Core/util/Memory.h

0001 // This file is part of Eigen, a lightweight C++ template library
0002 // for linear algebra.
0003 //
0004 // Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
0005 // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
0006 // Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
0007 // Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
0008 // Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
0009 // Copyright (C) 2013 Pavel Holoborodko <pavel@holoborodko.com>
0010 //
0011 // This Source Code Form is subject to the terms of the Mozilla
0012 // Public License v. 2.0. If a copy of the MPL was not distributed
0013 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
0014
0015
0016 /*****************************************************************************
0017 *** Platform checks for aligned malloc functions                           ***
0018 *****************************************************************************/
0019
0020 #ifndef EIGEN_MEMORY_H
0021 #define EIGEN_MEMORY_H
0022
0023 #ifndef EIGEN_MALLOC_ALREADY_ALIGNED
0024
0025 // Try to determine automatically if malloc is already aligned.
0026
0027 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
0028 //   http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
0029 // This is true at least since glibc 2.8.
0030 // This leaves the question how to detect 64-bit. According to this document,
0031 //   http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
0032 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
0033 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
0034 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
0035  && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
0036   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
0037 #else
0038   #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
0039 #endif
0040
0041 // FreeBSD 6 seems to have 16-byte aligned malloc
0042 //   See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
0043 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
0044 //   See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
0045 #if defined(__FreeBSD__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
0046   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
0047 #else
0048   #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
0049 #endif
0050
0051 #if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16))     \
0052  || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16))   \
0053  || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED              \
0054  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
0055   #define EIGEN_MALLOC_ALREADY_ALIGNED 1
0056 #else
0057   #define EIGEN_MALLOC_ALREADY_ALIGNED 0
0058 #endif
0059
0060 #endif
0061
0062 namespace Eigen {
0063
0064 namespace internal {
0065
0066 EIGEN_DEVICE_FUNC
0067 inline void throw_std_bad_alloc()
0068 {
0069   #ifdef EIGEN_EXCEPTIONS
0070     throw std::bad_alloc();
0071   #else
0072     std::size_t huge = static_cast<std::size_t>(-1);
0073     #if defined(EIGEN_HIPCC)
0074     //
0075     // calls to "::operator new" are to be treated as opaque function calls (i.e no inlining),
0076     // and as a consequence the code in the #else block triggers the hipcc warning :
0077     // "no overloaded function has restriction specifiers that are compatible with the ambient context"
0078     //
0079     // "throw_std_bad_alloc" has the EIGEN_DEVICE_FUNC attribute, so it seems that hipcc expects
0080     // the same on "operator new"
0081     // Reverting code back to the old version in this #if block for the hipcc compiler
0082     //
0083     new int[huge];
0084     #else
0085     void* unused = ::operator new(huge);
0086     EIGEN_UNUSED_VARIABLE(unused);
0087     #endif
0088   #endif
0089 }
0090
0091 /*****************************************************************************
0092 *** Implementation of handmade aligned functions                           ***
0093 *****************************************************************************/
0094
0095 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
0096
0097 /** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
0098   * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
0099   */
0100 EIGEN_DEVICE_FUNC inline void* handmade_aligned_malloc(std::size_t size, std::size_t alignment = EIGEN_DEFAULT_ALIGN_BYTES)
0101 {
0102   eigen_assert(alignment >= sizeof(void*) && (alignment & (alignment-1)) == 0 && "Alignment must be at least sizeof(void*) and a power of 2");
0103
0104   EIGEN_USING_STD(malloc)
0105   void *original = malloc(size+alignment);
0106
0107   if (original == 0) return 0;
0108   void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(alignment-1))) + alignment);
0109   *(reinterpret_cast<void**>(aligned) - 1) = original;
0110   return aligned;
0111 }
0112
0113 /** \internal Frees memory allocated with handmade_aligned_malloc */
0114 EIGEN_DEVICE_FUNC inline void handmade_aligned_free(void *ptr)
0115 {
0116   if (ptr) {
0117     EIGEN_USING_STD(free)
0118     free(*(reinterpret_cast<void**>(ptr) - 1));
0119   }
0120 }
0121
0122 /** \internal
0123   * \brief Reallocates aligned memory.
0124   * Since we know that our handmade version is based on std::malloc
0125   * we can use std::realloc to implement efficient reallocation.
0126   */
0127 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
0128 {
0129   if (ptr == 0) return handmade_aligned_malloc(size);
0130   void *original = *(reinterpret_cast<void**>(ptr) - 1);
0131   std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
0132   original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES);
0133   if (original == 0) return 0;
0134   void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
0135   void *previous_aligned = static_cast<char *>(original)+previous_offset;
0136   if(aligned!=previous_aligned)
0137     std::memmove(aligned, previous_aligned, size);
0138
0139   *(reinterpret_cast<void**>(aligned) - 1) = original;
0140   return aligned;
0141 }
0142
0143 /*****************************************************************************
0144 *** Implementation of portable aligned versions of malloc/free/realloc     ***
0145 *****************************************************************************/
0146
0147 #ifdef EIGEN_NO_MALLOC
0148 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
0149 {
0150   eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
0151 }
0152 #elif defined EIGEN_RUNTIME_NO_MALLOC
0153 EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
0154 {
0155   static bool value = true;
0156   if (update == 1)
0157     value = new_value;
0158   return value;
0159 }
0160 EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
0161 EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
0162 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
0163 {
0164   eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
0165 }
0166 #else
0167 EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
0168 {}
0169 #endif
0170
0171 /** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements.
