ruby-eigen 0.0.9 → 0.0.10.pre1

data/ext/eigen/eigen3/Eigen/src/Core/util/Macros.h

@@ -0,0 +1,451 @@
+// 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) 2006-2008 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_MACROS_H
+#define EIGEN_MACROS_H
+
+#define EIGEN_WORLD_VERSION 3
+#define EIGEN_MAJOR_VERSION 2
+#define EIGEN_MINOR_VERSION 7
+
+#define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \
+                                      (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \
+                                                                 EIGEN_MINOR_VERSION>=z))))
+#ifdef __GNUC__
+  #define EIGEN_GNUC_AT_LEAST(x,y) ((__GNUC__==x && __GNUC_MINOR__>=y) || __GNUC__>x)
+#else
+  #define EIGEN_GNUC_AT_LEAST(x,y) 0
+#endif
+ 
+#ifdef __GNUC__
+  #define EIGEN_GNUC_AT_MOST(x,y) ((__GNUC__==x && __GNUC_MINOR__<=y) || __GNUC__<x)
+#else
+  #define EIGEN_GNUC_AT_MOST(x,y) 0
+#endif
+
+#if EIGEN_GNUC_AT_MOST(4,3) && !defined(__clang__)
+  // see bug 89
+  #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 0
+#else
+  #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 1
+#endif
+
+#if defined(__GNUC__) && (__GNUC__ <= 3)
+#define EIGEN_GCC3_OR_OLDER 1
+#else
+#define EIGEN_GCC3_OR_OLDER 0
+#endif
+
+// 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
+// 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
+// enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
+// certain common platform (compiler+architecture combinations) to avoid these problems.
+// Only static alignment is really problematic (relies on nonstandard compiler extensions that don't
+// work everywhere, for example don't work on GCC/ARM), try to keep heap alignment even
+// when we have to disable static alignment.
+#if defined(__GNUC__) && !(defined(__i386__) || defined(__x86_64__) || defined(__powerpc__) || defined(__ppc__) || defined(__ia64__))
+#define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+#else
+#define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
+#endif
+
+// static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX
+#if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \
+ && !EIGEN_GCC3_OR_OLDER \
+ && !defined(__SUNPRO_CC) \
+ && !defined(__QNXNTO__)
+  #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1
+#else
+  #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0
+#endif
+
+#ifdef EIGEN_DONT_ALIGN
+  #ifndef EIGEN_DONT_ALIGN_STATICALLY
+    #define EIGEN_DONT_ALIGN_STATICALLY
+  #endif
+  #define EIGEN_ALIGN 0
+#else
+  #define EIGEN_ALIGN 1
+#endif
+
+// EIGEN_ALIGN_STATICALLY is the true test whether we want to align arrays on the stack or not. It takes into account both the user choice to explicitly disable
+// alignment (EIGEN_DONT_ALIGN_STATICALLY) and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT). Henceforth, only EIGEN_ALIGN_STATICALLY should be used.
+#if EIGEN_ARCH_WANTS_STACK_ALIGNMENT && !defined(EIGEN_DONT_ALIGN_STATICALLY)
+  #define EIGEN_ALIGN_STATICALLY 1
+#else
+  #define EIGEN_ALIGN_STATICALLY 0
+  #ifndef EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
+    #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
+  #endif
+#endif
+
+#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION RowMajor
+#else
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ColMajor
+#endif
+
+#ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
+#endif
+
+// A Clang feature extension to determine compiler features.
+// We use it to determine 'cxx_rvalue_references'
+#ifndef __has_feature
+# define __has_feature(x) 0
+#endif
+
+// Do we support r-value references?
+#if (__has_feature(cxx_rvalue_references) || \
+     defined(__GXX_EXPERIMENTAL_CXX0X__) || \
+     (defined(_MSC_VER) && _MSC_VER >= 1600))
+  #define EIGEN_HAVE_RVALUE_REFERENCES
+#endif
+
+
+// Cross compiler wrapper around LLVM's __has_builtin
+#ifdef __has_builtin
+#  define EIGEN_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define EIGEN_HAS_BUILTIN(x) 0
+#endif
+
+/** Allows to disable some optimizations which might affect the accuracy of the result.
+  * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them.
+  * They currently include:
+  *   - single precision Cwise::sin() and Cwise::cos() when SSE vectorization is enabled.
+  */
+#ifndef EIGEN_FAST_MATH
+#define EIGEN_FAST_MATH 1
+#endif
+
+#define EIGEN_DEBUG_VAR(x) std::cerr << #x << " = " << x << std::endl;
+
+// concatenate two tokens
+#define EIGEN_CAT2(a,b) a ## b
+#define EIGEN_CAT(a,b) EIGEN_CAT2(a,b)
+
+// convert a token to a string
+#define EIGEN_MAKESTRING2(a) #a
+#define EIGEN_MAKESTRING(a) EIGEN_MAKESTRING2(a)
+
+// EIGEN_STRONG_INLINE is a stronger version of the inline, using __forceinline on MSVC,
+// but it still doesn't use GCC's always_inline. This is useful in (common) situations where MSVC needs forceinline
+// but GCC is still doing fine with just inline.
+#if (defined _MSC_VER) || (defined __INTEL_COMPILER)
+#define EIGEN_STRONG_INLINE __forceinline
+#else
+#define EIGEN_STRONG_INLINE inline
+#endif
+
+// EIGEN_ALWAYS_INLINE is the stronget, it has the effect of making the function inline and adding every possible
+// attribute to maximize inlining. This should only be used when really necessary: in particular,
+// it uses __attribute__((always_inline)) on GCC, which most of the time is useless and can severely harm compile times.
