
 
 } // end namespace Eigen

Eigen/src/Core/Swap.h

@@ -12,129 +12,56 @@
 
 namespace Eigen { 
 
-/** \class SwapWrapper
-  * \ingroup Core_Module
-  *
-  * \internal
-  *
-  * \brief Internal helper class for swapping two expressions
-  */
 namespace internal {
-template<typename ExpressionType>
-struct traits<SwapWrapper<ExpressionType> > : traits<ExpressionType> {};
-}
 
-template<typename ExpressionType> class SwapWrapper
-  : public internal::dense_xpr_base<SwapWrapper<ExpressionType> >::type
+// Overload default assignPacket behavior for swapping them
+template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT>
+class generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, Specialized>
+ : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn>
 {
-  public:
-
-    typedef typename internal::dense_xpr_base<SwapWrapper>::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(SwapWrapper)
-    typedef typename internal::packet_traits<Scalar>::type Packet;
-
-    EIGEN_DEVICE_FUNC
-    inline SwapWrapper(ExpressionType& xpr) : m_expression(xpr) {}
-
-    EIGEN_DEVICE_FUNC
-    inline Index rows() const { return m_expression.rows(); }
-    EIGEN_DEVICE_FUNC
-    inline Index cols() const { return m_expression.cols(); }
-    EIGEN_DEVICE_FUNC
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    EIGEN_DEVICE_FUNC
-    inline Index innerStride() const { return m_expression.innerStride(); }
-    
-    typedef typename internal::conditional<
-                       internal::is_lvalue<ExpressionType>::value,
-                       Scalar,
-                       const Scalar
-                     >::type ScalarWithConstIfNotLvalue;
-                     
-    EIGEN_DEVICE_FUNC
-    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
-    EIGEN_DEVICE_FUNC
-    inline const Scalar* data() const { return m_expression.data(); }
-
-    EIGEN_DEVICE_FUNC
-    inline Scalar& coeffRef(Index rowId, Index colId)
-    {
-      return m_expression.const_cast_derived().coeffRef(rowId, colId);
-    }
-
-    EIGEN_DEVICE_FUNC
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_expression.const_cast_derived().coeffRef(index);
-    }
-
-    EIGEN_DEVICE_FUNC
-    inline Scalar& coeffRef(Index rowId, Index colId) const
-    {
-      return m_expression.coeffRef(rowId, colId);
-    }
-
-    EIGEN_DEVICE_FUNC
-    inline Scalar& coeffRef(Index index) const
-    {
-      return m_expression.coeffRef(index);
-    }
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    void copyCoeff(Index rowId, Index colId, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(rowId >= 0 && rowId < rows()
-                         && colId >= 0 && colId < cols());
-      Scalar tmp = m_expression.coeff(rowId, colId);
-      m_expression.coeffRef(rowId, colId) = _other.coeff(rowId, colId);
-      _other.coeffRef(rowId, colId) = tmp;
-    }
-
-    template<typename OtherDerived>
-    EIGEN_DEVICE_FUNC
-    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_expression.size());
-      Scalar tmp = m_expression.coeff(index);
-      m_expression.coeffRef(index) = _other.coeff(index);
-      _other.coeffRef(index) = tmp;
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index rowId, Index colId, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(rowId >= 0 && rowId < rows()
-                        && colId >= 0 && colId < cols());
-      Packet tmp = m_expression.template packet<StoreMode>(rowId, colId);
-      m_expression.template writePacket<StoreMode>(rowId, colId,
-        _other.template packet<LoadMode>(rowId, colId)
-      );
-      _other.template writePacket<LoadMode>(rowId, colId, tmp);
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_expression.size());
-      Packet tmp = m_expression.template packet<StoreMode>(index);
-      m_expression.template writePacket<StoreMode>(index,
-        _other.template packet<LoadMode>(index)
-      );
-      _other.template writePacket<LoadMode>(index, tmp);
-    }
-
-    EIGEN_DEVICE_FUNC
-    ExpressionType& expression() const { return m_expression; }
-
-  protected:
-    ExpressionType& m_expression;
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn> Base;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+  
+public:
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::DstXprType DstXprType;
+  typedef swap_assign_op<Scalar> Functor;
+  
+  EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index row, Index col)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(row,col);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(row,col, m_dst.template packet<StoreMode,PacketType>(row,col));
+    m_dst.template writePacket<StoreMode>(row,col,tmp);
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index index)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(index);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(index, m_dst.template packet<StoreMode,PacketType>(index));
+    m_dst.template writePacket<StoreMode>(index,tmp);
+  }
+  
+  // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael)
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacketByOuterInner(Index outer, Index inner)
+  {
+    Index row = Base::rowIndexByOuterInner(outer, inner); 
+    Index col = Base::colIndexByOuterInner(outer, inner);
+    assignPacket<StoreMode,LoadMode,PacketType>(row, col);
+  }
 };
 
+} // namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_SWAP_H

Eigen/src/Core/Transpose.h

@@ -2,7 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009-2014 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
@@ -29,13 +29,10 @@ namespace Eigen {
 
 namespace internal {
 template<typename MatrixType>
-struct traits<Transpose<MatrixType> > : traits<MatrixType>
+struct traits<Transpose<MatrixType> > : public traits<MatrixType>
 {
-  typedef typename MatrixType::Scalar Scalar;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedPlain;
-  typedef typename traits<MatrixType>::StorageKind StorageKind;
-  typedef typename traits<MatrixType>::XprKind XprKind;
   enum {
     RowsAtCompileTime = MatrixType::ColsAtCompileTime,
     ColsAtCompileTime = MatrixType::RowsAtCompileTime,
@@ -45,7 +42,6 @@ struct traits<Transpose<MatrixType> > : traits<MatrixType>
     Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit),
     Flags1 = Flags0 | FlagsLvalueBit,
     Flags = Flags1 ^ RowMajorBit,
-    CoeffReadCost = MatrixTypeNestedPlain::CoeffReadCost,
     InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret,
     OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret
   };
@@ -61,9 +57,10 @@ template<typename MatrixType> class Transpose
 
     typedef typename TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
     EIGEN_DEVICE_FUNC
-    inline Transpose(MatrixType& a_matrix) : m_matrix(a_matrix) {}
+    explicit inline Transpose(MatrixType& matrix) : m_matrix(matrix) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose)
 
@@ -100,12 +97,22 @@ struct TransposeImpl_base<MatrixType, false>
 
 } // end namespace internal
 
+// Generic API dispatcher
+template<typename XprType, typename StorageKind>
+class TransposeImpl
+  : public internal::generic_xpr_base<Transpose<XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<Transpose<XprType> >::type Base;
+};
+
 template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
   : public internal::TransposeImpl_base<MatrixType>::type
 {
   public:
 
     typedef typename internal::TransposeImpl_base<MatrixType>::type Base;
+    using Base::coeffRef;
     EIGEN_DENSE_PUBLIC_INTERFACE(Transpose<MatrixType>)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TransposeImpl)
 
@@ -118,23 +125,10 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
-    inline const Scalar* data() const { return derived().nestedExpression().data(); }
-
-    EIGEN_DEVICE_FUNC
-    inline ScalarWithConstIfNotLvalue& coeffRef(Index rowId, Index colId)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return derived().nestedExpression().const_cast_derived().coeffRef(colId, rowId);
-    }
-
-    EIGEN_DEVICE_FUNC
-    inline ScalarWithConstIfNotLvalue& coeffRef(Index index)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return derived().nestedExpression().const_cast_derived().coeffRef(index);
-    }
+    EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return derived().nestedExpression().data(); }
 
+    // FIXME: shall we keep the const version of coeffRef?
     EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
@@ -146,42 +140,6 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
     {
       return derived().nestedExpression().coeffRef(index);
     }
-
-    EIGEN_DEVICE_FUNC
-    inline CoeffReturnType coeff(Index rowId, Index colId) const
-    {
-      return derived().nestedExpression().coeff(colId, rowId);
-    }
-
-    EIGEN_DEVICE_FUNC
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return derived().nestedExpression().coeff(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().nestedExpression().template packet<LoadMode>(colId, rowId);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index rowId, Index colId, const PacketScalar& x)
-    {
-      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(colId, rowId, x);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return derived().nestedExpression().template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(index, x);
-    }
 };
 
 /** \returns an expression of the transpose of *this.
@@ -207,7 +165,7 @@ template<typename Derived>
 inline Transpose<Derived>
 DenseBase<Derived>::transpose()
 {
-  return derived();
+  return TransposeReturnType(derived());
 }
 
 /** This is the const version of transpose().
@@ -245,8 +203,7 @@ template<typename Derived>
 inline const typename MatrixBase<Derived>::AdjointReturnType
 MatrixBase<Derived>::adjoint() const
 {
-  return this->transpose(); // in the complex case, the .conjugate() is be implicit here
-                            // due to implicit conversion to return type
+  return AdjointReturnType(this->transpose());
 }
 
 /***************************************************************************
@@ -256,18 +213,38 @@ MatrixBase<Derived>::adjoint() const
 namespace internal {
 
 template<typename MatrixType,
-  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic>
+  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic,
+  bool MatchPacketSize =
+        (int(MatrixType::RowsAtCompileTime) == int(internal::packet_traits<typename MatrixType::Scalar>::size))
+    &&  (internal::evaluator<MatrixType>::Flags&PacketAccessBit) >
 struct inplace_transpose_selector;
 
 template<typename MatrixType>
-struct inplace_transpose_selector<MatrixType,true> { // square matrix
+struct inplace_transpose_selector<MatrixType,true,false> { // square matrix
   static void run(MatrixType& m) {
     m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
   }
 };
 
+// TODO: vectorized path is currently limited to LargestPacketSize x LargestPacketSize cases only.
 template<typename MatrixType>
-struct inplace_transpose_selector<MatrixType,false> { // non square matrix
+struct inplace_transpose_selector<MatrixType,true,true> { // PacketSize x PacketSize
+  static void run(MatrixType& m) {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename internal::packet_traits<typename MatrixType::Scalar>::type Packet;
+    const Index PacketSize = internal::packet_traits<Scalar>::size;
+    const Index Alignment = internal::evaluator<MatrixType>::Alignment;
+    PacketBlock<Packet> A;
+    for (Index i=0; i<PacketSize; ++i)
+      A.packet[i] = m.template packetByOuterInner<Alignment>(i,0);
+    internal::ptranspose(A);
+    for (Index i=0; i<PacketSize; ++i)
+      m.template writePacket<Alignment>(m.rowIndexByOuterInner(i,0), m.colIndexByOuterInner(i,0), A.packet[i]);
+  }
+};
+
+template<typename MatrixType,bool MatchPacketSize>
+struct inplace_transpose_selector<MatrixType,false,MatchPacketSize> { // non square matrix
   static void run(MatrixType& m) {
     if (m.rows()==m.cols())
       m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
@@ -340,14 +317,6 @@ inline void MatrixBase<Derived>::adjointInPlace()
 
 namespace internal {
 
-template<typename BinOp,typename NestedXpr,typename Rhs>
-struct blas_traits<SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> >
- : blas_traits<NestedXpr>
-{
-  typedef SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> XprType;
-  static inline const XprType extract(const XprType& x) { return x; }
-};
-
 template<bool DestIsTransposed, typename OtherDerived>
 struct check_transpose_aliasing_compile_time_selector
 {
@@ -413,15 +382,15 @@ struct checkTransposeAliasing_impl<Derived, OtherDerived, false>
     }
 };
 
+template<typename Dst, typename Src>
+void check_for_aliasing(const Dst &dst, const Src &src)
+{
+  internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src);
+}
+
 } // end namespace internal
 
-template<typename Derived>
-template<typename OtherDerived>
-void DenseBase<Derived>::checkTransposeAliasing(const OtherDerived& other) const
-{
-    internal::checkTransposeAliasing_impl<Derived, OtherDerived>::run(derived(), other);
-}
-#endif
+#endif // EIGEN_NO_DEBUG
 