0172   * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
0173   */
0174 EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
0175 {
0176   check_that_malloc_is_allowed();
0177
0178   void *result;
0179   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
0180
0181     EIGEN_USING_STD(malloc)
0182     result = malloc(size);
0183
0184     #if EIGEN_DEFAULT_ALIGN_BYTES==16
0185     eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade aligned memory allocator.");
0186     #endif
0187   #else
0188     result = handmade_aligned_malloc(size);
0189   #endif
0190
0191   if(!result && size)
0192     throw_std_bad_alloc();
0193
0194   return result;
0195 }
0196
0197 /** \internal Frees memory allocated with aligned_malloc. */
0198 EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
0199 {
0200   #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
0201
0202     EIGEN_USING_STD(free)
0203     free(ptr);
0204
0205   #else
0206     handmade_aligned_free(ptr);
0207   #endif
0208 }
0209
0210 /**
0211   * \internal
0212   * \brief Reallocates an aligned block of memory.
0213   * \throws std::bad_alloc on allocation failure
0214   */
0215 inline void* aligned_realloc(void *ptr, std::size_t new_size, std::size_t old_size)
0216 {
0217   EIGEN_UNUSED_VARIABLE(old_size)
0218
0219   void *result;
0220 #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
0221   result = std::realloc(ptr,new_size);
0222 #else
0223   result = handmade_aligned_realloc(ptr,new_size,old_size);
0224 #endif
0225
0226   if (!result && new_size)
0227     throw_std_bad_alloc();
0228
0229   return result;
0230 }
0231
0232 /*****************************************************************************
0233 *** Implementation of conditionally aligned functions                      ***
0234 *****************************************************************************/
0235
0236 /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
0237   * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
0238   */
0239 template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size)
0240 {
0241   return aligned_malloc(size);
0242 }
0243
0244 template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size)
0245 {
0246   check_that_malloc_is_allowed();
0247
0248   EIGEN_USING_STD(malloc)
0249   void *result = malloc(size);
0250
0251   if(!result && size)
0252     throw_std_bad_alloc();
0253   return result;
0254 }
0255
0256 /** \internal Frees memory allocated with conditional_aligned_malloc */
0257 template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr)
0258 {
0259   aligned_free(ptr);
0260 }
0261
0262 template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
0263 {
0264   EIGEN_USING_STD(free)
0265   free(ptr);
0266 }
0267
0268 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size)
0269 {
0270   return aligned_realloc(ptr, new_size, old_size);
0271 }
0272
0273 template<> inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size, std::size_t)
0274 {
0275   return std::realloc(ptr, new_size);
0276 }
0277
0278 /*****************************************************************************
0279 *** Construction/destruction of array elements                             ***
0280 *****************************************************************************/
0281
0282 /** \internal Destructs the elements of an array.
0283   * The \a size parameters tells on how many objects to call the destructor of T.
0284   */
0285 template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, std::size_t size)
0286 {
0287   // always destruct an array starting from the end.
0288   if(ptr)
0289     while(size) ptr[--size].~T();
0290 }
0291
0292 /** \internal Constructs the elements of an array.
0293   * The \a size parameter tells on how many objects to call the constructor of T.
0294   */
0295 template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, std::size_t size)
0296 {
0297   std::size_t i;
0298   EIGEN_TRY
0299   {
0300       for (i = 0; i < size; ++i) ::new (ptr + i) T;
0301       return ptr;
0302   }
0303   EIGEN_CATCH(...)
0304   {
0305     destruct_elements_of_array(ptr, i);
0306     EIGEN_THROW;
0307   }
0308   return NULL;
0309 }
0310
0311 /*****************************************************************************
0312 *** Implementation of aligned new/delete-like functions                    ***
0313 *****************************************************************************/
0314
0315 template<typename T>
0316 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size)
0317 {
0318   if(size > std::size_t(-1) / sizeof(T))
0319     throw_std_bad_alloc();
0320 }
0321
0322 /** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
0323   * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
0324   * The default constructor of T is called.