+// FIXME with the always_inline attribute,
+// gcc 3.4.x reports the following compilation error:
+//   Eval.h:91: sorry, unimplemented: inlining failed in call to 'const Eigen::Eval<Derived> Eigen::MatrixBase<Scalar, Derived>::eval() const'
+//    : function body not available
+#if EIGEN_GNUC_AT_LEAST(4,0)
+#define EIGEN_ALWAYS_INLINE __attribute__((always_inline)) inline
+#else
+#define EIGEN_ALWAYS_INLINE EIGEN_STRONG_INLINE
+#endif
+
+#if (defined __GNUC__)
+#define EIGEN_DONT_INLINE __attribute__((noinline))
+#elif (defined _MSC_VER)
+#define EIGEN_DONT_INLINE __declspec(noinline)
+#else
+#define EIGEN_DONT_INLINE
+#endif
+
+#if (defined __GNUC__)
+#define EIGEN_PERMISSIVE_EXPR __extension__
+#else
+#define EIGEN_PERMISSIVE_EXPR
+#endif
+
+// this macro allows to get rid of linking errors about multiply defined functions.
+//  - static is not very good because it prevents definitions from different object files to be merged.
+//           So static causes the resulting linked executable to be bloated with multiple copies of the same function.
+//  - inline is not perfect either as it unwantedly hints the compiler toward inlining the function.
+#define EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+#define EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS inline
+
+#ifdef NDEBUG
+# ifndef EIGEN_NO_DEBUG
+#  define EIGEN_NO_DEBUG
+# endif
+#endif
+
+// eigen_plain_assert is where we implement the workaround for the assert() bug in GCC <= 4.3, see bug 89
+#ifdef EIGEN_NO_DEBUG
+  #define eigen_plain_assert(x)
+#else
+  #if EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO
+    namespace Eigen {
+    namespace internal {
+    inline bool copy_bool(bool b) { return b; }
+    }
+    }
+    #define eigen_plain_assert(x) assert(x)
+  #else
+    // work around bug 89
+    #include <cstdlib>   // for abort
+    #include <iostream>  // for std::cerr
+
+    namespace Eigen {
+    namespace internal {
+    // trivial function copying a bool. Must be EIGEN_DONT_INLINE, so we implement it after including Eigen headers.
+    // see bug 89.
+    namespace {
+    EIGEN_DONT_INLINE bool copy_bool(bool b) { return b; }
+    }
+    inline void assert_fail(const char *condition, const char *function, const char *file, int line)
+    {
+      std::cerr << "assertion failed: " << condition << " in function " << function << " at " << file << ":" << line << std::endl;
+      abort();
+    }
+    }
+    }
+    #define eigen_plain_assert(x) \
+      do { \
+        if(!Eigen::internal::copy_bool(x)) \
+          Eigen::internal::assert_fail(EIGEN_MAKESTRING(x), __PRETTY_FUNCTION__, __FILE__, __LINE__); \
+      } while(false)
+  #endif
+#endif
+
+// eigen_assert can be overridden
+#ifndef eigen_assert
+#define eigen_assert(x) eigen_plain_assert(x)
+#endif
+
+#ifdef EIGEN_INTERNAL_DEBUGGING
+#define eigen_internal_assert(x) eigen_assert(x)
+#else
+#define eigen_internal_assert(x)
+#endif
+
+#ifdef EIGEN_NO_DEBUG
+#define EIGEN_ONLY_USED_FOR_DEBUG(x) (void)x
+#else
+#define EIGEN_ONLY_USED_FOR_DEBUG(x)
+#endif
+
+#ifndef EIGEN_NO_DEPRECATED_WARNING
+  #if (defined __GNUC__)
+    #define EIGEN_DEPRECATED __attribute__((deprecated))
+  #elif (defined _MSC_VER)
+    #define EIGEN_DEPRECATED __declspec(deprecated)
+  #else
+    #define EIGEN_DEPRECATED
+  #endif
+#else
+  #define EIGEN_DEPRECATED
+#endif
+
+#if (defined __GNUC__)
+#define EIGEN_UNUSED __attribute__((unused))
+#else
+#define EIGEN_UNUSED
+#endif
+
+// Suppresses 'unused variable' warnings.
+namespace Eigen {
+  namespace internal {
+    template<typename T> void ignore_unused_variable(const T&) {}
+  }
+}
+#define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var);
+
+#if !defined(EIGEN_ASM_COMMENT)
+  #if (defined __GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
+    #define EIGEN_ASM_COMMENT(X)  __asm__("#" X)
+  #else
+    #define EIGEN_ASM_COMMENT(X)
+  #endif
+#endif
+
+/* EIGEN_ALIGN_TO_BOUNDARY(n) forces data to be n-byte aligned. This is used to satisfy SIMD requirements.
+ * However, we do that EVEN if vectorization (EIGEN_VECTORIZE) is disabled,
+ * so that vectorization doesn't affect binary compatibility.
+ *
+ * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link
+ * vectorized and non-vectorized code.
+ */
+#if (defined __GNUC__) || (defined __PGI) || (defined __IBMCPP__) || (defined __ARMCC_VERSION)
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+#elif (defined _MSC_VER)
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n))
+#elif (defined __SUNPRO_CC)
+  // FIXME not sure about this one:
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+#else
+  #error Please tell me what is the equivalent of __attribute__((aligned(n))) for your compiler
+#endif
+
+#define EIGEN_ALIGN8  EIGEN_ALIGN_TO_BOUNDARY(8)
+#define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16)
+
+#if EIGEN_ALIGN_STATICALLY
+#define EIGEN_USER_ALIGN_TO_BOUNDARY(n) EIGEN_ALIGN_TO_BOUNDARY(n)
+#define EIGEN_USER_ALIGN16 EIGEN_ALIGN16
+#else
+#define EIGEN_USER_ALIGN_TO_BOUNDARY(n)
+#define EIGEN_USER_ALIGN16
+#endif
+
+#ifdef EIGEN_DONT_USE_RESTRICT_KEYWORD
+  #define EIGEN_RESTRICT
+#endif
+#ifndef EIGEN_RESTRICT
+  #define EIGEN_RESTRICT __restrict
+#endif
+
+#ifndef EIGEN_STACK_ALLOCATION_LIMIT
+// 131072 == 128 KB
+#define EIGEN_STACK_ALLOCATION_LIMIT 131072
+#endif
+
+#ifndef EIGEN_DEFAULT_IO_FORMAT
+#ifdef EIGEN_MAKING_DOCS
+// format used in Eigen's documentation
+// needed to define it here as escaping characters in CMake add_definition's argument seems very problematic.