 } // end namespace Eigen
 

Eigen/src/Core/Transpositions.h

@@ -41,10 +41,6 @@ namespace Eigen {
   * \sa class PermutationMatrix
   */
 
-namespace internal {
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed=false> struct transposition_matrix_product_retval;
-}
-
 template<typename Derived>
 class TranspositionsBase
 {
@@ -53,8 +49,8 @@ class TranspositionsBase
   public:
 
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar StorageIndexType;
-    typedef typename IndicesType::Index  Index;
+    typedef typename IndicesType::Scalar StorageIndex;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
 
     Derived& derived() { return *static_cast<Derived*>(this); }
     const Derived& derived() const { return *static_cast<const Derived*>(this); }
@@ -66,7 +62,7 @@ class TranspositionsBase
       indices() = other.indices();
       return derived();
     }
-
+    
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
@@ -79,20 +75,24 @@ class TranspositionsBase
     #endif
 
     /** \returns the number of transpositions */
-    inline Index size() const { return indices().size(); }
+    Index size() const { return indices().size(); }
+    /** \returns the number of rows of the equivalent permutation matrix */
+    Index rows() const { return indices().size(); }
+    /** \returns the number of columns of the equivalent permutation matrix */
+    Index cols() const { return indices().size(); }
 
     /** Direct access to the underlying index vector */
-    inline const StorageIndexType& coeff(Index i) const { return indices().coeff(i); }
+    inline const StorageIndex& coeff(Index i) const { return indices().coeff(i); }
     /** Direct access to the underlying index vector */
-    inline StorageIndexType& coeffRef(Index i) { return indices().coeffRef(i); }
+    inline StorageIndex& coeffRef(Index i) { return indices().coeffRef(i); }
     /** Direct access to the underlying index vector */
-    inline const StorageIndexType& operator()(Index i) const { return indices()(i); }
+    inline const StorageIndex& operator()(Index i) const { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline StorageIndexType& operator()(Index i) { return indices()(i); }
+    inline StorageIndex& operator()(Index i) { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline const StorageIndexType& operator[](Index i) const { return indices()(i); }
+    inline const StorageIndex& operator[](Index i) const { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline StorageIndexType& operator[](Index i) { return indices()(i); }
+    inline StorageIndex& operator[](Index i) { return indices()(i); }
 
     /** const version of indices(). */
     const IndicesType& indices() const { return derived().indices(); }
@@ -108,7 +108,7 @@ class TranspositionsBase
     /** Sets \c *this to represents an identity transformation */
     void setIdentity()
     {
-      for(StorageIndexType i = 0; i < indices().size(); ++i)
+      for(StorageIndex i = 0; i < indices().size(); ++i)
         coeffRef(i) = i;
     }
 
@@ -145,26 +145,24 @@ class TranspositionsBase
 };
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType>
-struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
-  typedef Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
-  typedef typename IndicesType::Index Index;
-  typedef _StorageIndexType StorageIndexType;
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType>
-class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
     typedef internal::traits<Transpositions> Traits;
   public:
 
     typedef TranspositionsBase<Transpositions> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar StorageIndexType;
-    typedef typename IndicesType::Index  Index;
-    
+    typedef typename IndicesType::Scalar StorageIndex;
 
     inline Transpositions() {}
 
@@ -181,7 +179,7 @@ class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTim
 
     /** Generic constructor from expression of the transposition indices. */
     template<typename Other>
-    explicit inline Transpositions(const MatrixBase<Other>& a_indices) : m_indices(a_indices)
+    explicit inline Transpositions(const MatrixBase<Other>& indices) : m_indices(indices)
     {}
 
     /** Copies the \a other transpositions into \c *this */
@@ -219,32 +217,32 @@ class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTim
 
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType, int _PacketAccess>
-struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndexType>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,_PacketAccess> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
-  typedef Map<const Matrix<_StorageIndexType,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
-  typedef typename IndicesType::Index Index;
-  typedef _StorageIndexType StorageIndexType;
+  typedef Map<const Matrix<_StorageIndex,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndexType, int PacketAccess>
-class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndexType>,PacketAccess>
- : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndexType>,PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int PacketAccess>
+class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess>
+ : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess> >
 {
     typedef internal::traits<Map> Traits;
   public:
 
     typedef TranspositionsBase<Map> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar StorageIndexType;
-    typedef typename IndicesType::Index  Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
-    inline Map(const StorageIndexType* indicesPtr)
+    explicit inline Map(const StorageIndex* indicesPtr)
       : m_indices(indicesPtr)
     {}
 
-    inline Map(const StorageIndexType* indicesPtr, Index size)
+    inline Map(const StorageIndex* indicesPtr, Index size)
       : m_indices(indicesPtr,size)
     {}
 
@@ -280,10 +278,9 @@ class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndexTyp
 namespace internal {
 template<typename _IndicesType>
 struct traits<TranspositionsWrapper<_IndicesType> >
+ : traits<PermutationWrapper<_IndicesType> >
 {
-  typedef typename _IndicesType::Scalar StorageIndexType;
-  typedef typename _IndicesType::Index Index;
-  typedef _IndicesType IndicesType;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
@@ -296,11 +293,10 @@ class TranspositionsWrapper
 
     typedef TranspositionsBase<TranspositionsWrapper> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar StorageIndexType;
-    typedef typename IndicesType::Index  Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
-    inline TranspositionsWrapper(IndicesType& a_indices)
-      : m_indices(a_indices)
+    explicit inline TranspositionsWrapper(IndicesType& indices)
+      : m_indices(indices)
     {}
 
     /** Copies the \a other transpositions into \c *this */
@@ -332,81 +328,43 @@ class TranspositionsWrapper
     const typename IndicesType::Nested m_indices;
 };
 
+
+
 /** \returns the \a matrix with the \a transpositions applied to the columns.
   */
-template<typename Derived, typename TranspositionsDerived>
-inline const internal::transposition_matrix_product_retval<TranspositionsDerived, Derived, OnTheRight>
-operator*(const MatrixBase<Derived>& matrix,
-          const TranspositionsBase<TranspositionsDerived> &transpositions)
+template<typename MatrixDerived, typename TranspositionsDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const TranspositionsBase<TranspositionsDerived>& transpositions)
 {
-  return internal::transposition_matrix_product_retval
-           <TranspositionsDerived, Derived, OnTheRight>
-           (transpositions.derived(), matrix.derived());
+  return Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+            (matrix.derived(), transpositions.derived());
 }
 
 /** \returns the \a matrix with the \a transpositions applied to the rows.
   */
-template<typename Derived, typename TranspositionDerived>
-inline const internal::transposition_matrix_product_retval
-               <TranspositionDerived, Derived, OnTheLeft>
-operator*(const TranspositionsBase<TranspositionDerived> &transpositions,
-          const MatrixBase<Derived>& matrix)
+template<typename TranspositionsDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+operator*(const TranspositionsBase<TranspositionsDerived> &transpositions,
+          const MatrixBase<MatrixDerived>& matrix)
 {
-  return internal::transposition_matrix_product_retval
-           <TranspositionDerived, Derived, OnTheLeft>
-           (transpositions.derived(), matrix.derived());
+  return Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+            (transpositions.derived(), matrix.derived());
 }
 
+// Template partial specialization for transposed/inverse transpositions
+
 namespace internal {
 
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
-struct traits<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
-{
-  typedef typename MatrixType::PlainObject ReturnType;
-};
-
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
-struct transposition_matrix_product_retval
- : public ReturnByValue<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
-{
-    typedef typename remove_all<typename MatrixType::Nested>::type MatrixTypeNestedCleaned;
-    typedef typename TranspositionType::Index Index;
-    typedef typename TranspositionType::StorageIndexType StorageIndexType;
-
-    transposition_matrix_product_retval(const TranspositionType& tr, const MatrixType& matrix)
-      : m_transpositions(tr), m_matrix(matrix)
-    {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-
-    template<typename Dest> inline void evalTo(Dest& dst) const
-    {
-      const Index size = m_transpositions.size();
-      StorageIndexType j = 0;
-
-      if(!(is_same<MatrixTypeNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_matrix)))
-        dst = m_matrix;
-
-      for(Index k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k)
-        if(Index(j=m_transpositions.coeff(k))!=k)
-        {
-          if(Side==OnTheLeft)
-            dst.row(k).swap(dst.row(j));
-          else if(Side==OnTheRight)
-            dst.col(k).swap(dst.col(j));
-        }
-    }
-
-  protected:
-    const TranspositionType& m_transpositions;
-    typename MatrixType::Nested m_matrix;
-};
+template<typename Derived>
+struct traits<Transpose<TranspositionsBase<Derived> > >
+ : traits<Derived>
+{};
 
 } // end namespace internal
 
-/* Template partial specialization for transposed/inverse transpositions */
-
 template<typename TranspositionsDerived>
 class Transpose<TranspositionsBase<TranspositionsDerived> >
 {
@@ -414,27 +372,31 @@ class Transpose<TranspositionsBase<TranspositionsDerived> >
     typedef typename TranspositionType::IndicesType IndicesType;
   public:
 
-    Transpose(const TranspositionType& t) : m_transpositions(t) {}
+    explicit Transpose(const TranspositionType& t) : m_transpositions(t) {}
 
-    inline int size() const { return m_transpositions.size(); }
+    Index size() const { return m_transpositions.size(); }
+    Index rows() const { return m_transpositions.size(); }
+    Index cols() const { return m_transpositions.size(); }
 
     /** \returns the \a matrix with the inverse transpositions applied to the columns.
       */
-    template<typename Derived> friend
-    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>
-    operator*(const MatrixBase<Derived>& matrix, const Transpose& trt)
+    template<typename OtherDerived> friend
+    const Product<OtherDerived, Transpose, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trt)
     {
-      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>(trt.m_transpositions, matrix.derived());
+      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trt.derived());
     }
 
     /** \returns the \a matrix with the inverse transpositions applied to the rows.
       */
-    template<typename Derived>
-    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>
-    operator*(const MatrixBase<Derived>& matrix) const
+    template<typename OtherDerived>
+    const Product<Transpose, OtherDerived, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix) const
     {
-      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>(m_transpositions, matrix.derived());
+      return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived());
     }
+    
+    const TranspositionType& nestedExpression() const { return m_transpositions; }
 
   protected:
     const TranspositionType& m_transpositions;

Eigen/src/Core/VectorwiseOp.h

@@ -41,32 +41,22 @@ 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 nested<MatrixType>::type MatrixTypeNested;
+  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,
     MaxRowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime,
-    Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits,
-    Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0),
+    Flags = RowsAtCompileTime == 1 ? RowMajorBit : 0,
     TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime :  MatrixType::ColsAtCompileTime
   };
-  #if EIGEN_GNUC_AT_LEAST(3,4)
-  typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
-  #else
-  typedef typename MemberOp::template Cost<InputScalar,TraversalSize> CostOpType;
-  #endif
-  enum {
-    CoeffReadCost = TraversalSize==Dynamic ? Dynamic
-                  : TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value)
-  };
 };
 }
 
 template< typename MatrixType, typename MemberOp, int Direction>
-class PartialReduxExpr : internal::no_assignment_operator,
-  public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type
+class PartialReduxExpr : public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type,
+                         internal::no_assignment_operator
 {
   public:
 
@@ -75,13 +65,16 @@ class PartialReduxExpr : internal::no_assignment_operator,
     typedef typename internal::traits<PartialReduxExpr>::MatrixTypeNested MatrixTypeNested;
     typedef typename internal::traits<PartialReduxExpr>::_MatrixTypeNested _MatrixTypeNested;
 
-    PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
+    EIGEN_DEVICE_FUNC
+    explicit PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
       : m_matrix(mat), m_functor(func) {}
 