0325   */
0326 template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size)
0327 {
0328   check_size_for_overflow<T>(size);
0329   T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
0330   EIGEN_TRY
0331   {
0332     return construct_elements_of_array(result, size);
0333   }
0334   EIGEN_CATCH(...)
0335   {
0336     aligned_free(result);
0337     EIGEN_THROW;
0338   }
0339   return result;
0340 }
0341
0342 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size)
0343 {
0344   check_size_for_overflow<T>(size);
0345   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
0346   EIGEN_TRY
0347   {
0348     return construct_elements_of_array(result, size);
0349   }
0350   EIGEN_CATCH(...)
0351   {
0352     conditional_aligned_free<Align>(result);
0353     EIGEN_THROW;
0354   }
0355   return result;
0356 }
0357
0358 /** \internal Deletes objects constructed with aligned_new
0359   * The \a size parameters tells on how many objects to call the destructor of T.
0360   */
0361 template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, std::size_t size)
0362 {
0363   destruct_elements_of_array<T>(ptr, size);
0364   Eigen::internal::aligned_free(ptr);
0365 }
0366
0367 /** \internal Deletes objects constructed with conditional_aligned_new
0368   * The \a size parameters tells on how many objects to call the destructor of T.
0369   */
0370 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, std::size_t size)
0371 {
0372   destruct_elements_of_array<T>(ptr, size);
0373   conditional_aligned_free<Align>(ptr);
0374 }
0375
0376 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size)
0377 {
0378   check_size_for_overflow<T>(new_size);
0379   check_size_for_overflow<T>(old_size);
0380   if(new_size < old_size)
0381     destruct_elements_of_array(pts+new_size, old_size-new_size);
0382   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
0383   if(new_size > old_size)
0384   {
0385     EIGEN_TRY
0386     {
0387       construct_elements_of_array(result+old_size, new_size-old_size);
0388     }
0389     EIGEN_CATCH(...)
0390     {
0391       conditional_aligned_free<Align>(result);
0392       EIGEN_THROW;
0393     }
0394   }
0395   return result;
0396 }
0397
0398
0399 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size)
0400 {
0401   if(size==0)
0402     return 0; // short-cut. Also fixes Bug 884
0403   check_size_for_overflow<T>(size);
0404   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
0405   if(NumTraits<T>::RequireInitialization)
0406   {
0407     EIGEN_TRY
0408     {
0409       construct_elements_of_array(result, size);
0410     }
0411     EIGEN_CATCH(...)
0412     {
0413       conditional_aligned_free<Align>(result);
0414       EIGEN_THROW;
0415     }
0416   }
0417   return result;
0418 }
0419
0420 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size)
0421 {
0422   check_size_for_overflow<T>(new_size);
0423   check_size_for_overflow<T>(old_size);
0424   if(NumTraits<T>::RequireInitialization && (new_size < old_size))
0425     destruct_elements_of_array(pts+new_size, old_size-new_size);
0426   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
0427   if(NumTraits<T>::RequireInitialization && (new_size > old_size))
0428   {
0429     EIGEN_TRY
0430     {
0431       construct_elements_of_array(result+old_size, new_size-old_size);
0432     }
0433     EIGEN_CATCH(...)
0434     {
0435       conditional_aligned_free<Align>(result);
0436       EIGEN_THROW;
0437     }
0438   }
0439   return result;
0440 }
0441
0442 template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, std::size_t size)
0443 {
0444   if(NumTraits<T>::RequireInitialization)
0445     destruct_elements_of_array<T>(ptr, size);
0446   conditional_aligned_free<Align>(ptr);
0447 }
0448
0449 /****************************************************************************/
0450
0451 /** \internal Returns the index of the first element of the array that is well aligned with respect to the requested \a Alignment.
0452   *
0453   * \tparam Alignment requested alignment in Bytes.
0454   * \param array the address of the start of the array
0455   * \param size the size of the array
0456   *
0457   * \note If no element of the array is well aligned or the requested alignment is not a multiple of a scalar,
0458   * the size of the array is returned. For example with SSE, the requested alignment is typically 16-bytes. If
0459   * packet size for the given scalar type is 1, then everything is considered well-aligned.
0460   *
0461   * \note Otherwise, if the Alignment is larger that the scalar size, we rely on the assumptions that sizeof(Scalar) is a
0462   * power of 2. On the other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
0463   * example with Scalar=double on certain 32-bit platforms, see bug #79.
0464   *
0465   * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
0466   * \sa first_default_aligned()
0467   */
0468 template<int Alignment, typename Scalar, typename Index>
0469 EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size)
0470 {
0471   const Index ScalarSize = sizeof(Scalar);
0472   const Index AlignmentSize = Alignment / ScalarSize;
0473   const Index AlignmentMask = AlignmentSize-1;
0474
0475   if(AlignmentSize<=1)
0476   {
0477     // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar
0478     // so that all elements of the array have the same alignment.
0479     return 0;
0480   }
0481   else if( (UIntPtr(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0)
0482   {
0483     // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size.
0484     // Consequently, no element of the array is well aligned.
0485     return size;
0486   }
0487   else
0488   {
0489     Index first = (AlignmentSize - (Index((UIntPtr(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask;
0490     return (first < size) ? first : size;
0491   }
0492 }
0493
0494 /** \internal Returns the index of the first element of the array that is well aligned with respect the largest packet requirement.