+#define EIGEN_DEFAULT_IO_FORMAT Eigen::IOFormat(3, 0, " ", "\n", "", "")
+#else
+#define EIGEN_DEFAULT_IO_FORMAT Eigen::IOFormat()
+#endif
+#endif
+
+// just an empty macro !
+#define EIGEN_EMPTY
+
+#if defined(_MSC_VER) && (_MSC_VER < 1900) && (!defined(__INTEL_COMPILER))
+#define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+  using Base::operator =;
+#elif defined(__clang__) // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653)
+#define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+  using Base::operator =; \
+  EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \
+  template <typename OtherDerived> \
+  EIGEN_STRONG_INLINE Derived& operator=(const DenseBase<OtherDerived>& other) { Base::operator=(other.derived()); return *this; }
+#else
+#define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+  using Base::operator =; \
+  EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) \
+  { \
+    Base::operator=(other); \
+    return *this; \
+  }
+#endif
+
+/** \internal
+ * \brief Macro to manually inherit assignment operators.
+ * This is necessary, because the implicitly defined assignment operator gets deleted when a custom operator= is defined.
+ */
+#define EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Derived) EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived)
+
+/**
+* Just a side note. Commenting within defines works only by documenting
+* behind the object (via '!<'). Comments cannot be multi-line and thus
+* we have these extra long lines. What is confusing doxygen over here is
+* that we use '\' and basically have a bunch of typedefs with their
+* documentation in a single line.
+**/
+
+#define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \
+  typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex<float>. */ \
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex<T>, T were corresponding to RealScalar. */ \
+  typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \
+  typedef typename Eigen::internal::nested<Derived>::type Nested; \
+  typedef typename Eigen::internal::traits<Derived>::StorageKind StorageKind; \
+  typedef typename Eigen::internal::traits<Derived>::Index Index; \
+  enum { RowsAtCompileTime = Eigen::internal::traits<Derived>::RowsAtCompileTime, \
+        ColsAtCompileTime = Eigen::internal::traits<Derived>::ColsAtCompileTime, \
+        Flags = Eigen::internal::traits<Derived>::Flags, \
+        CoeffReadCost = Eigen::internal::traits<Derived>::CoeffReadCost, \
+        SizeAtCompileTime = Base::SizeAtCompileTime, \
+        MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \
+        IsVectorAtCompileTime = Base::IsVectorAtCompileTime };
+
+
+#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
+  typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex<float>. */ \
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex<T>, T were corresponding to RealScalar. */ \
+  typedef typename Base::PacketScalar PacketScalar; \
+  typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \
+  typedef typename Eigen::internal::nested<Derived>::type Nested; \
+  typedef typename Eigen::internal::traits<Derived>::StorageKind StorageKind; \
+  typedef typename Eigen::internal::traits<Derived>::Index Index; \
+  enum { RowsAtCompileTime = Eigen::internal::traits<Derived>::RowsAtCompileTime, \
+        ColsAtCompileTime = Eigen::internal::traits<Derived>::ColsAtCompileTime, \
+        MaxRowsAtCompileTime = Eigen::internal::traits<Derived>::MaxRowsAtCompileTime, \
+        MaxColsAtCompileTime = Eigen::internal::traits<Derived>::MaxColsAtCompileTime, \
+        Flags = Eigen::internal::traits<Derived>::Flags, \
+        CoeffReadCost = Eigen::internal::traits<Derived>::CoeffReadCost, \
+        SizeAtCompileTime = Base::SizeAtCompileTime, \
+        MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \
+        IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \
+  using Base::derived; \
+  using Base::const_cast_derived;
+
+
+#define EIGEN_PLAIN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b)
+#define EIGEN_PLAIN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b)
+
+// EIGEN_SIZE_MIN_PREFER_DYNAMIC gives the min between compile-time sizes. 0 has absolute priority, followed by 1,
+// followed by Dynamic, followed by other finite values. The reason for giving Dynamic the priority over
+// finite values is that min(3, Dynamic) should be Dynamic, since that could be anything between 0 and 3.
+#define EIGEN_SIZE_MIN_PREFER_DYNAMIC(a,b) (((int)a == 0 || (int)b == 0) ? 0 \
+                           : ((int)a == 1 || (int)b == 1) ? 1 \
+                           : ((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \
+                           : ((int)a <= (int)b) ? (int)a : (int)b)
+
+// EIGEN_SIZE_MIN_PREFER_FIXED is a variant of EIGEN_SIZE_MIN_PREFER_DYNAMIC comparing MaxSizes. The difference is that finite values
+// now have priority over Dynamic, so that min(3, Dynamic) gives 3. Indeed, whatever the actual value is
+// (between 0 and 3), it is not more than 3.
+#define EIGEN_SIZE_MIN_PREFER_FIXED(a,b)  (((int)a == 0 || (int)b == 0) ? 0 \
+                           : ((int)a == 1 || (int)b == 1) ? 1 \
+                           : ((int)a == Dynamic && (int)b == Dynamic) ? Dynamic \
+                           : ((int)a == Dynamic) ? (int)b \
+                           : ((int)b == Dynamic) ? (int)a \
+                           : ((int)a <= (int)b) ? (int)a : (int)b)
+
+// see EIGEN_SIZE_MIN_PREFER_DYNAMIC. No need for a separate variant for MaxSizes here.