+    EIGEN_DEVICE_FUNC
     Index rows() const { return (Direction==Vertical   ? 1 : m_matrix.rows()); }
+    EIGEN_DEVICE_FUNC
     Index cols() const { return (Direction==Horizontal ? 1 : m_matrix.cols()); }
 
-    EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
     {
       if (Direction==Vertical)
         return m_functor(m_matrix.col(j));
@@ -89,7 +82,7 @@ class PartialReduxExpr : internal::no_assignment_operator,
         return m_functor(m_matrix.row(i));
     }
 
-    const Scalar coeff(Index index) const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
     {
       if (Direction==Vertical)
         return m_functor(m_matrix.col(index));
@@ -110,7 +103,8 @@ class PartialReduxExpr : internal::no_assignment_operator,
     template<typename Scalar, int Size> struct Cost                     \
     { enum { value = COST }; };                                         \
     template<typename XprType>                                          \
-    EIGEN_STRONG_INLINE ResultType operator()(const XprType& mat) const \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                               \
+    ResultType operator()(const XprType& mat) const                     \
     { return mat.MEMBER(); } \
   }
 
@@ -134,13 +128,13 @@ EIGEN_MEMBER_FUNCTOR(prod, (Size-1)*NumTraits<Scalar>::MulCost);
 template <typename BinaryOp, typename Scalar>
 struct member_redux {
   typedef typename result_of<
-                     BinaryOp(Scalar)
+                     BinaryOp(Scalar,Scalar)
                    >::type  result_type;
   template<typename _Scalar, int Size> struct Cost
   { enum { value = (Size-1) * functor_traits<BinaryOp>::Cost }; };
-  member_redux(const BinaryOp func) : m_functor(func) {}
+  EIGEN_DEVICE_FUNC explicit member_redux(const BinaryOp func) : m_functor(func) {}
   template<typename Derived>
-  inline result_type operator()(const DenseBase<Derived>& mat) const
+  EIGEN_DEVICE_FUNC inline result_type operator()(const DenseBase<Derived>& mat) const
   { return mat.redux(m_functor); }
   const BinaryOp m_functor;
 };
@@ -169,16 +163,15 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     typedef typename ExpressionType::Scalar Scalar;
     typedef typename ExpressionType::RealScalar RealScalar;
-    typedef typename ExpressionType::Index Index;
-    typedef typename internal::conditional<internal::must_nest_by_value<ExpressionType>::ret,
-        ExpressionType, ExpressionType&>::type ExpressionTypeNested;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type ExpressionTypeNested;
     typedef typename internal::remove_all<ExpressionTypeNested>::type ExpressionTypeNestedCleaned;
 
     template<template<typename _Scalar> class Functor,
-                      typename Scalar=typename internal::traits<ExpressionType>::Scalar> struct ReturnType
+                      typename Scalar_=Scalar> struct ReturnType
     {
       typedef PartialReduxExpr<ExpressionType,
-                               Functor<Scalar>,
+                               Functor<Scalar_>,
                                Direction
                               > Type;
     };
@@ -186,23 +179,24 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     template<typename BinaryOp> struct ReduxReturnType
     {
       typedef PartialReduxExpr<ExpressionType,
-                               internal::member_redux<BinaryOp,typename internal::traits<ExpressionType>::Scalar>,
+                               internal::member_redux<BinaryOp,Scalar>,
                                Direction
                               > Type;
     };
 
     enum {
-      IsVertical   = (Direction==Vertical) ? 1 : 0,
-      IsHorizontal = (Direction==Horizontal) ? 1 : 0
+      isVertical   = (Direction==Vertical) ? 1 : 0,
+      isHorizontal = (Direction==Horizontal) ? 1 : 0
     };
 
   protected:
 
     /** \internal
       * \returns the i-th subvector according to the \c Direction */
-    typedef typename internal::conditional<Direction==Vertical,
+    typedef typename internal::conditional<isVertical,
                                typename ExpressionType::ColXpr,
                                typename ExpressionType::RowXpr>::type SubVector;
+    EIGEN_DEVICE_FUNC
     SubVector subVector(Index i)
     {
       return SubVector(m_matrix.derived(),i);
@@ -210,58 +204,62 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** \internal
       * \returns the number of subvectors in the direction \c Direction */
+    EIGEN_DEVICE_FUNC
     Index subVectors() const
-    { return Direction==Vertical?m_matrix.cols():m_matrix.rows(); }
+    { return isVertical?m_matrix.cols():m_matrix.rows(); }
 
     template<typename OtherDerived> struct ExtendedType {
       typedef Replicate<OtherDerived,
-                        Direction==Vertical   ? 1 : ExpressionType::RowsAtCompileTime,
-                        Direction==Horizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
+                        isVertical   ? 1 : ExpressionType::RowsAtCompileTime,
+                        isHorizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
     };
 
     /** \internal
       * Replicates a vector to match the size of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename ExtendedType<OtherDerived>::Type
     extendedTo(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxColsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxColsAtCompileTime==1),
                           YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxRowsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxRowsAtCompileTime==1),
                           YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
       return typename ExtendedType<OtherDerived>::Type
                       (other.derived(),
-                       Direction==Vertical   ? 1 : m_matrix.rows(),
-                       Direction==Horizontal ? 1 : m_matrix.cols());
+                       isVertical   ? 1 : m_matrix.rows(),
+                       isHorizontal ? 1 : m_matrix.cols());
     }
     
     template<typename OtherDerived> struct OppositeExtendedType {
       typedef Replicate<OtherDerived,
-                        Direction==Horizontal ? 1 : ExpressionType::RowsAtCompileTime,
-                        Direction==Vertical   ? 1 : ExpressionType::ColsAtCompileTime> Type;
+                        isHorizontal ? 1 : ExpressionType::RowsAtCompileTime,
+                        isVertical   ? 1 : ExpressionType::ColsAtCompileTime> Type;
     };
 
     /** \internal
       * Replicates a vector in the opposite direction to match the size of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename OppositeExtendedType<OtherDerived>::Type
     extendedToOpposite(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxColsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxColsAtCompileTime==1),
                           YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxRowsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxRowsAtCompileTime==1),
                           YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
       return typename OppositeExtendedType<OtherDerived>::Type
                       (other.derived(),
-                       Direction==Horizontal  ? 1 : m_matrix.rows(),
-                       Direction==Vertical    ? 1 : m_matrix.cols());
+                       isHorizontal  ? 1 : m_matrix.rows(),
+                       isVertical    ? 1 : m_matrix.cols());
     }
 
   public:
-
-    inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
 
     /** \internal */
+    EIGEN_DEVICE_FUNC
     inline const ExpressionType& _expression() const { return m_matrix; }
 
     /** \returns a row or column vector expression of \c *this reduxed by \a func
@@ -272,9 +270,25 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * \sa class VectorwiseOp, DenseBase::colwise(), DenseBase::rowwise()
       */
     template<typename BinaryOp>
+    EIGEN_DEVICE_FUNC
     const typename ReduxReturnType<BinaryOp>::Type
     redux(const BinaryOp& func = BinaryOp()) const
-    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), func); }
+    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), internal::member_redux<BinaryOp,Scalar>(func)); }
+
+    typedef typename ReturnType<internal::member_minCoeff>::Type MinCoeffReturnType;
+    typedef typename ReturnType<internal::member_maxCoeff>::Type MaxCoeffReturnType;
+    typedef typename ReturnType<internal::member_squaredNorm,RealScalar>::Type SquaredNormReturnType;
+    typedef typename ReturnType<internal::member_norm,RealScalar>::Type NormReturnType;
+    typedef typename ReturnType<internal::member_blueNorm,RealScalar>::Type BlueNormReturnType;
+    typedef typename ReturnType<internal::member_stableNorm,RealScalar>::Type StableNormReturnType;
+    typedef typename ReturnType<internal::member_hypotNorm,RealScalar>::Type HypotNormReturnType;
+    typedef typename ReturnType<internal::member_sum>::Type SumReturnType;
+    typedef typename ReturnType<internal::member_mean>::Type MeanReturnType;
+    typedef typename ReturnType<internal::member_all>::Type AllReturnType;
+    typedef typename ReturnType<internal::member_any>::Type AnyReturnType;
+    typedef PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> CountReturnType;
+    typedef typename ReturnType<internal::member_prod>::Type ProdReturnType;
+    typedef Reverse<ExpressionType, Direction> ReverseReturnType;
 
     /** \returns a row (or column) vector expression of the smallest coefficient
       * of each column (or row) of the referenced expression.
@@ -285,8 +299,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_minCoeff.out
       *
       * \sa DenseBase::minCoeff() */
-    const typename ReturnType<internal::member_minCoeff>::Type minCoeff() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MinCoeffReturnType minCoeff() const
+    { return MinCoeffReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the largest coefficient
       * of each column (or row) of the referenced expression.
@@ -297,8 +312,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_maxCoeff.out
       *
       * \sa DenseBase::maxCoeff() */
-    const typename ReturnType<internal::member_maxCoeff>::Type maxCoeff() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MaxCoeffReturnType maxCoeff() const
+    { return MaxCoeffReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the squared norm
       * of each column (or row) of the referenced expression.
@@ -308,8 +324,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_squaredNorm.out
       *
       * \sa DenseBase::squaredNorm() */
-    const typename ReturnType<internal::member_squaredNorm,RealScalar>::Type squaredNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const SquaredNormReturnType squaredNorm() const
+    { return SquaredNormReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression.
@@ -319,8 +336,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_norm.out
       *
       * \sa DenseBase::norm() */
-    const typename ReturnType<internal::member_norm,RealScalar>::Type norm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const NormReturnType norm() const
+    { return NormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
@@ -329,8 +347,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::blueNorm() */
-    const typename ReturnType<internal::member_blueNorm,RealScalar>::Type blueNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const BlueNormReturnType blueNorm() const
+    { return BlueNormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
@@ -339,8 +358,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::stableNorm() */
-    const typename ReturnType<internal::member_stableNorm,RealScalar>::Type stableNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const StableNormReturnType stableNorm() const
+    { return StableNormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
@@ -349,8 +369,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::hypotNorm() */
-    const typename ReturnType<internal::member_hypotNorm,RealScalar>::Type hypotNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const HypotNormReturnType hypotNorm() const
+    { return HypotNormReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the sum
       * of each column (or row) of the referenced expression.
@@ -359,31 +380,35 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_sum.out
       *
       * \sa DenseBase::sum() */
-    const typename ReturnType<internal::member_sum>::Type sum() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const SumReturnType sum() const
+    { return SumReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the mean
     * of each column (or row) of the referenced expression.
     *
     * \sa DenseBase::mean() */
-    const typename ReturnType<internal::member_mean>::Type mean() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MeanReturnType mean() const
+    { return MeanReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * whether \b all coefficients of each respective column (or row) are \c true.
       * This expression can be assigned to a vector with entries of type \c bool.
       *
       * \sa DenseBase::all() */
-    const typename ReturnType<internal::member_all>::Type all() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const AllReturnType all() const
+    { return AllReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * whether \b at \b least one coefficient of each respective column (or row) is \c true.
       * This expression can be assigned to a vector with entries of type \c bool.
       *
       * \sa DenseBase::any() */
-    const typename ReturnType<internal::member_any>::Type any() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const AnyReturnType any() const
+    { return Any(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * the number of \c true coefficients of each respective column (or row).
@@ -394,8 +419,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_count.out
       *
       * \sa DenseBase::count() */
-    const PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> count() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const CountReturnType count() const
+    { return CountReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the product
       * of each column (or row) of the referenced expression.
@@ -404,8 +430,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_prod.out
       *
       * \sa DenseBase::prod() */
-    const typename ReturnType<internal::member_prod>::Type prod() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const ProdReturnType prod() const
+    { return ProdReturnType(_expression()); }
 