0495    * \sa first_aligned(Scalar*,Index) and first_default_aligned(DenseBase<Derived>) */
0496 template<typename Scalar, typename Index>
0497 EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size)
0498 {
0499   typedef typename packet_traits<Scalar>::type DefaultPacketType;
0500   return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size);
0501 }
0502
0503 /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
0504   */
0505 template<typename Index>
0506 inline Index first_multiple(Index size, Index base)
0507 {
0508   return ((size+base-1)/base)*base;
0509 }
0510
0511 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
0512 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
0513 template<typename T, bool UseMemcpy> struct smart_copy_helper;
0514
0515 template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target)
0516 {
0517   smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
0518 }
0519
0520 template<typename T> struct smart_copy_helper<T,true> {
0521   EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
0522   {
0523     IntPtr size = IntPtr(end)-IntPtr(start);
0524     if(size==0) return;
0525     eigen_internal_assert(start!=0 && end!=0 && target!=0);
0526     EIGEN_USING_STD(memcpy)
0527     memcpy(target, start, size);
0528   }
0529 };
0530
0531 template<typename T> struct smart_copy_helper<T,false> {
0532   EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
0533   { std::copy(start, end, target); }
0534 };
0535
0536 // intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise.
0537 template<typename T, bool UseMemmove> struct smart_memmove_helper;
0538
0539 template<typename T> void smart_memmove(const T* start, const T* end, T* target)
0540 {
0541   smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
0542 }
0543
0544 template<typename T> struct smart_memmove_helper<T,true> {
0545   static inline void run(const T* start, const T* end, T* target)
0546   {
0547     IntPtr size = IntPtr(end)-IntPtr(start);
0548     if(size==0) return;
0549     eigen_internal_assert(start!=0 && end!=0 && target!=0);
0550     std::memmove(target, start, size);
0551   }
0552 };
0553
0554 template<typename T> struct smart_memmove_helper<T,false> {
0555   static inline void run(const T* start, const T* end, T* target)
0556   {
0557     if (UIntPtr(target) < UIntPtr(start))
0558     {
0559       std::copy(start, end, target);
0560     }
0561     else
0562     {
0563       std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T);
0564       std::copy_backward(start, end, target + count);
0565     }
0566   }
0567 };
0568
0569 #if EIGEN_HAS_RVALUE_REFERENCES
0570 template<typename T> EIGEN_DEVICE_FUNC T* smart_move(T* start, T* end, T* target)
0571 {
0572   return std::move(start, end, target);
0573 }
0574 #else
0575 template<typename T> EIGEN_DEVICE_FUNC T* smart_move(T* start, T* end, T* target)
0576 {
0577   return std::copy(start, end, target);
0578 }
0579 #endif
0580
0581 /*****************************************************************************
0582 *** Implementation of runtime stack allocation (falling back to malloc)    ***
0583 *****************************************************************************/
0584
0585 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
0586 // to the appropriate stack allocation function
0587 #if ! defined EIGEN_ALLOCA && ! defined EIGEN_GPU_COMPILE_PHASE
0588   #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
0589     #define EIGEN_ALLOCA alloca
0590   #elif EIGEN_COMP_MSVC
0591     #define EIGEN_ALLOCA _alloca
0592   #endif
0593 #endif
0594
0595 // With clang -Oz -mthumb, alloca changes the stack pointer in a way that is
0596 // not allowed in Thumb2. -DEIGEN_STACK_ALLOCATION_LIMIT=0 doesn't work because
0597 // the compiler still emits bad code because stack allocation checks use "<=".
0598 // TODO: Eliminate after https://bugs.llvm.org/show_bug.cgi?id=23772
0599 // is fixed.
0600 #if defined(__clang__) && defined(__thumb__)
0601   #undef EIGEN_ALLOCA
0602 #endif
0603
0604 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
0605 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
0606 template<typename T> class aligned_stack_memory_handler : noncopyable
0607 {
0608   public:
0609     /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
0610      * Note that \a ptr can be 0 regardless of the other parameters.
0611      * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
0612      * In this case, the buffer elements will also be destructed when this handler will be destructed.