+#define EIGEN_SIZE_MAX(a,b) (((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \
+                           : ((int)a >= (int)b) ? (int)a : (int)b)
+
+#define EIGEN_ADD_COST(a,b) int(a)==Dynamic || int(b)==Dynamic ? Dynamic : int(a)+int(b)
+
+#define EIGEN_LOGICAL_XOR(a,b) (((a) || (b)) && !((a) && (b)))
+
+#define EIGEN_IMPLIES(a,b) (!(a) || (b))
+
+#define EIGEN_MAKE_CWISE_BINARY_OP(METHOD,FUNCTOR) \
+  template<typename OtherDerived> \
+  EIGEN_STRONG_INLINE const CwiseBinaryOp<FUNCTOR<Scalar>, const Derived, const OtherDerived> \
+  (METHOD)(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const \
+  { \
+    return CwiseBinaryOp<FUNCTOR<Scalar>, const Derived, const OtherDerived>(derived(), other.derived()); \
+  }
+
+// the expression type of a cwise product
+#define EIGEN_CWISE_PRODUCT_RETURN_TYPE(LHS,RHS) \
+    CwiseBinaryOp< \
+      internal::scalar_product_op< \
+          typename internal::traits<LHS>::Scalar, \
+          typename internal::traits<RHS>::Scalar \
+      >, \
+      const LHS, \
+      const RHS \
+    >
+
+#endif // EIGEN_MACROS_H

data/ext/eigen/eigen3/Eigen/src/Core/util/Memory.h

@@ -0,0 +1,977 @@
+// 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) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
+// Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
+// Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
+//
+// 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/.
+
+
+/*****************************************************************************
+*** Platform checks for aligned malloc functions                           ***
+*****************************************************************************/
+
+#ifndef EIGEN_MEMORY_H
+#define EIGEN_MEMORY_H
+
+#ifndef EIGEN_MALLOC_ALREADY_ALIGNED
+
+// Try to determine automatically if malloc is already aligned.
+
+// On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
+//   http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
+// This is true at least since glibc 2.8.
+// This leaves the question how to detect 64-bit. According to this document,
+//   http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
+// page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
+// quite safe, at least within the context of glibc, to equate 64-bit with LP64.
+#if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
+ && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ )
+  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
+#else
+  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
+#endif
+
+// FreeBSD 6 seems to have 16-byte aligned malloc
+//   See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
+// FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
+//   See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
+#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
+  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
+#else
+  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
+#endif
+
+#if defined(__APPLE__) \
+ || defined(_WIN64) \
+ || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
+ || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
+  #define EIGEN_MALLOC_ALREADY_ALIGNED 1
+#else
+  #define EIGEN_MALLOC_ALREADY_ALIGNED 0
+#endif
+
+#endif
+
+// See bug 554 (http://eigen.tuxfamily.org/bz/show_bug.cgi?id=554)
+// It seems to be unsafe to check _POSIX_ADVISORY_INFO without including unistd.h first.
+// Currently, let's include it only on unix systems:
+#if defined(__unix__) || defined(__unix)
+  #include <unistd.h>
+  #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || (defined __PGI) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
+    #define EIGEN_HAS_POSIX_MEMALIGN 1
+  #endif
+#endif
+
+#ifndef EIGEN_HAS_POSIX_MEMALIGN
+  #define EIGEN_HAS_POSIX_MEMALIGN 0
+#endif
+
+#ifdef EIGEN_VECTORIZE_SSE
+  #define EIGEN_HAS_MM_MALLOC 1
+#else
+  #define EIGEN_HAS_MM_MALLOC 0
+#endif
+
+namespace Eigen {
+
+namespace internal {
+
+inline void throw_std_bad_alloc()
+{
+  #ifdef EIGEN_EXCEPTIONS
+    throw std::bad_alloc();
+  #else
+    std::size_t huge = -1;
+    new int[huge];
+  #endif
+}
+
+/*****************************************************************************
+*** Implementation of handmade aligned functions                           ***
+*****************************************************************************/
+
+/* ----- Hand made implementations of aligned malloc/free and realloc ----- */
+
+/** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
+  * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
+  */
+inline void* handmade_aligned_malloc(std::size_t size)
+{
+  void *original = std::malloc(size+16);
+  if (original == 0) return 0;
+  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
+  *(reinterpret_cast<void**>(aligned) - 1) = original;
+  return aligned;
+}
+
+/** \internal Frees memory allocated with handmade_aligned_malloc */
+inline void handmade_aligned_free(void *ptr)
+{
+  if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
+}
+
+/** \internal
+  * \brief Reallocates aligned memory.
+  * Since we know that our handmade version is based on std::realloc
+  * we can use std::realloc to implement efficient reallocation.
+  */
+inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
+{
+  if (ptr == 0) return handmade_aligned_malloc(size);
+  void *original = *(reinterpret_cast<void**>(ptr) - 1);
+  std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
+  original = std::realloc(original,size+16);
+  if (original == 0) return 0;
+  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
+  void *previous_aligned = static_cast<char *>(original)+previous_offset;
+  if(aligned!=previous_aligned)
+    std::memmove(aligned, previous_aligned, size);
+  
+  *(reinterpret_cast<void**>(aligned) - 1) = original;
+  return aligned;
+}
+
+/*****************************************************************************
+*** Implementation of generic aligned realloc (when no realloc can be used)***
+*****************************************************************************/
+
+void* aligned_malloc(std::size_t size);
+void  aligned_free(void *ptr);
+
+/** \internal
+  * \brief Reallocates aligned memory.
+  * Allows reallocation with aligned ptr types. This implementation will
+  * always create a new memory chunk and copy the old data.