 
     /** \returns a matrix expression
@@ -415,10 +442,12 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude Vectorwise_reverse.out
       *
       * \sa DenseBase::reverse() */
-    const Reverse<ExpressionType, Direction> reverse() const
-    { return Reverse<ExpressionType, Direction>( _expression() ); }
+    EIGEN_DEVICE_FUNC
+    const ReverseReturnType reverse() const
+    { return ReverseReturnType( _expression() ); }
 
-    typedef Replicate<ExpressionType,Direction==Vertical?Dynamic:1,Direction==Horizontal?Dynamic:1> ReplicateReturnType;
+    typedef Replicate<ExpressionType,(isVertical?Dynamic:1),(isHorizontal?Dynamic:1)> ReplicateReturnType;
+    EIGEN_DEVICE_FUNC
     const ReplicateReturnType replicate(Index factor) const;
 
     /**
@@ -430,17 +459,20 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * \sa VectorwiseOp::replicate(Index), DenseBase::replicate(), class Replicate
       */
     // NOTE implemented here because of sunstudio's compilation errors
-    template<int Factor> const Replicate<ExpressionType,(IsVertical?Factor:1),(IsHorizontal?Factor:1)>
+    // isVertical*Factor+isHorizontal instead of (isVertical?Factor:1) to handle CUDA bug with ternary operator
+    template<int Factor> const Replicate<ExpressionType,isVertical*Factor+isHorizontal,isHorizontal*Factor+isVertical>
+    EIGEN_DEVICE_FUNC
     replicate(Index factor = Factor) const
     {
-      return Replicate<ExpressionType,Direction==Vertical?Factor:1,Direction==Horizontal?Factor:1>
-          (_expression(),Direction==Vertical?factor:1,Direction==Horizontal?factor:1);
+      return Replicate<ExpressionType,(isVertical?Factor:1),(isHorizontal?Factor:1)>
+          (_expression(),isVertical?factor:1,isHorizontal?factor:1);
     }
 
 /////////// Artithmetic operators ///////////
 
     /** Copies the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -451,6 +483,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Adds the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator+=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -460,6 +493,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Substracts the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator-=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -469,6 +503,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Multiples each subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator*=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -480,6 +515,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Divides each subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator/=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -490,7 +526,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     }
 
     /** Returns the expression of the sum of the vector \a other to each subvector of \c *this */
-    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_sum_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
@@ -503,6 +539,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Returns the expression of the difference between each subvector of \c *this and the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_difference_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
@@ -515,10 +552,11 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Returns the expression where each subvector is the product of the vector \a other
       * by the corresponding subvector of \c *this */
-    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_product_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
+    EIGEN_DEVICE_FUNC
     operator*(const DenseBase<OtherDerived>& other) const
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -530,6 +568,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     /** Returns the expression where each subvector is the quotient of the corresponding
       * subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
@@ -545,6 +584,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * The referenced matrix is \b not modified.
       * \sa MatrixBase::normalized(), normalize()
       */
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename OppositeExtendedType<typename ReturnType<internal::member_norm,RealScalar>::Type>::Type>
@@ -554,16 +594,20 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     /** Normalize in-place each row or columns of the referenced matrix.
       * \sa MatrixBase::normalize(), normalized()
       */
-    void normalize() {
+    EIGEN_DEVICE_FUNC void normalize() {
       m_matrix = this->normalized();
     }
 
+    EIGEN_DEVICE_FUNC inline void reverseInPlace();
+
 /////////// Geometry module ///////////
 
-    Homogeneous<ExpressionType,Direction> homogeneous() const;
+    typedef Homogeneous<ExpressionType,Direction> HomogeneousReturnType;
+    HomogeneousReturnType homogeneous() const;
 
     typedef typename ExpressionType::PlainObject CrossReturnType;
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     const CrossReturnType cross(const MatrixBase<OtherDerived>& other) const;
 
     enum {
@@ -594,19 +638,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     ExpressionTypeNested m_matrix;
 };
 
-/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
-  *
-  * Example: \include MatrixBase_colwise.cpp
-  * Output: \verbinclude MatrixBase_colwise.out
-  *
-  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
-  */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstColwiseReturnType
-DenseBase<Derived>::colwise() const
-{
-  return derived();
-}
+//const colwise moved to DenseBase.h due to CUDA compiler bug
+
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
@@ -616,22 +649,11 @@ template<typename Derived>
 inline typename DenseBase<Derived>::ColwiseReturnType
 DenseBase<Derived>::colwise()
 {
-  return derived();
+  return ColwiseReturnType(derived());
 }
 
-/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
-  *
-  * Example: \include MatrixBase_rowwise.cpp
-  * Output: \verbinclude MatrixBase_rowwise.out
-  *
-  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
-  */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstRowwiseReturnType
-DenseBase<Derived>::rowwise() const
-{
-  return derived();
-}
+//const rowwise moved to DenseBase.h due to CUDA compiler bug
+
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
@@ -641,7 +663,7 @@ template<typename Derived>
 inline typename DenseBase<Derived>::RowwiseReturnType
 DenseBase<Derived>::rowwise()
 {
-  return derived();
+  return RowwiseReturnType(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/Visitor.h

@@ -22,6 +22,7 @@ struct visitor_impl
     row = (UnrollCount-1) % Derived::RowsAtCompileTime
   };
 
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived &mat, Visitor& visitor)
   {
     visitor_impl<Visitor, Derived, UnrollCount-1>::run(mat, visitor);
@@ -32,6 +33,7 @@ struct visitor_impl
 template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, 1>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived &mat, Visitor& visitor)
   {
     return visitor.init(mat.coeff(0, 0), 0, 0);
@@ -41,7 +43,7 @@ struct visitor_impl<Visitor, Derived, 1>
 template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, Dynamic>
 {
-  typedef typename Derived::Index Index;
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived& mat, Visitor& visitor)
   {
     visitor.init(mat.coeff(0,0), 0, 0);
@@ -53,6 +55,33 @@ struct visitor_impl<Visitor, Derived, Dynamic>
   }
 };
 
+// evaluator adaptor
+template<typename XprType>
+class visitor_evaluator
+{
+public:
+  EIGEN_DEVICE_FUNC
+  explicit visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  
+  enum {
+    RowsAtCompileTime = XprType::RowsAtCompileTime,
+    CoeffReadCost = internal::evaluator<XprType>::CoeffReadCost
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
 } // end namespace internal
 
 /** Applies the visitor \a visitor to the whole coefficients of the matrix or vector.
@@ -74,16 +103,20 @@ struct visitor_impl<Visitor, Derived, Dynamic>
   */
 template<typename Derived>
 template<typename Visitor>
+EIGEN_DEVICE_FUNC
 void DenseBase<Derived>::visit(Visitor& visitor) const
 {
-  enum { unroll = SizeAtCompileTime != Dynamic
-                   && CoeffReadCost != Dynamic
-                   && (SizeAtCompileTime == 1 || internal::functor_traits<Visitor>::Cost != Dynamic)
-                   && SizeAtCompileTime * CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost
-                      <= EIGEN_UNROLLING_LIMIT };
-  return internal::visitor_impl<Visitor, Derived,
+  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(derived(), visitor);
+    >::run(thisEval, visitor);
 }
 
 namespace internal {
@@ -94,10 +127,10 @@ namespace internal {
 template <typename Derived>
 struct coeff_visitor
 {
-  typedef typename Derived::Index Index;
   typedef typename Derived::Scalar Scalar;
   Index row, col;
   Scalar res;
+  EIGEN_DEVICE_FUNC
   inline void init(const Scalar& value, Index i, Index j)
   {
     res = value;
@@ -114,8 +147,8 @@ struct coeff_visitor
 template <typename Derived>
 struct min_coeff_visitor : coeff_visitor<Derived>
 {
-  typedef typename Derived::Index Index;
   typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC
   void operator() (const Scalar& value, Index i, Index j)
   {
     if(value < this->res)
@@ -142,8 +175,8 @@ struct functor_traits<min_coeff_visitor<Scalar> > {
 template <typename Derived>
 struct max_coeff_visitor : coeff_visitor<Derived>
 {
-  typedef typename Derived::Index Index;
-  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Scalar Scalar; 
+  EIGEN_DEVICE_FUNC
   void operator() (const Scalar& value, Index i, Index j)
   {
     if(value > this->res)
@@ -171,6 +204,7 @@ struct functor_traits<max_coeff_visitor<Scalar> > {
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
 {
@@ -188,6 +222,7 @@ DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff(IndexType* index) const
 {
@@ -205,6 +240,7 @@ DenseBase<Derived>::minCoeff(IndexType* index) const
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const
 {
@@ -222,6 +258,7 @@ DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff(IndexType* index) const
 {

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

@@ -45,7 +45,7 @@ template<> struct packet_traits<std::complex<float> >  : default_packet_traits
   };
 };
 
-template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4}; typedef Packet2cf half; };
+template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4, alignment=Aligned32}; typedef Packet2cf half; };
 
 template<> EIGEN_STRONG_INLINE Packet4cf padd<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_add_ps(a.v,b.v)); }
 template<> EIGEN_STRONG_INLINE Packet4cf psub<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_sub_ps(a.v,b.v)); }
@@ -92,7 +92,7 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploaddup<Packet4cf>(const std::complex<
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); }
 
-template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather<std::complex<float>, Packet4cf>(const std::complex<float>* from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather<std::complex<float>, Packet4cf>(const std::complex<float>* from, Index stride)
 {
   return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]),
                                  std::imag(from[2*stride]), std::real(from[2*stride]),
@@ -100,7 +100,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather<std::complex<float>, Packe
                                  std::imag(from[0*stride]), std::real(from[0*stride])));
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet4cf>(std::complex<float>* to, const Packet4cf& from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet4cf>(std::complex<float>* to, const Packet4cf& from, Index stride)
 {
   __m128 low = _mm256_extractf128_ps(from.v, 0);
   to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)),
@@ -267,7 +267,7 @@ template<> struct packet_traits<std::complex<double> >  : default_packet_traits
   };
 };
 
-template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2}; typedef Packet1cd half; };
+template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2, alignment=Aligned32}; typedef Packet1cd half; };
 
 template<> EIGEN_STRONG_INLINE Packet2cd padd<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_add_pd(a.v,b.v)); }
 template<> EIGEN_STRONG_INLINE Packet2cd psub<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_sub_pd(a.v,b.v)); }
@@ -310,13 +310,13 @@ template<> EIGEN_STRONG_INLINE Packet2cd ploaddup<Packet2cd>(const std::complex<
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
 
-template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather<std::complex<double>, Packet2cd>(const std::complex<double>* from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather<std::complex<double>, Packet2cd>(const std::complex<double>* from, Index stride)
 {
   return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]),
 				 std::imag(from[0*stride]), std::real(from[0*stride])));
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet2cd>(std::complex<double>* to, const Packet2cd& from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet2cd>(std::complex<double>* to, const Packet2cd& from, Index stride)
 {
   __m128d low = _mm256_extractf128_pd(from.v, 0);
   to[stride*0] = std::complex<double>(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1)));