0613      * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
0614      **/
0615     EIGEN_DEVICE_FUNC
0616     aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc)
0617       : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
0618     {
0619       if(NumTraits<T>::RequireInitialization && m_ptr)
0620         Eigen::internal::construct_elements_of_array(m_ptr, size);
0621     }
0622     EIGEN_DEVICE_FUNC
0623     ~aligned_stack_memory_handler()
0624     {
0625       if(NumTraits<T>::RequireInitialization && m_ptr)
0626         Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
0627       if(m_deallocate)
0628         Eigen::internal::aligned_free(m_ptr);
0629     }
0630   protected:
0631     T* m_ptr;
0632     std::size_t m_size;
0633     bool m_deallocate;
0634 };
0635
0636 #ifdef EIGEN_ALLOCA
0637
0638 template<typename Xpr, int NbEvaluations,
0639          bool MapExternalBuffer = nested_eval<Xpr,NbEvaluations>::Evaluate && Xpr::MaxSizeAtCompileTime==Dynamic
0640          >
0641 struct local_nested_eval_wrapper
0642 {
0643   static const bool NeedExternalBuffer = false;
0644   typedef typename Xpr::Scalar Scalar;
0645   typedef typename nested_eval<Xpr,NbEvaluations>::type ObjectType;
0646   ObjectType object;
0647
0648   EIGEN_DEVICE_FUNC
0649   local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr) : object(xpr)
0650   {
0651     EIGEN_UNUSED_VARIABLE(ptr);
0652     eigen_internal_assert(ptr==0);
0653   }
0654 };
0655
0656 template<typename Xpr, int NbEvaluations>
0657 struct local_nested_eval_wrapper<Xpr,NbEvaluations,true>
0658 {
0659   static const bool NeedExternalBuffer = true;
0660   typedef typename Xpr::Scalar Scalar;
0661   typedef typename plain_object_eval<Xpr>::type PlainObject;
0662   typedef Map<PlainObject,EIGEN_DEFAULT_ALIGN_BYTES> ObjectType;
0663   ObjectType object;
0664
0665   EIGEN_DEVICE_FUNC
0666   local_nested_eval_wrapper(const Xpr& xpr, Scalar* ptr)
0667     : object(ptr==0 ? reinterpret_cast<Scalar*>(Eigen::internal::aligned_malloc(sizeof(Scalar)*xpr.size())) : ptr, xpr.rows(), xpr.cols()),
0668       m_deallocate(ptr==0)
0669   {
0670     if(NumTraits<Scalar>::RequireInitialization && object.data())
0671       Eigen::internal::construct_elements_of_array(object.data(), object.size());
0672     object = xpr;
0673   }
0674
0675   EIGEN_DEVICE_FUNC
0676   ~local_nested_eval_wrapper()
0677   {
0678     if(NumTraits<Scalar>::RequireInitialization && object.data())
0679       Eigen::internal::destruct_elements_of_array(object.data(), object.size());
0680     if(m_deallocate)
0681       Eigen::internal::aligned_free(object.data());
0682   }
0683
0684 private:
0685   bool m_deallocate;
0686 };
0687
0688 #endif // EIGEN_ALLOCA
0689
0690 template<typename T> class scoped_array : noncopyable
0691 {
0692   T* m_ptr;
0693 public:
0694   explicit scoped_array(std::ptrdiff_t size)
0695   {
0696     m_ptr = new T[size];
0697   }
0698   ~scoped_array()
0699   {
0700     delete[] m_ptr;
0701   }
0702   T& operator[](std::ptrdiff_t i) { return m_ptr[i]; }
0703   const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; }
0704   T* &ptr() { return m_ptr; }
0705   const T* ptr() const { return m_ptr; }
0706   operator const T*() const { return m_ptr; }
0707 };
0708
0709 template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
0710 {
0711   std::swap(a.ptr(),b.ptr());
0712 }
0713
0714 } // end namespace internal
0715
0716 /** \internal
0717   *
0718   * The macro ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) declares, allocates,
0719   * and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
0720   * if the size in bytes is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
0721   * (currently, this is Linux, OSX and Visual Studio only). Otherwise the memory is allocated on the heap.
0722   * The allocated buffer is automatically deleted when exiting the scope of this declaration.
0723   * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
0724   * Here is an example:
0725   * \code
0726   * {
0727   *   ei_declare_aligned_stack_constructed_variable(float,data,size,0);
0728   *   // use data[0] to data[size-1]
0729   * }
0730   * \endcode
0731   * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
0732   *
0733   * The macro ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) is analogue to
0734   * \code
0735   *   typename internal::nested_eval<XPRT_T,N>::type NAME(XPR);
0736   * \endcode
0737   * with the advantage of using aligned stack allocation even if the maximal size of XPR at compile time is unknown.
0738   * This is accomplished through alloca if this later is supported and if the required number of bytes
0739   * is below EIGEN_STACK_ALLOCATION_LIMIT.
0740   */
0741 #ifdef EIGEN_ALLOCA
0742
0743   #if EIGEN_DEFAULT_ALIGN_BYTES>0
0744     // We always manually re-align the result of EIGEN_ALLOCA.