+  */
+inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
+{
+  if (ptr==0)
+    return aligned_malloc(size);
+
+  if (size==0)
+  {
+    aligned_free(ptr);
+    return 0;
+  }
+
+  void* newptr = aligned_malloc(size);
+  if (newptr == 0)
+  {
+    #ifdef EIGEN_HAS_ERRNO
+    errno = ENOMEM; // according to the standard
+    #endif
+    return 0;
+  }
+
+  if (ptr != 0)
+  {
+    std::memcpy(newptr, ptr, (std::min)(size,old_size));
+    aligned_free(ptr);
+  }
+
+  return newptr;
+}
+
+/*****************************************************************************
+*** Implementation of portable aligned versions of malloc/free/realloc     ***
+*****************************************************************************/
+
+#ifdef EIGEN_NO_MALLOC
+inline void check_that_malloc_is_allowed()
+{
+  eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
+}
+#elif defined EIGEN_RUNTIME_NO_MALLOC
+inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
+{
+  static bool value = true;
+  if (update == 1)
+    value = new_value;
+  return value;
+}
+inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
+inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
+inline void check_that_malloc_is_allowed()
+{
+  eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
+}
+#else 
+inline void check_that_malloc_is_allowed()
+{}
+#endif
+
+/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment.
+  * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
+  */
+inline void* aligned_malloc(size_t size)
+{
+  check_that_malloc_is_allowed();
+
+  void *result;
+  #if !EIGEN_ALIGN
+    result = std::malloc(size);
+  #elif EIGEN_MALLOC_ALREADY_ALIGNED
+    result = std::malloc(size);
+  #elif EIGEN_HAS_POSIX_MEMALIGN
+    if(posix_memalign(&result, 16, size)) result = 0;
+  #elif EIGEN_HAS_MM_MALLOC
+    result = _mm_malloc(size, 16);
+  #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
+    result = _aligned_malloc(size, 16);
+  #else
+    result = handmade_aligned_malloc(size);
+  #endif
+
+  if(!result && size)
+    throw_std_bad_alloc();
+
+  return result;
+}
+
+/** \internal Frees memory allocated with aligned_malloc. */
+inline void aligned_free(void *ptr)
+{
+  #if !EIGEN_ALIGN
+    std::free(ptr);
+  #elif EIGEN_MALLOC_ALREADY_ALIGNED
+    std::free(ptr);
+  #elif EIGEN_HAS_POSIX_MEMALIGN
+    std::free(ptr);
+  #elif EIGEN_HAS_MM_MALLOC
+    _mm_free(ptr);
+  #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
+    _aligned_free(ptr);
+  #else
+    handmade_aligned_free(ptr);
+  #endif
+}
+
+/**
+* \internal
+* \brief Reallocates an aligned block of memory.
+* \throws std::bad_alloc on allocation failure
+**/
+inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
+{
+  EIGEN_UNUSED_VARIABLE(old_size);
+
+  void *result;
+#if !EIGEN_ALIGN
+  result = std::realloc(ptr,new_size);
+#elif EIGEN_MALLOC_ALREADY_ALIGNED
+  result = std::realloc(ptr,new_size);
+#elif EIGEN_HAS_POSIX_MEMALIGN
+  result = generic_aligned_realloc(ptr,new_size,old_size);
+#elif EIGEN_HAS_MM_MALLOC
+  // The defined(_mm_free) is just here to verify that this MSVC version
+  // implements _mm_malloc/_mm_free based on the corresponding _aligned_
+  // functions. This may not always be the case and we just try to be safe.
+  #if defined(_MSC_VER) && (!defined(_WIN32_WCE)) && defined(_mm_free)
+    result = _aligned_realloc(ptr,new_size,16);
+  #else
+    result = generic_aligned_realloc(ptr,new_size,old_size);
+  #endif
+#elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
+  result = _aligned_realloc(ptr,new_size,16);
+#else
+  result = handmade_aligned_realloc(ptr,new_size,old_size);
+#endif
+
+  if (!result && new_size)
+    throw_std_bad_alloc();
+
+  return result;
+}
+
+/*****************************************************************************
+*** Implementation of conditionally aligned functions                      ***
+*****************************************************************************/
+
+/** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
+  * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
+  */
+template<bool Align> inline void* conditional_aligned_malloc(size_t size)
+{
+  return aligned_malloc(size);
+}
+
+template<> inline void* conditional_aligned_malloc<false>(size_t size)
+{
+  check_that_malloc_is_allowed();
+
+  void *result = std::malloc(size);
+  if(!result && size)
+    throw_std_bad_alloc();
+  return result;
+}
+
+/** \internal Frees memory allocated with conditional_aligned_malloc */
+template<bool Align> inline void conditional_aligned_free(void *ptr)
+{
+  aligned_free(ptr);
+}
+
+template<> inline void conditional_aligned_free<false>(void *ptr)
+{
+  std::free(ptr);
+}
+
+template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
+{
+  return aligned_realloc(ptr, new_size, old_size);
+}
+
+template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
+{
+  return std::realloc(ptr, new_size);
+}
+
+/*****************************************************************************
+*** Construction/destruction of array elements                             ***
+*****************************************************************************/
+
+/** \internal Constructs the elements of an array.
+  * The \a size parameter tells on how many objects to call the constructor of T.
+  */
+template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
+{
+  for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
+  return ptr;
+}
+
+/** \internal Destructs the elements of an array.
+  * The \a size parameters tells on how many objects to call the destructor of T.
+  */
+template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
+{
+  // always destruct an array starting from the end.
+  if(ptr)
+    while(size) ptr[--size].~T();
+}
+
+/*****************************************************************************
+*** Implementation of aligned new/delete-like functions                    ***
+*****************************************************************************/
+
+template<typename T>
+EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
+{
+  if(size > size_t(-1) / sizeof(T))
+    throw_std_bad_alloc();
+}
+
+/** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
+  * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
+  * The default constructor of T is called.