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

@@ -0,0 +1,441 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@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_MATH_FUNCTIONS_AVX_H
+#define EIGEN_MATH_FUNCTIONS_AVX_H
+
+// For some reason, this function didn't make it into the avxintirn.h
+// used by the compiler, so we'll just wrap it.
+#define _mm256_setr_m128(lo, hi) \
+  _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1)
+
+/* The sin, cos, exp, and log functions of this file are loosely derived from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+namespace Eigen {
+
+namespace internal {
+
+// Sine function
+// Computes sin(x) by wrapping x to the interval [-Pi/4,3*Pi/4] and
+// evaluating interpolants in [-Pi/4,Pi/4] or [Pi/4,3*Pi/4]. The interpolants
+// are (anti-)symmetric and thus have only odd/even coefficients
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+psin<Packet8f>(const Packet8f& _x) {
+  Packet8f x = _x;
+
+  // Some useful values.
+  _EIGEN_DECLARE_CONST_Packet8i(one, 1);
+  _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
+  _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);
+
+  // 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);
+  Packet8f shift = _mm256_floor_ps(padd(z, p8f_one_over_four));
+  x = pmadd(shift, p8f_neg_pi_first, x);
+  x = pmadd(shift, p8f_neg_pi_second, x);
+  x = pmadd(shift, p8f_neg_pi_third, x);
+  z = pmul(x, p8f_four_over_pi);
+
+  // Make a mask for the entries that need flipping, i.e. wherever the shift
+  // is odd.
+  Packet8i shift_ints = _mm256_cvtps_epi32(shift);
+  Packet8i shift_isodd =
+      (__m256i)_mm256_and_ps((__m256)shift_ints, (__m256)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);
+  __m128i hi =
+      _mm_slli_epi32(_mm256_extractf128_si256((__m256i)shift_isodd, 1), 31);
+  Packet8i sign_flip_mask = _mm256_setr_m128(lo, hi);
+#endif
+
+  // Create a mask for which interpolant to use, i.e. if z > 1, then the mask
+  // is set to ones for that entry.
+  Packet8f ival_mask = _mm256_cmp_ps(z, p8f_one, _CMP_GT_OQ);
+
+  // 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);
+  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);
+  right = pmadd(right, z_minus_two2, p8f_coeff_right_2);
+  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);
+  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);
+  left = pmadd(left, z2, p8f_coeff_left_1);
+  left = pmul(left, z);
+
+  // Assemble the results, i.e. select the left and right polynomials.
+  left = _mm256_andnot_ps(ival_mask, left);
+  right = _mm256_and_ps(ival_mask, right);
+  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);
+  return res;
+}
+
+// Natural logarithm
+// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2)
+// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can
+// be easily approximated by a polynomial centered on m=1 for stability.
+// TODO(gonnet): Further reduce the interval allowing for lower-degree
+//               polynomial interpolants -> ... -> profit!
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+plog<Packet8f>(const Packet8f& _x) {
+  Packet8f x = _x;
+  _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(126f, 126.0f);
+
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  // The smallest non denormalized float number.
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(min_norm_pos, 0x00800000);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(minus_inf, 0xff800000);
+
+  // Polynomial coefficients.
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p1, -1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p3, -1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p4, +1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p5, -1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p6, +2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p7, -2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p8, +3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q2, 0.693359375f);
+
+  Packet8f invalid_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_NGE_UQ); // not greater equal is true if x is NaN
+  Packet8f iszero_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_EQ_OQ);
+
+  // Truncate input values to the minimum positive normal.
+  x = pmax(x, p8f_min_norm_pos);
+
+// 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));
+#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);
+  Packet8f emm0 = _mm256_cvtepi32_ps(_mm256_setr_m128(lo, hi));
+#endif
+  Packet8f e = _mm256_sub_ps(emm0, p8f_126f);
+
+  // Set the exponents to -1, i.e. x are in the range [0.5,1).
+  x = _mm256_and_ps(x, p8f_inv_mant_mask);
+  x = _mm256_or_ps(x, p8f_half);
+
+  // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2))
+  // and shift by -1. The values are then centered around 0, which improves
+  // the stability of the polynomial evaluation.
+  //   if( x < SQRTHF ) {
+  //     e -= 1;
+  //     x = x + x - 1.0;
+  //   } else { x = x - 1.0; }
+  Packet8f mask = _mm256_cmp_ps(x, p8f_cephes_SQRTHF, _CMP_LT_OQ);
+  Packet8f tmp = _mm256_and_ps(x, mask);
+  x = psub(x, p8f_1);
+  e = psub(e, _mm256_and_ps(p8f_1, mask));
+  x = padd(x, tmp);
+
+  Packet8f x2 = pmul(x, x);
+  Packet8f x3 = pmul(x2, x);
+
+  // Evaluate the polynomial approximant of degree 8 in three parts, probably
+  // to improve instruction-level parallelism.
+  Packet8f y, y1, y2;
+  y = pmadd(p8f_cephes_log_p0, x, p8f_cephes_log_p1);
+  y1 = pmadd(p8f_cephes_log_p3, x, p8f_cephes_log_p4);
+  y2 = pmadd(p8f_cephes_log_p6, x, p8f_cephes_log_p7);
+  y = pmadd(y, x, p8f_cephes_log_p2);
+  y1 = pmadd(y1, x, p8f_cephes_log_p5);
+  y2 = pmadd(y2, x, p8f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  // Add the logarithm of the exponent back to the result of the interpolation.
+  y1 = pmul(e, p8f_cephes_log_q1);
+  tmp = pmul(x2, p8f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p8f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+
+  // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF.
+  return _mm256_or_ps(
+      _mm256_andnot_ps(iszero_mask, _mm256_or_ps(x, invalid_mask)),
+      _mm256_and_ps(iszero_mask, p8f_minus_inf));
+}
+
+// Exponential function. Works by writing "x = m*log(2) + r" where
+// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
+// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1).
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+pexp<Packet8f>(const Packet8f& _x) {
+  _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(127, 127.0f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(exp_hi, 88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet8f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_LOG2EF, 1.44269504088896341f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p5, 5.0000001201E-1f);
+
+  // Clamp x.
+  Packet8f x = pmax(pmin(_x, p8f_exp_hi), p8f_exp_lo);
+
+  // Express exp(x) as exp(m*ln(2) + r), start by extracting
+  // m = floor(x/ln(2) + 0.5).
+  Packet8f m = _mm256_floor_ps(pmadd(x, p8f_cephes_LOG2EF, p8f_half));
+
+// Get r = x - m*ln(2). If no FMA instructions are available, m*ln(2) is
+// subtracted out in two parts, m*C1+m*C2 = m*ln(2), to avoid accumulating
+// truncation errors. Note that we don't use the "pmadd" function here to
+// ensure that a precision-preserving FMA instruction is used.
+#ifdef EIGEN_VECTORIZE_FMA
+  _EIGEN_DECLARE_CONST_Packet8f(nln2, -0.6931471805599453f);
+  Packet8f r = _mm256_fmadd_ps(m, p8f_nln2, x);
+#else
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C1, 0.693359375f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C2, -2.12194440e-4f);
+  Packet8f r = psub(x, pmul(m, p8f_cephes_exp_C1));
+  r = psub(r, pmul(m, p8f_cephes_exp_C2));
+#endif
+
+  Packet8f r2 = pmul(r, r);
+
+  // TODO(gonnet): Split into odd/even polynomials and try to exploit
+  //               instruction-level parallelism.
+  Packet8f y = p8f_cephes_exp_p0;
+  y = pmadd(y, r, p8f_cephes_exp_p1);
+  y = pmadd(y, r, p8f_cephes_exp_p2);
+  y = pmadd(y, r, p8f_cephes_exp_p3);
+  y = pmadd(y, r, p8f_cephes_exp_p4);
+  y = pmadd(y, r, p8f_cephes_exp_p5);
+  y = pmadd(y, r2, r);
+  y = padd(y, p8f_1);
+
+  // Build emm0 = 2^m.
+  Packet8i emm0 = _mm256_cvttps_epi32(padd(m, p8f_127));
+#ifdef EIGEN_VECTORIZE_AVX2
+  emm0 = _mm256_slli_epi32(emm0, 23);
+#else
+  __m128i lo = _mm_slli_epi32(_mm256_extractf128_si256(emm0, 0), 23);
+  __m128i hi = _mm_slli_epi32(_mm256_extractf128_si256(emm0, 1), 23);
+  emm0 = _mm256_setr_m128(lo, hi);
+#endif
+
+  // Return 2^m * exp(r).
+  return pmax(pmul(y, _mm256_castsi256_ps(emm0)), _x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4d
+pexp<Packet4d>(const Packet4d& _x) {
+  Packet4d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet4d(1, 1.0);
+  _EIGEN_DECLARE_CONST_Packet4d(2, 2.0);
+  _EIGEN_DECLARE_CONST_Packet4d(half, 0.5);
+
+  _EIGEN_DECLARE_CONST_Packet4d(exp_hi, 709.437);
+  _EIGEN_DECLARE_CONST_Packet4d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C2, 1.42860682030941723212e-6);
+  _EIGEN_DECLARE_CONST_Packet4i(1023, 1023);
+
+  Packet4d tmp, fx;
+
+  // clamp x
+  x = pmax(pmin(x, p4d_exp_hi), p4d_exp_lo);
+  // Express exp(x) as exp(g + n*log(2)).
+  fx = pmadd(p4d_cephes_LOG2EF, x, p4d_half);
+
+  // Get the integer modulus of log(2), i.e. the "n" described above.
+  fx = _mm256_floor_pd(fx);
+
+  // Get the remainder modulo log(2), i.e. the "g" described above. Subtract
+  // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last
+  // digits right.
+  tmp = pmul(fx, p4d_cephes_exp_C1);
+  Packet4d z = pmul(fx, p4d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet4d x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial of the rational interpolant.
+  Packet4d px = p4d_cephes_exp_p0;
+  px = pmadd(px, x2, p4d_cephes_exp_p1);
+  px = pmadd(px, x2, p4d_cephes_exp_p2);
+  px = pmul(px, x);
+
+  // Evaluate the denominator polynomial of the rational interpolant.
+  Packet4d qx = p4d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p4d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p4d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p4d_cephes_exp_q3);
+
+  // I don't really get this bit, copied from the SSE2 routines, so...
+  // TODO(gonnet): Figure out what is going on here, perhaps find a better
+  // rational interpolant?
+  x = _mm256_div_pd(px, psub(qx, px));
+  x = pmadd(p4d_2, x, p4d_1);
+
+  // Build e=2^n by constructing the exponents in a 128-bit vector and
+  // shifting them to where they belong in double-precision values.
+  __m128i emm0 = _mm256_cvtpd_epi32(fx);
+  emm0 = _mm_add_epi32(emm0, p4i_1023);
+  emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(3, 1, 2, 0));
+  __m128i lo = _mm_slli_epi64(emm0, 52);
+  __m128i hi = _mm_slli_epi64(_mm_srli_epi64(emm0, 32), 52);
+  __m256i e = _mm256_insertf128_si256(_mm256_setzero_si256(), lo, 0);
+  e = _mm256_insertf128_si256(e, hi, 1);
+
+  // 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);
+}
+
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. The main advantage of this approach is not just speed, but
+// also the fact that it can be inlined and pipelined with other computations,
+// further reducing its effective latency.
+#if EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+psqrt<Packet8f>(const Packet8f& _x) {
+  _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
+
+  Packet8f neg_half = pmul(_x, p8f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  Packet8f non_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_GE_OQ);
+  Packet8f x = _mm256_and_ps(non_zero_mask, _mm256_rsqrt_ps(_x));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five));
+
+  // Multiply the original _x by it's reciprocal square root to extract the
+  // square root.
+  return pmul(_x, x);
+}
+#else
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f psqrt<Packet8f>(const Packet8f& x) {
+  return _mm256_sqrt_ps(x);
+}
+#endif
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4d psqrt<Packet4d>(const Packet4d& x) {
+  return _mm256_sqrt_pd(x);
+}
+#if EIGEN_FAST_MATH
+
+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(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
+
+  Packet8f neg_half = pmul(_x, p8f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  Packet8f le_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_LT_OQ);
+  Packet8f x = _mm256_andnot_ps(le_zero_mask, _mm256_rsqrt_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  Packet8f neg_mask = _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_LT_OQ);
+  Packet8f zero_mask = _mm256_andnot_ps(neg_mask, le_zero_mask);
+  Packet8f infs_and_nans = _mm256_or_ps(_mm256_and_ps(neg_mask, p8f_nan),
+                                        _mm256_and_ps(zero_mask, p8f_inf));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm256_or_ps(x, infs_and_nans);
+}
+
+#else
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f prsqrt<Packet8f>(const Packet8f& x) {
+  _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
+  return _mm256_div_ps(p8f_one, _mm256_sqrt_ps(x));
+}
+#endif
+
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4d prsqrt<Packet4d>(const Packet4d& x) {
+  _EIGEN_DECLARE_CONST_Packet4d(one, 1.0);
+  return _mm256_div_pd(p4d_one, _mm256_sqrt_pd(x));
+}
+
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_AVX_H