0745     // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
0746     #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((internal::UIntPtr(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)))
0747   #else
0748     #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE)
0749   #endif
0750
0751   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
0752     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
0753     TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
0754                : reinterpret_cast<TYPE*>( \
0755                       (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
0756                     : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
0757     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)
0758
0759
0760   #define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) \
0761     Eigen::internal::local_nested_eval_wrapper<XPR_T,N> EIGEN_CAT(NAME,_wrapper)(XPR, reinterpret_cast<typename XPR_T::Scalar*>( \
0762       ( (Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::NeedExternalBuffer) && ((sizeof(typename XPR_T::Scalar)*XPR.size())<=EIGEN_STACK_ALLOCATION_LIMIT) ) \
0763         ? EIGEN_ALIGNED_ALLOCA( sizeof(typename XPR_T::Scalar)*XPR.size() ) : 0 ) ) ; \
0764     typename Eigen::internal::local_nested_eval_wrapper<XPR_T,N>::ObjectType NAME(EIGEN_CAT(NAME,_wrapper).object)
0765
0766 #else
0767
0768   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
0769     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
0770     TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
0771     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
0772
0773
0774 #define ei_declare_local_nested_eval(XPR_T,XPR,N,NAME) typename Eigen::internal::nested_eval<XPR_T,N>::type NAME(XPR)
0775
0776 #endif
0777
0778
0779 /*****************************************************************************
0780 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
0781 *****************************************************************************/
0782
0783 #if EIGEN_HAS_CXX17_OVERALIGN
0784
0785 // C++17 -> no need to bother about alignment anymore :)
0786
0787 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign)
0788 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
0789 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW
0790 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size)
0791
0792 #else
0793
0794 // HIP does not support new/delete on device.
0795 #if EIGEN_MAX_ALIGN_BYTES!=0 && !defined(EIGEN_HIP_DEVICE_COMPILE)
0796   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
0797       EIGEN_DEVICE_FUNC \
0798       void* operator new(std::size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \
0799         EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
0800         EIGEN_CATCH (...) { return 0; } \
0801       }
0802   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
0803       EIGEN_DEVICE_FUNC \
0804       void *operator new(std::size_t size) { \
0805         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
0806       } \
0807       EIGEN_DEVICE_FUNC \
0808       void *operator new[](std::size_t size) { \
0809         return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
0810       } \
0811       EIGEN_DEVICE_FUNC \
0812       void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
0813       EIGEN_DEVICE_FUNC \
0814       void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
0815       EIGEN_DEVICE_FUNC \
0816       void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
0817       EIGEN_DEVICE_FUNC \
0818       void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
0819       /* in-place new and delete. since (at least afaik) there is no actual   */ \
0820       /* memory allocated we can safely let the default implementation handle */ \
0821       /* this particular case. */ \
0822       EIGEN_DEVICE_FUNC \
0823       static void *operator new(std::size_t size, void *ptr) { return ::operator new(size,ptr); } \
0824       EIGEN_DEVICE_FUNC \
0825       static void *operator new[](std::size_t size, void* ptr) { return ::operator new[](size,ptr); } \
0826       EIGEN_DEVICE_FUNC \
0827       void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \
0828       EIGEN_DEVICE_FUNC \
0829       void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \
0830       /* nothrow-new (returns zero instead of std::bad_alloc) */ \
0831       EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
0832       EIGEN_DEVICE_FUNC \
0833       void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \
0834         Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
0835       } \
0836       typedef void eigen_aligned_operator_new_marker_type;
0837 #else
0838   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
0839 #endif
0840
0841 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
0842 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size)                        \
0843   EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(                                                             \
0844         ((Size)!=Eigen::Dynamic) &&                                                                    \
0845         (((EIGEN_MAX_ALIGN_BYTES>=16) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES  )==0)) ||    \
0846          ((EIGEN_MAX_ALIGN_BYTES>=32) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/2)==0)) ||    \
0847          ((EIGEN_MAX_ALIGN_BYTES>=64) && ((sizeof(Scalar)*(Size))%(EIGEN_MAX_ALIGN_BYTES/4)==0))   )))
0848
0849 #endif
0850
0851 /****************************************************************************/
0852
0853 /** \class aligned_allocator
0854 * \ingroup Core_Module
0855 *
0856 * \brief STL compatible allocator to use with types requiring a non standrad alignment.
0857 *
0858 * The memory is aligned as for dynamically aligned matrix/array types such as MatrixXd.
0859 * By default, it will thus provide at least 16 bytes alignment and more in following cases:
0860 *  - 32 bytes alignment if AVX is enabled.
0861 *  - 64 bytes alignment if AVX512 is enabled.
0862 *
0863 * This can be controlled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented
0864 * \link TopicPreprocessorDirectivesPerformance there \endlink.
0865 *
0866 * Example:
0867 * \code
0868 * // Matrix4f requires 16 bytes alignment:
0869 * std::map< int, Matrix4f, std::less<int>,
0870 *           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
0871 * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
0872 * std::map< int, Vector3f > my_map_vec3;
0873 * \endcode
0874 *
0875 * \sa \blank \ref TopicStlContainers.