+  */
+template<typename T> inline T* aligned_new(size_t size)
+{
+  check_size_for_overflow<T>(size);
+  T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
+  return construct_elements_of_array(result, size);
+}
+
+template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
+{
+  check_size_for_overflow<T>(size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
+  return construct_elements_of_array(result, size);
+}
+
+/** \internal Deletes objects constructed with aligned_new
+  * The \a size parameters tells on how many objects to call the destructor of T.
+  */
+template<typename T> inline void aligned_delete(T *ptr, size_t size)
+{
+  destruct_elements_of_array<T>(ptr, size);
+  aligned_free(ptr);
+}
+
+/** \internal Deletes objects constructed with conditional_aligned_new
+  * The \a size parameters tells on how many objects to call the destructor of T.
+  */
+template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
+{
+  destruct_elements_of_array<T>(ptr, size);
+  conditional_aligned_free<Align>(ptr);
+}
+
+template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
+{
+  check_size_for_overflow<T>(new_size);
+  check_size_for_overflow<T>(old_size);
+  if(new_size < old_size)
+    destruct_elements_of_array(pts+new_size, old_size-new_size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
+  if(new_size > old_size)
+    construct_elements_of_array(result+old_size, new_size-old_size);
+  return result;
+}
+
+
+template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
+{
+  if(size==0)
+    return 0; // short-cut. Also fixes Bug 884
+  check_size_for_overflow<T>(size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
+  if(NumTraits<T>::RequireInitialization)
+    construct_elements_of_array(result, size);
+  return result;
+}
+
+template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
+{
+  check_size_for_overflow<T>(new_size);
+  check_size_for_overflow<T>(old_size);
+  if(NumTraits<T>::RequireInitialization && (new_size < old_size))
+    destruct_elements_of_array(pts+new_size, old_size-new_size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
+  if(NumTraits<T>::RequireInitialization && (new_size > old_size))
+    construct_elements_of_array(result+old_size, new_size-old_size);
+  return result;
+}
+
+template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
+{
+  if(NumTraits<T>::RequireInitialization)
+    destruct_elements_of_array<T>(ptr, size);
+  conditional_aligned_free<Align>(ptr);
+}
+
+/****************************************************************************/
+
+/** \internal Returns the index of the first element of the array that is well aligned for vectorization.
+  *
+  * \param array the address of the start of the array
+  * \param size the size of the array
+  *
+  * \note If no element of the array is well aligned, the size of the array is returned. Typically,
+  * for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the
+  * packet size for the given scalar type is 1, then everything is considered well-aligned.
+  *
+  * \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a
+  * power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the
+  * other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
+  * example with Scalar=double on certain 32-bit platforms, see bug #79.
+  *
+  * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
+  */
+template<typename Scalar, typename Index>
+static inline Index first_aligned(const Scalar* array, Index size)
+{
+  static const Index PacketSize = packet_traits<Scalar>::size;
+  static const Index PacketAlignedMask = PacketSize-1;
+
+  if(PacketSize==1)
+  {
+    // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
+    // of the array have the same alignment.
+    return 0;
+  }
+  else if(size_t(array) & (sizeof(Scalar)-1))
+  {
+    // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
+    // Consequently, no element of the array is well aligned.
+    return size;
+  }
+  else
+  {
+    return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
+                           & PacketAlignedMask, size);
+  }
+}
+
+/** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
+  */ 
+template<typename Index> 
+inline static Index first_multiple(Index size, Index base)
+{
+  return ((size+base-1)/base)*base;
+}
+
+// std::copy is much slower than memcpy, so let's introduce a smart_copy which
+// use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
+template<typename T, bool UseMemcpy> struct smart_copy_helper;
+
+template<typename T> void smart_copy(const T* start, const T* end, T* target)
+{
+  smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
+}
+
+template<typename T> struct smart_copy_helper<T,true> {
+  static inline void run(const T* start, const T* end, T* target)
+  { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
+};
+
+template<typename T> struct smart_copy_helper<T,false> {
+  static inline void run(const T* start, const T* end, T* target)
+  { std::copy(start, end, target); }
+};
+
+
+/*****************************************************************************
+*** Implementation of runtime stack allocation (falling back to malloc)    ***
+*****************************************************************************/
+
+// you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
+// to the appropriate stack allocation function
+#ifndef EIGEN_ALLOCA
+  #if (defined __linux__) || (defined __APPLE__) || (defined alloca)
+    #define EIGEN_ALLOCA alloca
+  #elif defined(_MSC_VER)
+    #define EIGEN_ALLOCA _alloca
+  #endif
+#endif
+
+// This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
+// at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
+template<typename T> class aligned_stack_memory_handler
+{
+  public:
+    /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
+     * Note that \a ptr can be 0 regardless of the other parameters.
+     * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
+     * In this case, the buffer elements will also be destructed when this handler will be destructed.
+     * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
+     **/
+    aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
+      : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
+    {
+      if(NumTraits<T>::RequireInitialization && m_ptr)
+        Eigen::internal::construct_elements_of_array(m_ptr, size);
+    }
+    ~aligned_stack_memory_handler()
+    {
+      if(NumTraits<T>::RequireInitialization && m_ptr)
+        Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
+      if(m_deallocate)
+        Eigen::internal::aligned_free(m_ptr);
+    }
+  protected:
+    T* m_ptr;
+    size_t m_size;
+    bool m_deallocate;
+};
+
+} // end namespace internal
+
+/** \internal
+  * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
+  * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
+  * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
+  * The allocated buffer is automatically deleted when exiting the scope of this declaration.