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

@@ -22,9 +22,9 @@ namespace internal {
 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
 #endif
 
-#ifdef EIGEN_VECTORIZE_FMA
-#ifndef EIGEN_HAS_FUSED_MADD
-#define EIGEN_HAS_FUSED_MADD 1
+#ifdef __FMA__
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
 #endif
 #endif
 
@@ -42,6 +42,12 @@ template<> struct is_arithmetic<__m256d> { enum { value = true }; };
 #define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \
   const Packet4d p4d_##NAME = pset1<Packet4d>(X)
 
+#define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \
+  const Packet8f p8f_##NAME = (__m256)pset1<Packet8i>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \
+  const Packet8i p8i_##NAME = pset1<Packet8i>(X)
+
 
 template<> struct packet_traits<float>  : default_packet_traits
 {
@@ -54,13 +60,15 @@ template<> struct packet_traits<float>  : default_packet_traits
     HasHalfPacket = 1,
 
     HasDiv  = 1,
-    HasSin  = 0,
+    HasSin  = EIGEN_FAST_MATH,
     HasCos  = 0,
-    HasLog  = 0,
-    HasExp  = 0,
-    HasSqrt = 0
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasBlend = 1
   };
- };
+};
 template<> struct packet_traits<double> : default_packet_traits
 {
   typedef Packet4d type;
@@ -72,7 +80,10 @@ template<> struct packet_traits<double> : default_packet_traits
     HasHalfPacket = 1,
 
     HasDiv  = 1,
-    HasExp  = 0
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasBlend = 1
   };
 };
 
@@ -89,9 +100,9 @@ template<> struct packet_traits<int>    : default_packet_traits
 };
 */
 
-template<> struct unpacket_traits<Packet8f> { typedef float  type; typedef Packet4f half; enum {size=8}; };
-template<> struct unpacket_traits<Packet4d> { typedef double type; typedef Packet2d half; enum {size=4}; };
-template<> struct unpacket_traits<Packet8i> { typedef int    type; typedef Packet4i half; enum {size=8}; };
+template<> struct unpacket_traits<Packet8f> { typedef float  type; typedef Packet4f half; enum {size=8, alignment=Aligned32}; };
+template<> struct unpacket_traits<Packet4d> { typedef double type; typedef Packet2d half; enum {size=4, alignment=Aligned32}; };
+template<> struct unpacket_traits<Packet8i> { typedef int    type; typedef Packet4i half; enum {size=8, alignment=Aligned32}; };
 
 template<> EIGEN_STRONG_INLINE Packet8f pset1<Packet8f>(const float&  from) { return _mm256_set1_ps(from); }
 template<> EIGEN_STRONG_INLINE Packet4d pset1<Packet4d>(const double& from) { return _mm256_set1_pd(from); }
@@ -100,8 +111,8 @@ template<> EIGEN_STRONG_INLINE Packet8i pset1<Packet8i>(const int&    from) { re
 template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float*  from) { return _mm256_broadcast_ss(from); }
 template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); }
 
-template<> EIGEN_STRONG_INLINE Packet8f plset<float>(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); }
-template<> EIGEN_STRONG_INLINE Packet4d plset<double>(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); }
+template<> EIGEN_STRONG_INLINE Packet8f plset<Packet8f>(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); }
+template<> EIGEN_STRONG_INLINE Packet4d plset<Packet4d>(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); }
 
 template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); }
 template<> EIGEN_STRONG_INLINE Packet4d padd<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_add_pd(a,b); }
@@ -133,9 +144,9 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, co
   return pset1<Packet8i>(0);
 }
 
-#ifdef EIGEN_VECTORIZE_FMA
+#ifdef __FMA__
 template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) {
-#if defined(__clang__) || defined(__GNUC__)
+#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG
   // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
   // and gcc stupidly generates a vfmadd132ps instruction,
   // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate
@@ -148,7 +159,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f&
 #endif
 }
 template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) {
-#if defined(__clang__) || defined(__GNUC__)
+#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG
   // see above
   Packet4d res = c;
   __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
@@ -224,17 +235,17 @@ template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet8i&
 
 // NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available
 // NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4);
-template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, Index stride)
 {
   return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride],
                        from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4d pgather<double, Packet4d>(const double* from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline Packet4d pgather<double, Packet4d>(const double* from, Index stride)
 {
   return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet8f>(float* to, const Packet8f& from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet8f>(float* to, const Packet8f& from, Index stride)
 {
   __m128 low = _mm256_extractf128_ps(from, 0);
   to[stride*0] = _mm_cvtss_f32(low);
@@ -248,7 +259,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet8f>(float* to, co
   to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2));
   to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3));
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet4d>(double* to, const Packet4d& from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet4d>(double* to, const Packet4d& from, Index stride)
 {
   __m128d low = _mm256_extractf128_pd(from, 0);
   to[stride*0] = _mm_cvtsd_f64(low);
@@ -423,26 +434,30 @@ struct palign_impl<Offset,Packet8f>
     if (Offset==1)
     {
       first = _mm256_blend_ps(first, second, 1);
-      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1));
-      first = _mm256_blend_ps(tmp, _mm256_permute2f128_ps (tmp, tmp, 1), 0x88);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0x88);
     }
     else if (Offset==2)
     {
       first = _mm256_blend_ps(first, second, 3);
-      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2));
-      first = _mm256_blend_ps(tmp, _mm256_permute2f128_ps (tmp, tmp, 1), 0xcc);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0xcc);
     }
     else if (Offset==3)
     {
       first = _mm256_blend_ps(first, second, 7);
-      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3));
-      first = _mm256_blend_ps(tmp, _mm256_permute2f128_ps (tmp, tmp, 1), 0xee);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0xee);
     }
     else if (Offset==4)
     {
       first = _mm256_blend_ps(first, second, 15);
-      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0));
-      first = _mm256_permute_ps(_mm256_permute2f128_ps (tmp, tmp, 1), _MM_SHUFFLE(3,2,1,0));
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_permute_ps(tmp2, _MM_SHUFFLE(3,2,1,0));
     }
     else if (Offset==5)
     {
@@ -557,6 +572,19 @@ ptranspose(PacketBlock<Packet4d,4>& kernel) {
   kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49);
 }
 
+template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) {
+  const __m256 zero = _mm256_setzero_ps();
+  const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ);
+  return _mm256_blendv_ps(thenPacket, elsePacket, false_mask);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) {
+  const __m256d zero = _mm256_setzero_pd();
+  const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ);
+  return _mm256_blendv_pd(thenPacket, elsePacket, false_mask);
+}
+
 } // end namespace internal
 
 } // end namespace Eigen

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

@@ -0,0 +1,51 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@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_TYPE_CASTING_AVX_H
+#define EIGEN_TYPE_CASTING_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+// For now we use SSE to handle integers, so we can't use AVX instructions to cast
+// from int to float
+template <>
+struct type_casting_traits<float, int> {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template <>
+struct type_casting_traits<int, float> {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+
+template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
+  return _mm256_cvtps_epi32(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
+  return _mm256_cvtepi32_ps(a);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_AVX_H

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

@@ -7,23 +7,21 @@
 // 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_COMPLEX_ALTIVEC_H
-#define EIGEN_COMPLEX_ALTIVEC_H
+#ifndef EIGEN_COMPLEX32_ALTIVEC_H
+#define EIGEN_COMPLEX32_ALTIVEC_H
 
 namespace Eigen {
 
 namespace internal {
 
 static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
-static Packet16uc p16uc_COMPLEX_RE   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
-static Packet16uc p16uc_COMPLEX_IM   = vec_sld(p16uc_DUPLICATE, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
-static Packet16uc p16uc_COMPLEX_REV  = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
-static Packet16uc p16uc_COMPLEX_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);//{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
-static Packet16uc p16uc_PSET_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
-static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
-static Packet16uc p16uc_COMPLEX_MASK16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);//{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
-static Packet16uc p16uc_COMPLEX_TRANSPOSE_0 = vec_add(p16uc_PSET_HI, p16uc_COMPLEX_MASK16);//{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
-static Packet16uc p16uc_COMPLEX_TRANSPOSE_1 = vec_add(p16uc_PSET_LO, p16uc_COMPLEX_MASK16);//{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
+#ifdef _BIG_ENDIAN
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+#else
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+#endif
 
 //---------- float ----------
 struct Packet2cf
@@ -55,7 +53,7 @@ template<> struct packet_traits<std::complex<float> >  : default_packet_traits
   };
 };
 
-template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; typedef Packet2cf half; };
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
 
 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
 {
@@ -65,18 +63,18 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<flo
     res.v = pload<Packet4f>((const float *)&from);
   else
     res.v = ploadu<Packet4f>((const float *)&from);
-  res.v = vec_perm(res.v, res.v, p16uc_PSET_HI);
+  res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI);
   return res;
 }
 
-template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
 {
   std::complex<float> EIGEN_ALIGN16 af[2];
   af[0] = from[0*stride];
   af[1] = from[1*stride];
   return Packet2cf(vec_ld(0, (const float*)af));
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
 {
   std::complex<float> EIGEN_ALIGN16 af[2];
   vec_st(from.v, 0, (float*)af);
@@ -95,16 +93,16 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
   Packet4f v1, v2;
 
   // Permute and multiply the real parts of a and b
-  v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE);
+  v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
   // Get the imaginary parts of a
-  v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM);
+  v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
   // multiply a_re * b 
   v1 = vec_madd(v1, b.v, p4f_ZERO);
   // multiply a_im * b and get the conjugate result
   v2 = vec_madd(v2, b.v, p4f_ZERO);
   v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR);
   // permute back to a proper order
-  v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV);
+  v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV);
   
   return Packet2cf(vec_add(v1, v2));
 }
@@ -138,7 +136,7 @@ template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Pack
 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
 {
   Packet4f rev_a;
-  rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2);
+  rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2);
   return Packet2cf(rev_a);
 }
 
@@ -153,9 +151,13 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe
 template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
 {
   Packet4f b1, b2;
-  
+#ifdef _BIG_ENDIAN  
   b1 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8);
   b2 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8);
+#else
+  b1 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8);
+  b2 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8);
+#endif
   b2 = (Packet4f) vec_sld(b2, b2, 8);
   b2 = padd(b1, b2);
 
@@ -179,7 +181,11 @@ struct palign_impl<Offset,Packet2cf>
   {
     if (Offset==1)
     {
+#ifdef _BIG_ENDIAN
       first.v = vec_sld(first.v, second.v, 8);
+#else
+      first.v = vec_sld(second.v, first.v, 8);
+#endif
     }
   }
 };
@@ -222,23 +228,203 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con
   // TODO optimize it for AltiVec
   Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
   Packet4f s = vec_madd(b.v, b.v, p4f_ZERO);
-  return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV))));
+  return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX32_REV))));
 }
 
 template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x)
 {
-  return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV));
+  return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV));
 }
 
 EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
 {
-  Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0);
-  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1);
+  Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
   kernel.packet[0].v = tmp;
 }
 
+//---------- double ----------
+#ifdef __VSX__
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
+  Packet2d v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride)
+{
+  std::complex<double> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet1cd>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride)
+{
+  std::complex<double> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<double> >(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  Packet2d a_re, a_im, v1, v2;
+
+  // Permute and multiply the real parts of a and b
+  a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI);
+  // Get the imaginary parts of a
+  a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO);
+  // multiply a_re * b
+  v1 = vec_madd(a_re, b.v, p2d_ZERO);
+  // multiply a_im * b and get the conjugate result
+  v2 = vec_madd(a_im, b.v, p2d_ZERO);
+  v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8);
+  v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR1);
+
+  return Packet1cd(vec_add(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from)
+{
+  return pset1<Packet1cd>(*from);
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { vec_dstt((long *)addr, DST_CTRL(2,2,32), DST_CHAN); }
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  std::complex<double> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<double> >(res, a);
+
+  return res[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
+{
+  return vecs[0];
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for AltiVec
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_);
+  return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_REVERSE64))));
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+#endif // __VSX__
 } // end namespace internal
 