0876 */
0877 template<class T>
0878 class aligned_allocator : public std::allocator<T>
0879 {
0880 public:
0881   typedef std::size_t     size_type;
0882   typedef std::ptrdiff_t  difference_type;
0883   typedef T*              pointer;
0884   typedef const T*        const_pointer;
0885   typedef T&              reference;
0886   typedef const T&        const_reference;
0887   typedef T               value_type;
0888
0889   template<class U>
0890   struct rebind
0891   {
0892     typedef aligned_allocator<U> other;
0893   };
0894
0895   aligned_allocator() : std::allocator<T>() {}
0896
0897   aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {}
0898
0899   template<class U>
0900   aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {}
0901
0902   ~aligned_allocator() {}
0903
0904   #if EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_LEAST(7,0)
0905   // In gcc std::allocator::max_size() is bugged making gcc triggers a warning:
0906   // eigen/Eigen/src/Core/util/Memory.h:189:12: warning: argument 1 value '18446744073709551612' exceeds maximum object size 9223372036854775807
0907   // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87544
0908   size_type max_size() const {
0909     return (std::numeric_limits<std::ptrdiff_t>::max)()/sizeof(T);
0910   }
0911   #endif
0912
0913   pointer allocate(size_type num, const void* /*hint*/ = 0)
0914   {
0915     internal::check_size_for_overflow<T>(num);
0916     return static_cast<pointer>( internal::aligned_malloc(num * sizeof(T)) );
0917   }
0918
0919   void deallocate(pointer p, size_type /*num*/)
0920   {
0921     internal::aligned_free(p);
0922   }
0923 };
0924
0925 //---------- Cache sizes ----------
0926
0927 #if !defined(EIGEN_NO_CPUID)
0928 #  if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64
0929 #    if defined(__PIC__) && EIGEN_ARCH_i386
0930        // Case for x86 with PIC
0931 #      define EIGEN_CPUID(abcd,func,id) \
0932          __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));
0933 #    elif defined(__PIC__) && EIGEN_ARCH_x86_64
0934        // 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.
0935        // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
0936 #      define EIGEN_CPUID(abcd,func,id) \
0937         __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));
0938 #    else
0939        // Case for x86_64 or x86 w/o PIC
0940 #      define EIGEN_CPUID(abcd,func,id) \
0941          __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
0942 #    endif
0943 #  elif EIGEN_COMP_MSVC
0944 #    if (EIGEN_COMP_MSVC > 1500) && EIGEN_ARCH_i386_OR_x86_64
0945 #      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
0946 #    endif
0947 #  endif
0948 #endif
0949
0950 namespace internal {
0951
0952 #ifdef EIGEN_CPUID
0953
0954 inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
0955 {
0956   return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
0957 }
0958
0959 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
0960 {
0961   int abcd[4];
0962   l1 = l2 = l3 = 0;
0963   int cache_id = 0;
0964   int cache_type = 0;
0965   do {
0966     abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
0967     EIGEN_CPUID(abcd,0x4,cache_id);
0968     cache_type  = (abcd[0] & 0x0F) >> 0;
0969     if(cache_type==1||cache_type==3) // data or unified cache
0970     {
0971       int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
0972       int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
0973       int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
0974       int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
0975       int sets        = (abcd[2]);                    // C[31:0]
0976
0977       int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
0978
0979       switch(cache_level)
0980       {
0981         case 1: l1 = cache_size; break;
0982         case 2: l2 = cache_size; break;
0983         case 3: l3 = cache_size; break;
0984         default: break;
0985       }
0986     }
0987     cache_id++;
0988   } while(cache_type>0 && cache_id<16);
0989 }
0990
0991 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
0992 {
0993   int abcd[4];
0994   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
0995   l1 = l2 = l3 = 0;
0996   EIGEN_CPUID(abcd,0x00000002,0);
0997   unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
0998   bool check_for_p2_core2 = false;
0999   for(int i=0; i<14; ++i)
1000   {
1001     switch(bytes[i])
1002     {
1003       case 0x0A: l1 = 8; break;   // 0Ah   data L1 cache, 8 KB, 2 ways, 32 byte lines
1004       case 0x0C: l1 = 16; break;  // 0Ch   data L1 cache, 16 KB, 4 ways, 32 byte lines
1005       case 0x0E: l1 = 24; break;  // 0Eh   data L1 cache, 24 KB, 6 ways, 64 byte lines
1006       case 0x10: l1 = 16; break;  // 10h   data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
1007       case 0x15: l1 = 16; break;  // 15h   code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
1008       case 0x2C: l1 = 32; break;  // 2Ch   data L1 cache, 32 KB, 8 ways, 64 byte lines
1009       case 0x30: l1 = 32; break;  // 30h   code L1 cache, 32 KB, 8 ways, 64 byte lines
1010       case 0x60: l1 = 16; break;  // 60h   data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
1011       case 0x66: l1 = 8; break;   // 66h   data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
1012       case 0x67: l1 = 16; break;  // 67h   data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
1013       case 0x68: l1 = 32; break;  // 68h   data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
1014       case 0x1A: l2 = 96; break;   // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
1015       case 0x22: l3 = 512; break;   // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
1016       case 0x23: l3 = 1024; break;   // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
1017       case 0x25: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
1018       case 0x29: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
1019       case 0x39: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
1020       case 0x3A: l2 = 192; break;   // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
1021       case 0x3B: l2 = 128; break;   // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
1022       case 0x3C: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