+  * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
+  * Here is an example:
+  * \code
+  * {
+  *   ei_declare_aligned_stack_constructed_variable(float,data,size,0);
+  *   // use data[0] to data[size-1]
+  * }
+  * \endcode
+  * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
+  */
+#ifdef EIGEN_ALLOCA
+
+  #if defined(__arm__) || defined(_WIN32)
+    #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
+  #else
+    #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
+  #endif
+
+  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
+    Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
+    TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
+               : reinterpret_cast<TYPE*>( \
+                      (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
+                    : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
+    Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,sizeof(TYPE)*SIZE>EIGEN_STACK_ALLOCATION_LIMIT)
+
+#else
+
+  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
+    Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
+    TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
+    Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
+    
+#endif
+
+
+/*****************************************************************************
+*** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
+*****************************************************************************/
+
+#if EIGEN_ALIGN
+  #ifdef EIGEN_EXCEPTIONS
+    #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
+      void* operator new(size_t size, const std::nothrow_t&) throw() { \
+        try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
+        catch (...) { return 0; } \
+      }
+  #else
+    #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
+      void* operator new(size_t size, const std::nothrow_t&) throw() { \
+        return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
+      }
+  #endif
+
+  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
+      void *operator new(size_t size) { \
+        return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
+      } \
+      void *operator new[](size_t size) { \
+        return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
+      } \
+      void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      /* in-place new and delete. since (at least afaik) there is no actual   */ \
+      /* memory allocated we can safely let the default implementation handle */ \
+      /* this particular case. */ \
+      static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
+      static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \
+      void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
+      void operator delete[](void * memory, void *ptr) throw() { return ::operator delete[](memory,ptr); } \
+      /* nothrow-new (returns zero instead of std::bad_alloc) */ \
+      EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
+      void operator delete(void *ptr, const std::nothrow_t&) throw() { \
+        Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
+      } \
+      typedef void eigen_aligned_operator_new_marker_type;
+#else
+  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
+#endif
+
+#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
+#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
+
+/****************************************************************************/
+
+/** \class aligned_allocator
+* \ingroup Core_Module
+*
+* \brief STL compatible allocator to use with with 16 byte aligned types
+*
+* Example:
+* \code
+* // Matrix4f requires 16 bytes alignment:
+* std::map< int, Matrix4f, std::less<int>, 
+*           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
+* // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
+* std::map< int, Vector3f > my_map_vec3;
+* \endcode
+*
+* \sa \ref TopicStlContainers.
+*/
+template<class T>
+class aligned_allocator
+{
+public:
+    typedef size_t    size_type;
+    typedef std::ptrdiff_t difference_type;
+    typedef T*        pointer;
+    typedef const T*  const_pointer;
+    typedef T&        reference;
+    typedef const T&  const_reference;
+    typedef T         value_type;
+
+    template<class U>
+    struct rebind
+    {
+        typedef aligned_allocator<U> other;
+    };
+
+    pointer address( reference value ) const
+    {
+        return &value;
+    }
+
+    const_pointer address( const_reference value ) const
+    {
+        return &value;
+    }
+
+    aligned_allocator()
+    {
+    }
+
+    aligned_allocator( const aligned_allocator& )
+    {
+    }
+
+    template<class U>
+    aligned_allocator( const aligned_allocator<U>& )
+    {
+    }
+
+    ~aligned_allocator()
+    {
+    }
+
+    size_type max_size() const
+    {
+        return (std::numeric_limits<size_type>::max)();
+    }
+
+    pointer allocate( size_type num, const void* hint = 0 )
+    {
+        EIGEN_UNUSED_VARIABLE(hint);
+        internal::check_size_for_overflow<T>(num);
+        return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
+    }
+
+    void construct( pointer p, const T& value )
+    {
+        ::new( p ) T( value );
+    }
+
+    void destroy( pointer p )
+    {
+        p->~T();
+    }
+
+    void deallocate( pointer p, size_type /*num*/ )
+    {
+        internal::aligned_free( p );
+    }
+
+    bool operator!=(const aligned_allocator<T>& ) const
+    { return false; }
+
+    bool operator==(const aligned_allocator<T>& ) const
+    { return true; }
+};
+
+//---------- Cache sizes ----------
+
+#if !defined(EIGEN_NO_CPUID)
+#  if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
+#    if defined(__PIC__) && defined(__i386__)
+       // Case for x86 with PIC
+#      define EIGEN_CPUID(abcd,func,id) \
+         __asm__ __volatile__ ("xchgl %%ebx, %k1;cpuid; xchgl %%ebx,%k1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id));
+#    elif defined(__PIC__) && defined(__x86_64__)
+       // Case for x64 with PIC. In theory this is only a problem with recent gcc and with medium or large code model, not with the default small code model.