 } // end namespace Eigen
 
-#endif // EIGEN_COMPLEX_ALTIVEC_H
+#endif // EIGEN_COMPLEX32_ALTIVEC_H

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

@@ -0,0 +1,290 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Julien Pommier
+// Copyright (C) 2009 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/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+#define EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f plog<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+  _EIGEN_DECLARE_CONST_Packet4i(23, 23);
+
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  /* the smallest non denormalized float number */
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000); // -1.f/0.f
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan,     0xffffffff);
+  
+  /* natural logarithm computed for 4 simultaneous float
+    return NaN for x <= 0
+  */
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+
+
+  Packet4i emm0;
+
+  /* isvalid_mask is 0 if x < 0 or x is NaN. */
+  Packet4ui isvalid_mask = reinterpret_cast<Packet4ui>(vec_cmpge(x, p4f_ZERO));
+  Packet4ui iszero_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(x, p4f_ZERO));
+
+  x = pmax(x, p4f_min_norm_pos);  /* cut off denormalized stuff */
+  emm0 = vec_sr(reinterpret_cast<Packet4i>(x),
+                reinterpret_cast<Packet4ui>(p4i_23));
+
+  /* keep only the fractional part */
+  x = pand(x, p4f_inv_mant_mask);
+  x = por(x, p4f_half);
+
+  emm0 = psub(emm0, p4i_0x7f);
+  Packet4f e = padd(vec_ctf(emm0, 0), p4f_1);
+
+  /* part2:
+     if( x < SQRTHF ) {
+       e -= 1;
+       x = x + x - 1.0;
+     } else { x = x - 1.0; }
+  */
+  Packet4f mask = reinterpret_cast<Packet4f>(vec_cmplt(x, p4f_cephes_SQRTHF));
+  Packet4f tmp = pand(x, mask);
+  x = psub(x, p4f_1);
+  e = psub(e, pand(p4f_1, mask));
+  x = padd(x, tmp);
+
+  Packet4f x2 = pmul(x,x);
+  Packet4f x3 = pmul(x2,x);
+
+  Packet4f y, y1, y2;
+  y  = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
+  y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
+  y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
+  y  = pmadd(y , x, p4f_cephes_log_p2);
+  y1 = pmadd(y1, x, p4f_cephes_log_p5);
+  y2 = pmadd(y2, x, p4f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  y1 = pmul(e, p4f_cephes_log_q1);
+  tmp = pmul(x2, p4f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p4f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+  // negative arg will be NAN, 0 will be -INF
+  x = vec_sel(x, p4f_minus_inf, iszero_mask);
+  x = vec_sel(p4f_minus_nan, x, isvalid_mask);
+  return x;
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+  _EIGEN_DECLARE_CONST_Packet4i(23, 23);
+
+
+  _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+
+  Packet4f tmp, fx;
+  Packet4i emm0;
+
+  // clamp x
+  x = vec_max(vec_min(x, p4f_exp_hi), p4f_exp_lo);
+
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half);
+
+  fx = vec_floor(fx);
+
+  tmp = pmul(fx, p4f_cephes_exp_C1);
+  Packet4f z = pmul(fx, p4f_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  z = pmul(x,x);
+
+  Packet4f y = p4f_cephes_exp_p0;
+  y = pmadd(y, x, p4f_cephes_exp_p1);
+  y = pmadd(y, x, p4f_cephes_exp_p2);
+  y = pmadd(y, x, p4f_cephes_exp_p3);
+  y = pmadd(y, x, p4f_cephes_exp_p4);
+  y = pmadd(y, x, p4f_cephes_exp_p5);
+  y = pmadd(y, z, x);
+  y = padd(y, p4f_1);
+
+  // build 2^n
+  emm0 = vec_cts(fx, 0);
+  emm0 = vec_add(emm0, p4i_0x7f);
+  emm0 = vec_sl(emm0, reinterpret_cast<Packet4ui>(p4i_23));
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet4ui isnumber_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(y, reinterpret_cast<Packet4f>(emm0)), _x),
+                 isnumber_mask);
+}
+
+#ifdef __VSX__
+// VSX support varies between different compilers and even different
+// versions of the same compiler.  For gcc version >= 4.9.3, we can use
+// vec_cts to efficiently convert Packet2d to Packet2l.  Otherwise, use
+// a slow version that works with older compilers. 
+static inline Packet2l ConvertToPacket2l(const Packet2d& x) {
+#if EIGEN_GNUC_AT_LEAST(5, 0) || \
+    (EIGEN_GNUC_AT(4, 9) && __GNUC_PATCHLEVEL__ >= 3)
+  return vec_cts(x, 0);    // TODO: check clang version.
+#else
+  double tmp[2];
+  memcpy(tmp, &x, sizeof(tmp));
+  Packet2l l = { static_cast<long long>(tmp[0]),
+                 static_cast<long long>(tmp[1]) };
+  return l;
+#endif
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d pexp<Packet2d>(const Packet2d& _x)
+{
+  Packet2d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+  _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+  _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+  _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+  _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+  Packet2d tmp, fx;
+  Packet2l emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(p2d_cephes_LOG2EF, x, p2d_half);
+
+  fx = vec_floor(fx);
+
+  tmp = pmul(fx, p2d_cephes_exp_C1);
+  Packet2d z = pmul(fx, p2d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet2d x2 = pmul(x,x);
+
+  Packet2d px = p2d_cephes_exp_p0;
+  px = pmadd(px, x2, p2d_cephes_exp_p1);
+  px = pmadd(px, x2, p2d_cephes_exp_p2);
+  px = pmul (px, x);
+
+  Packet2d qx = p2d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
+
+  x = pdiv(px,psub(qx,px));
+  x = pmadd(p2d_2,x,p2d_1);
+
+  // build 2^n
+  emm0 = ConvertToPacket2l(fx);
+
+#ifdef __POWER8_VECTOR__ 
+  static const Packet2l p2l_1023 = { 1023, 1023 };
+  static const Packet2ul p2ul_52 = { 52, 52 };
+
+  emm0 = vec_add(emm0, p2l_1023);
+  emm0 = vec_sl(emm0, p2ul_52);
+#else
+  // Code is a bit complex for POWER7.  There is actually a
+  // vec_xxsldi intrinsic but it is not supported by some gcc versions.
+  // So we shift (52-32) bits and do a word swap with zeros.
+  _EIGEN_DECLARE_CONST_Packet4i(1023, 1023);
+  _EIGEN_DECLARE_CONST_Packet4i(20, 20);    // 52 - 32
+
+  Packet4i emm04i = reinterpret_cast<Packet4i>(emm0);
+  emm04i = vec_add(emm04i, p4i_1023);
+  emm04i = vec_sl(emm04i, reinterpret_cast<Packet4ui>(p4i_20));
+  static const Packet16uc perm = {
+    0x14, 0x15, 0x16, 0x17, 0x00, 0x01, 0x02, 0x03, 
+    0x1c, 0x1d, 0x1e, 0x1f, 0x08, 0x09, 0x0a, 0x0b };
+#ifdef  _BIG_ENDIAN
+  emm0 = reinterpret_cast<Packet2l>(vec_perm(p4i_ZERO, emm04i, perm));
+#else
+  emm0 = reinterpret_cast<Packet2l>(vec_perm(emm04i, p4i_ZERO, perm));
+#endif
+
+#endif
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet2ul isnumber_mask = reinterpret_cast<Packet2ul>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(x, reinterpret_cast<Packet2d>(emm0)), _x),
+                 isnumber_mask);
+}
+#endif
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_ALTIVEC_H

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

@@ -18,17 +18,17 @@ namespace internal {
 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
 #endif
 
-#ifndef EIGEN_HAS_FUSED_MADD
-#define EIGEN_HAS_FUSED_MADD 1
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
 #endif
 
-#ifndef EIGEN_HAS_FUSE_CJMADD
-#define EIGEN_HAS_FUSE_CJMADD 1
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
 #endif
 
 // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
-#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  32
 #endif
 
 typedef __vector float          Packet4f;
@@ -50,30 +50,77 @@ typedef __vector unsigned char  Packet16uc;
 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
   Packet4f p4f_##NAME = pset1<Packet4f>(X)
 
-#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
-  Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
-
 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
   Packet4i p4i_##NAME = pset1<Packet4i>(X)
 
+#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+  Packet2d p2d_##NAME = pset1<Packet2d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
+  Packet2l p2l_##NAME = pset1<Packet2l>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+  const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
+
 #define DST_CHAN 1
 #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
 
-// Define global static constants:
-static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
-static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
-static Packet16uc p16uc_REVERSE = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3};
-static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); //{ 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15}
-static Packet16uc p16uc_DUPLICATE = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7};
 
+// These constants are endian-agnostic
 static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+#ifndef __VSX__
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
+static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
+#endif
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
-static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
 static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
 
+static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
+static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
+
+static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 };
+static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
+
+// Mask alignment
+#ifdef __PPC64__
+#define _EIGEN_MASK_ALIGNMENT	0xfffffffffffffff0
+#else
+#define _EIGEN_MASK_ALIGNMENT	0xfffffff0
+#endif
+
+#define _EIGEN_ALIGNED_PTR(x)	((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+
+// Handle endianness properly while loading constants
+// Define global static constants:
+#ifdef _BIG_ENDIAN
+static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); 
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET32_WODD   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN  = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+#else
+static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; 
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+#endif // _BIG_ENDIAN
+
+static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
+static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16);                                         //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
+static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16);                                         //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
+
+static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
+
+#ifdef _BIG_ENDIAN
+static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#else
+static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_PSET64_HI, p16uc_PSET64_LO, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#endif // _BIG_ENDIAN
+
 template<> struct packet_traits<float>  : default_packet_traits
 {
   typedef Packet4f type;
@@ -88,8 +135,8 @@ template<> struct packet_traits<float>  : default_packet_traits
     HasDiv  = 1,
     HasSin  = 0,
     HasCos  = 0,
-    HasLog  = 0,
-    HasExp  = 0,
+    HasLog  = 1,
+    HasExp  = 1,
     HasSqrt = 0
   };
 };
@@ -105,9 +152,22 @@ template<> struct packet_traits<int>    : default_packet_traits
   };
 };
 
-template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}; typedef Packet4f half; };
-template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; typedef Packet4i half; };
-/*
+
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+
+inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v)
+{
+  union {
+    Packet16uc   v;
+    unsigned char n[16];
+  } vt;
+  vt.v = v;
+  for (int i=0; i< 16; i++)
+    s << (int)vt.n[i] << ", ";
+  return s;
+}
+
 inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
 {
   union {
@@ -140,7 +200,7 @@ inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
   return s;
 }
-
+/*
 inline std::ostream & operator <<(std::ostream & s, const Packetbi & v)
 {
   union {
@@ -150,14 +210,21 @@ inline std::ostream & operator <<(std::ostream & s, const Packetbi & v)
   vt.v = v;
   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
   return s;
-}
-*/
+}*/
+
+
+// Need to define them first or we get specialization after instantiation errors
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
 