1023       case 0x3D: l2 = 384; break;   // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
1024       case 0x3E: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
1025       case 0x40: l2 = 0; break;   // no integrated L2 cache (P6 core) or L3 cache (P4 core)
1026       case 0x41: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
1027       case 0x42: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
1028       case 0x43: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
1029       case 0x44: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
1030       case 0x45: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
1031       case 0x46: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
1032       case 0x47: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
1033       case 0x48: l2 = 3072; break;   // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
1034       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
1035       case 0x4A: l3 = 6144; break;   // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
1036       case 0x4B: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
1037       case 0x4C: l3 = 12288; break;   // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
1038       case 0x4D: l3 = 16384; break;   // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
1039       case 0x4E: l2 = 6144; break;   // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
1040       case 0x78: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
1041       case 0x79: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
1042       case 0x7A: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
1043       case 0x7B: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
1044       case 0x7C: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
1045       case 0x7D: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
1046       case 0x7E: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
1047       case 0x7F: l2 = 512; break;   // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
1048       case 0x80: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
1049       case 0x81: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
1050       case 0x82: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
1051       case 0x83: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
1052       case 0x84: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
1053       case 0x85: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
1054       case 0x86: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
1055       case 0x87: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
1056       case 0x88: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
1057       case 0x89: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
1058       case 0x8A: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
1059       case 0x8D: l3 = 3072; break;   // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
1060
1061       default: break;
1062     }
1063   }
1064   if(check_for_p2_core2 && l2 == l3)
1065     l3 = 0;
1066   l1 *= 1024;
1067   l2 *= 1024;
1068   l3 *= 1024;
1069 }
1070
1071 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
1072 {
1073   if(max_std_funcs>=4)
1074     queryCacheSizes_intel_direct(l1,l2,l3);
1075   else if(max_std_funcs>=2)
1076     queryCacheSizes_intel_codes(l1,l2,l3);
1077   else
1078     l1 = l2 = l3 = 0;
1079 }
1080
1081 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
1082 {
1083   int abcd[4];
1084   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1085
1086   // First query the max supported function.
1087   EIGEN_CPUID(abcd,0x80000000,0);
1088   if(static_cast<numext::uint32_t>(abcd[0]) >= static_cast<numext::uint32_t>(0x80000006))
1089   {
1090     EIGEN_CPUID(abcd,0x80000005,0);
1091     l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
1092     abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
1093     EIGEN_CPUID(abcd,0x80000006,0);
1094     l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
1095     l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
1096   }
1097   else
1098   {
1099     l1 = l2 = l3 = 0;
1100   }
1101 }
1102 #endif
1103
1104 /** \internal
1105  * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
1106 inline void queryCacheSizes(int& l1, int& l2, int& l3)
1107 {
1108   #ifdef EIGEN_CPUID
1109   int abcd[4];
1110   const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
1111   const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
1112   const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
1113
1114   // identify the CPU vendor
1115   EIGEN_CPUID(abcd,0x0,0);
1116   int max_std_funcs = abcd[0];
1117   if(cpuid_is_vendor(abcd,GenuineIntel))
1118     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
1119   else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
1120     queryCacheSizes_amd(l1,l2,l3);
1121   else
1122     // by default let's use Intel's API
1123     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
1124
1125   // here is the list of other vendors:
1126 //   ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
1127 //   ||cpuid_is_vendor(abcd,"CyrixInstead")
1128 //   ||cpuid_is_vendor(abcd,"CentaurHauls")
1129 //   ||cpuid_is_vendor(abcd,"GenuineTMx86")
1130 //   ||cpuid_is_vendor(abcd,"TransmetaCPU")
1131 //   ||cpuid_is_vendor(abcd,"RiseRiseRise")
1132 //   ||cpuid_is_vendor(abcd,"Geode by NSC")
1133 //   ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
1134 //   ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
1135 //   ||cpuid_is_vendor(abcd,"NexGenDriven")
1136   #else
1137   l1 = l2 = l3 = -1;
1138   #endif
1139 }
1140
1141 /** \internal
1142  * \returns the size in Bytes of the L1 data cache */
1143 inline int queryL1CacheSize()
1144 {
1145   int l1(-1), l2, l3;
1146   queryCacheSizes(l1,l2,l3);
1147   return l1;
1148 }
1149
1150 /** \internal
1151  * \returns the size in Bytes of the L2 or L3 cache if this later is present */
1152 inline int queryTopLevelCacheSize()
1153 {
1154   int l1, l2(-1), l3(-1);
1155   queryCacheSizes(l1,l2,l3);
1156   return (std::max)(l2,l3);
1157 }
1158
1159 } // end namespace internal
1160
1161 } // end namespace Eigen
1162
1163 #endif // EIGEN_MEMORY_H