+       // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
+#      define EIGEN_CPUID(abcd,func,id) \
+        __asm__ __volatile__ ("xchg{q}\t{%%}rbx, %q1; cpuid; xchg{q}\t{%%}rbx, %q1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id));
+#    else
+       // Case for x86_64 or x86 w/o PIC
+#      define EIGEN_CPUID(abcd,func,id) \
+         __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
+#    endif
+#  elif defined(_MSC_VER)
+#    if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) )
+#      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
+#    endif
+#  endif
+#endif
+
+namespace internal {
+
+#ifdef EIGEN_CPUID
+
+inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
+{
+  return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
+}
+
+inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
+{
+  int abcd[4];
+  l1 = l2 = l3 = 0;
+  int cache_id = 0;
+  int cache_type = 0;
+  do {
+    abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+    EIGEN_CPUID(abcd,0x4,cache_id);
+    cache_type  = (abcd[0] & 0x0F) >> 0;
+    if(cache_type==1||cache_type==3) // data or unified cache
+    {
+      int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
+      int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
+      int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
+      int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
+      int sets        = (abcd[2]);                    // C[31:0]
+
+      int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
+
+      switch(cache_level)
+      {
+        case 1: l1 = cache_size; break;
+        case 2: l2 = cache_size; break;
+        case 3: l3 = cache_size; break;
+        default: break;
+      }
+    }
+    cache_id++;
+  } while(cache_type>0 && cache_id<16);
+}
+
+inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
+{
+  int abcd[4];
+  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+  l1 = l2 = l3 = 0;
+  EIGEN_CPUID(abcd,0x00000002,0);
+  unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
+  bool check_for_p2_core2 = false;
+  for(int i=0; i<14; ++i)
+  {
+    switch(bytes[i])
+    {
+      case 0x0A: l1 = 8; break;   // 0Ah   data L1 cache, 8 KB, 2 ways, 32 byte lines
+      case 0x0C: l1 = 16; break;  // 0Ch   data L1 cache, 16 KB, 4 ways, 32 byte lines
+      case 0x0E: l1 = 24; break;  // 0Eh   data L1 cache, 24 KB, 6 ways, 64 byte lines
+      case 0x10: l1 = 16; break;  // 10h   data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
+      case 0x15: l1 = 16; break;  // 15h   code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
+      case 0x2C: l1 = 32; break;  // 2Ch   data L1 cache, 32 KB, 8 ways, 64 byte lines
+      case 0x30: l1 = 32; break;  // 30h   code L1 cache, 32 KB, 8 ways, 64 byte lines
+      case 0x60: l1 = 16; break;  // 60h   data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
+      case 0x66: l1 = 8; break;   // 66h   data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
+      case 0x67: l1 = 16; break;  // 67h   data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
+      case 0x68: l1 = 32; break;  // 68h   data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
+      case 0x1A: l2 = 96; break;   // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
+      case 0x22: l3 = 512; break;   // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
+      case 0x23: l3 = 1024; break;   // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x25: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x29: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x39: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
+      case 0x3A: l2 = 192; break;   // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
+      case 0x3B: l2 = 128; break;   // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
+      case 0x3C: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
+      case 0x3D: l2 = 384; break;   // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
+      case 0x3E: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
+      case 0x40: l2 = 0; break;   // no integrated L2 cache (P6 core) or L3 cache (P4 core)
+      case 0x41: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
+      case 0x42: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
+      case 0x43: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
+      case 0x44: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
+      case 0x45: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
+      case 0x46: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
+      case 0x47: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
+      case 0x48: l2 = 3072; break;   // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
+      case 0x49: if(l2!=0) l3 = 4096; else {check_for_p2_core2=true; l3 = l2 = 4096;} break;// code and data L3 cache, 4096 KB, 16 ways, 64 byte lines (P4) or L2 for core2
+      case 0x4A: l3 = 6144; break;   // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
+      case 0x4B: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
+      case 0x4C: l3 = 12288; break;   // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
+      case 0x4D: l3 = 16384; break;   // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
+      case 0x4E: l2 = 6144; break;   // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
+      case 0x78: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
+      case 0x79: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7A: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7B: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7C: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7D: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
+      case 0x7E: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
+      case 0x7F: l2 = 512; break;   // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
+      case 0x80: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
+      case 0x81: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
+      case 0x82: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
+      case 0x83: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
+      case 0x84: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
+      case 0x85: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
+      case 0x86: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
+      case 0x87: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
+      case 0x88: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
+      case 0x89: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
+      case 0x8A: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
+      case 0x8D: l3 = 3072; break;   // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
+
+      default: break;
+    }
+  }
+  if(check_for_p2_core2 && l2 == l3)
+    l3 = 0;
+  l1 *= 1024;
+  l2 *= 1024;
+  l3 *= 1024;
+}
+
+inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
+{
+  if(max_std_funcs>=4)
+    queryCacheSizes_intel_direct(l1,l2,l3);
+  else
+    queryCacheSizes_intel_codes(l1,l2,l3);
+}
+
+inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
+{
+  int abcd[4];
+  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+  EIGEN_CPUID(abcd,0x80000005,0);
+  l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
+  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+  EIGEN_CPUID(abcd,0x80000006,0);
+  l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
+  l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
+}
+#endif
+
+/** \internal
+ * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
+inline void queryCacheSizes(int& l1, int& l2, int& l3)
+{
+  #ifdef EIGEN_CPUID
+  int abcd[4];
+  const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
+  const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
+  const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
+
+  // identify the CPU vendor
+  EIGEN_CPUID(abcd,0x0,0);
+  int max_std_funcs = abcd[1];
+  if(cpuid_is_vendor(abcd,GenuineIntel))
+    queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
+  else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
+    queryCacheSizes_amd(l1,l2,l3);
+  else
+    // by default let's use Intel's API
+    queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
+
+  // here is the list of other vendors:
+//   ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
+//   ||cpuid_is_vendor(abcd,"CyrixInstead")
+//   ||cpuid_is_vendor(abcd,"CentaurHauls")
+//   ||cpuid_is_vendor(abcd,"GenuineTMx86")
+//   ||cpuid_is_vendor(abcd,"TransmetaCPU")
+//   ||cpuid_is_vendor(abcd,"RiseRiseRise")
+//   ||cpuid_is_vendor(abcd,"Geode by NSC")
+//   ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
+//   ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
+//   ||cpuid_is_vendor(abcd,"NexGenDriven")
+  #else
+  l1 = l2 = l3 = -1;
+  #endif
+}
+
+/** \internal
+ * \returns the size in Bytes of the L1 data cache */
+inline int queryL1CacheSize()
+{
+  int l1(-1), l2, l3;
+  queryCacheSizes(l1,l2,l3);
+  return l1;
+}
+
+/** \internal
+ * \returns the size in Bytes of the L2 or L3 cache if this later is present */
+inline int queryTopLevelCacheSize()
+{
+  int l1, l2(-1), l3(-1);
+  queryCacheSizes(l1,l2,l3);
+  return (std::max)(l2,l3);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MEMORY_H