 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) {
   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
   float EIGEN_ALIGN16 af[4];
   af[0] = from;
-  Packet4f vc = vec_ld(0, af);
+  Packet4f vc = pload<Packet4f>(af);
   vc = vec_splat(vc, 0);
   return vc;
 }
@@ -165,17 +232,15 @@ template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) {
 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   {
   int EIGEN_ALIGN16 ai[4];
   ai[0] = from;
-  Packet4i vc = vec_ld(0, ai);
+  Packet4i vc = pload<Packet4i>(ai);
   vc = vec_splat(vc, 0);
   return vc;
 }
-
-
 template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet4f>(const float *a,
                       Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
 {
-  a3 = vec_ld(0,a);
+  a3 = pload<Packet4f>(a);
   a0 = vec_splat(a3, 0);
   a1 = vec_splat(a3, 1);
   a2 = vec_splat(a3, 2);
@@ -185,52 +250,52 @@ template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet4i>(const int *a,
                       Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3)
 {
-  a3 = vec_ld(0,a);
+  a3 = pload<Packet4i>(a);
   a0 = vec_splat(a3, 0);
   a1 = vec_splat(a3, 1);
   a2 = vec_splat(a3, 2);
   a3 = vec_splat(a3, 3);
 }
 
-template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
 {
   float EIGEN_ALIGN16 af[4];
   af[0] = from[0*stride];
   af[1] = from[1*stride];
   af[2] = from[2*stride];
   af[3] = from[3*stride];
- return vec_ld(0, af);
+ return pload<Packet4f>(af);
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
 {
   int EIGEN_ALIGN16 ai[4];
   ai[0] = from[0*stride];
   ai[1] = from[1*stride];
   ai[2] = from[2*stride];
   ai[3] = from[3*stride];
- return vec_ld(0, ai);
+ return pload<Packet4i>(ai);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
 {
   float EIGEN_ALIGN16 af[4];
-  vec_st(from, 0, af);
+  pstore<float>(af, from);
   to[0*stride] = af[0];
   to[1*stride] = af[1];
   to[2*stride] = af[2];
   to[3*stride] = af[3];
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, DenseIndex stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
 {
   int EIGEN_ALIGN16 ai[4];
-  vec_st(from, 0, ai);
+  pstore<int>((int *)ai, from);
   to[0*stride] = ai[0];
   to[1*stride] = ai[1];
   to[2*stride] = ai[2];
   to[3*stride] = ai[3];
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); }
-template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a)     { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)   { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); }
 
 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); }
 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); }
@@ -283,7 +348,8 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const
 */
 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
 {
-  Packet4f t, y_0, y_1, res;
+#ifndef __VSX__  // VSX actually provides a div instruction
+  Packet4f t, y_0, y_1;
 
   // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
   y_0 = vec_re(b);
@@ -292,8 +358,10 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const
   t   = vec_nmsub(y_0, b, p4f_ONE);
   y_1 = vec_madd(y_0, t, y_0);
 
-  res = vec_madd(a, y_1, p4f_ZERO);
-  return res;
+  return vec_madd(a, y_1, p4f_ZERO);
+#else
+  return vec_div(a, b);
+#endif
 }
 
 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
@@ -311,7 +379,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const
 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
 
-// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
 
@@ -324,13 +391,10 @@ template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const
 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); }
 
-template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
-template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
-
+#ifdef _BIG_ENDIAN
 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
 {
   EIGEN_DEBUG_ALIGNED_LOAD
-  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
   Packet16uc MSQ, LSQ;
   Packet16uc mask;
   MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
@@ -350,25 +414,36 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
   mask = vec_lvsl(0, from);                        // create the permute mask
   return (Packet4i) vec_perm(MSQ, LSQ, mask);    // align the data
 }
+#else
+// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+  return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from));
+}
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+  return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from));
+}
+#endif
 
 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
 {
   Packet4f p;
   if((ptrdiff_t(from) % 16) == 0)  p = pload<Packet4f>(from);
   else                             p = ploadu<Packet4f>(from);
-  return vec_perm(p, p, p16uc_DUPLICATE);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
 {
   Packet4i p;
   if((ptrdiff_t(from) % 16) == 0)  p = pload<Packet4i>(from);
   else                             p = ploadu<Packet4i>(from);
-  return vec_perm(p, p, p16uc_DUPLICATE);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 
-template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
-template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
-
+#ifdef _BIG_ENDIAN
 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from)
 {
   EIGEN_DEBUG_UNALIGNED_STORE
@@ -405,15 +480,30 @@ template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& f
   vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
 }
+#else
+// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to));
+}
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet4f& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
+}
+#endif
 
+#ifndef __VSX__
 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
+#endif
 
 template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
 
-template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
-template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); }
+template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); }
 
 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); }
 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
@@ -460,7 +550,11 @@ template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
 {
   Packet4i sum;
   sum = vec_sums(a, p4i_ZERO);
+#ifdef _BIG_ENDIAN
   sum = vec_sld(sum, p4i_ZERO, 12);
+#else
+  sum = vec_sld(p4i_ZERO, sum, 4);
+#endif
   return pfirst(sum);
 }
 
@@ -547,8 +641,25 @@ struct palign_impl<Offset,Packet4f>
 {
   static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
   {
-    if (Offset!=0)
-      first = vec_sld(first, second, Offset*4);
+#ifdef _BIG_ENDIAN
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+#else
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(second, first, 12); break;
+    case 2:
+      first = vec_sld(second, first, 8); break;
+    case 3:
+      first = vec_sld(second, first, 4); break;
+    }
+#endif
   }
 };
 
@@ -557,8 +668,25 @@ struct palign_impl<Offset,Packet4i>
 {
   static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
   {
-    if (Offset!=0)
-      first = vec_sld(first, second, Offset*4);
+#ifdef _BIG_ENDIAN
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+#else
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(second, first, 12); break;
+    case 2:
+      first = vec_sld(second, first, 8); break;
+    case 3:
+      first = vec_sld(second, first, 4); break;
+    }
+#endif
   }
 };
 
@@ -588,6 +716,222 @@ ptranspose(PacketBlock<Packet4i,4>& kernel) {
   kernel.packet[3] = vec_mergel(t1, t3);
 }
 
+
+//---------- double ----------
+#ifdef __VSX__
+typedef __vector double              Packet2d;
+typedef __vector unsigned long long  Packet2ul;
+typedef __vector long long           Packet2l;
+
+static Packet2l p2l_ZERO = (Packet2l) p4i_ZERO;
+static Packet2d p2d_ONE = { 1.0, 1.0 }; 
+static Packet2d p2d_ZERO = (Packet2d) p4f_ZERO;
+static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
+
+#ifdef _BIG_ENDIAN
+static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ZERO, (Packet16uc) p2d_ONE, 8);
+#else
+static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ONE, (Packet16uc) p2d_ZERO, 8);
+#endif
+
+static EIGEN_STRONG_INLINE Packet2d vec_splat_dbl(Packet2d& a, int index)
+{
+  switch (index) {
+  case 0:
+    return (Packet2d) vec_perm(a, a, p16uc_PSET64_HI);
+  case 1:
+    return (Packet2d) vec_perm(a, a, p16uc_PSET64_LO);
+  }
+  return a;
+}
+
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 0,
+
+    HasDiv  = 1,
+    HasExp  = 1,
+    HasSqrt = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
+{
+  union {
+    Packet2d   v;
+    double n[2];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1];
+  return s;
+}
+
+// Need to define them first or we get specialization after instantiation errors
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return (Packet2d) vec_ld(0, (const float *) from); } //FIXME
+
+template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st((Packet4f)from, 0, (float *)to); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double&  from) {
+  double EIGEN_ALIGN16 af[2];
+  af[0] = from;
+  Packet2d vc = pload<Packet2d>(af);
+  vc = vec_splat_dbl(vc, 0);
+  return vc;
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+  a1 = pload<Packet2d>(a);
+  a0 = vec_splat_dbl(a1, 0);
+  a1 = vec_splat_dbl(a1, 1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec_splat_dbl(a3, 0);
+  a3 = vec_splat_dbl(a3, 1);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+ return pload<Packet2d>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  pstore<double>(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return vec_add(pset1<Packet2d>(a), p2d_COUNTDOWN); }
+
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_add(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_sub(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return psub<Packet2d>(p2d_ZERO, a); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); }
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); }
+
+// for some weird raisons, it has to be overloaded for packet of integers
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
+
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+  return (Packet2d) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from));
+}
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
+{
+  Packet2d p;
+  if((ptrdiff_t(from) % 16) == 0)  p = pload<Packet2d>(from);
+  else                             p = ploadu<Packet2d>(from);
+  return vec_perm(p, p, p16uc_PSET64_HI);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { vec_dstt((const float *) addr, DST_CTRL(2,2,32), DST_CHAN); }
+
+template<> EIGEN_STRONG_INLINE double  pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return (Packet2d)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE64); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); }
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  Packet2d b, sum;
+  b   = (Packet2d) vec_sld((Packet4ui) a, (Packet4ui)a, 8);
+  sum = vec_add(a, b);
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  Packet2d v[2], sum;
+  v[0] = vec_add(vecs[0], (Packet2d) vec_sld((Packet4ui) vecs[0], (Packet4ui) vecs[0], 8));
+  v[1] = vec_add(vecs[1], (Packet2d) vec_sld((Packet4ui) vecs[1], (Packet4ui) vecs[1], 8));
+ 
+#ifdef _BIG_ENDIAN
+ sum = (Packet2d) vec_sld((Packet4ui) v[0], (Packet4ui) v[1], 8);
+#else
+  sum = (Packet2d) vec_sld((Packet4ui) v[1], (Packet4ui) v[0], 8);
+#endif
+
+  return sum;
+}
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmul(a, (Packet2d)vec_sld((Packet4ui) a, (Packet4ui) a, 8)));
+}
+
+// min
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
+{
+  return pfirst(vec_min(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)));
+}
+
+// max
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
+{
+  return pfirst(vec_max(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
+  {
+    if (Offset == 1)
+#ifdef _BIG_ENDIAN
+      first = (Packet2d) vec_sld((Packet4ui) first, (Packet4ui) second, 8);
+#else
+      first = (Packet2d) vec_sld((Packet4ui) second, (Packet4ui) first, 8);
+#endif
+  }
+};
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  Packet2d t0, t1;
+  t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
+  t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
+  kernel.packet[0] = t0;
+  kernel.packet[1] = t1;
+}
+
+#endif // __VSX__
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/arch/CMakeLists.txt

@@ -1,5 +1,9 @@
-ADD_SUBDIRECTORY(SSE)
 ADD_SUBDIRECTORY(AltiVec)
-ADD_SUBDIRECTORY(NEON)
 ADD_SUBDIRECTORY(AVX)
+ADD_SUBDIRECTORY(CUDA)
 ADD_SUBDIRECTORY(Default)
+ADD_SUBDIRECTORY(NEON)
+ADD_SUBDIRECTORY(SSE)
+
+
+

Eigen/src/Core/arch/CUDA/CMakeLists.txt

@@ -0,0 +1,6 @@
+FILE(GLOB Eigen_Core_arch_CUDA_SRCS "*.h")
+
+INSTALL(FILES
+  ${Eigen_Core_arch_CUDA_SRCS}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/CUDA COMPONENT Devel
+)

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

@@ -0,0 +1,75 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@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_MATH_FUNCTIONS_CUDA_H
+#define EIGEN_MATH_FUNCTIONS_CUDA_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plog<float4>(const float4& a)
+{
+  return make_float4(logf(a.x), logf(a.y), logf(a.z), logf(a.w));
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plog<double2>(const double2& a)
+{
+  return make_double2(log(a.x), log(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pexp<float4>(const float4& a)
+{
+  return make_float4(expf(a.x), expf(a.y), expf(a.z), expf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pexp<double2>(const double2& a)
+{
+  return make_double2(exp(a.x), exp(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 psqrt<float4>(const float4& a)
+{
+  return make_float4(sqrtf(a.x), sqrtf(a.y), sqrtf(a.z), sqrtf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 psqrt<double2>(const double2& a)
+{
+  return make_double2(sqrt(a.x), sqrt(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 prsqrt<float4>(const float4& a)
+{
+  return make_float4(rsqrtf(a.x), rsqrtf(a.y), rsqrtf(a.z), rsqrtf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 prsqrt<double2>(const double2& a)
+{
+  return make_double2(rsqrt(a.x), rsqrt(a.y));
+}
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATH_FUNCTIONS_CUDA_H