Memory.h
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
5 // Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
6 // Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
7 // Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
8 // Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
9 //
10 // This Source Code Form is subject to the terms of the Mozilla
11 // Public License v. 2.0. If a copy of the MPL was not distributed
12 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
13 
14 
15 /*****************************************************************************
16 *** Platform checks for aligned malloc functions ***
17 *****************************************************************************/
18 
19 #ifndef EIGEN_MEMORY_H
20 #define EIGEN_MEMORY_H
21 
22 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
23 // http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
24 // This is true at least since glibc 2.8.
25 // This leaves the question how to detect 64-bit. According to this document,
26 // http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
27 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
28 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
29 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
30  && defined(__LP64__)
31  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
32 #else
33  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
34 #endif
35 
36 // FreeBSD 6 seems to have 16-byte aligned malloc
37 // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
38 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
39 // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
40 #if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
41  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
42 #else
43  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
44 #endif
45 
46 #if defined(__APPLE__) \
47  || defined(_WIN64) \
48  || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
49  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
50  #define EIGEN_MALLOC_ALREADY_ALIGNED 1
51 #else
52  #define EIGEN_MALLOC_ALREADY_ALIGNED 0
53 #endif
54 
55 #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) \
56  && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
57  #define EIGEN_HAS_POSIX_MEMALIGN 1
58 #else
59  #define EIGEN_HAS_POSIX_MEMALIGN 0
60 #endif
61 
62 #ifdef EIGEN_VECTORIZE_SSE
63  #define EIGEN_HAS_MM_MALLOC 1
64 #else
65  #define EIGEN_HAS_MM_MALLOC 0
66 #endif
67 
68 namespace Eigen {
69 
70 namespace internal {
71 
72 inline void throw_std_bad_alloc()
73 {
74  #ifdef EIGEN_EXCEPTIONS
75  throw std::bad_alloc();
76  #else
77  std::size_t huge = -1;
78  new int[huge];
79  #endif
80 }
81 
82 /*****************************************************************************
83 *** Implementation of handmade aligned functions ***
84 *****************************************************************************/
85 
86 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
87 
91 inline void* handmade_aligned_malloc(std::size_t size)
92 {
93  void *original = std::malloc(size+16);
94  if (original == 0) return 0;
95  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
96  *(reinterpret_cast<void**>(aligned) - 1) = original;
97  return aligned;
98 }
99 
101 inline void handmade_aligned_free(void *ptr)
102 {
103  if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
104 }
105 
111 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
112 {
113  if (ptr == 0) return handmade_aligned_malloc(size);
114  void *original = *(reinterpret_cast<void**>(ptr) - 1);
115  std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
116  original = std::realloc(original,size+16);
117  if (original == 0) return 0;
118  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
119  void *previous_aligned = static_cast<char *>(original)+previous_offset;
120  if(aligned!=previous_aligned)
121  std::memmove(aligned, previous_aligned, size);
122 
123  *(reinterpret_cast<void**>(aligned) - 1) = original;
124  return aligned;
125 }
126 
127 /*****************************************************************************
128 *** Implementation of generic aligned realloc (when no realloc can be used)***
129 *****************************************************************************/
130 
131 void* aligned_malloc(std::size_t size);
132 void aligned_free(void *ptr);
133 
139 inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
140 {
141  if (ptr==0)
142  return aligned_malloc(size);
143 
144  if (size==0)
145  {
146  aligned_free(ptr);
147  return 0;
148  }
149 
150  void* newptr = aligned_malloc(size);
151  if (newptr == 0)
152  {
153  #ifdef EIGEN_HAS_ERRNO
154  errno = ENOMEM; // according to the standard
155  #endif
156  return 0;
157  }
158 
159  if (ptr != 0)
160  {
161  std::memcpy(newptr, ptr, (std::min)(size,old_size));
162  aligned_free(ptr);
163  }
164 
165  return newptr;
166 }
167 
168 /*****************************************************************************
169 *** Implementation of portable aligned versions of malloc/free/realloc ***
170 *****************************************************************************/
171 
172 #ifdef EIGEN_NO_MALLOC
173 inline void check_that_malloc_is_allowed()
174 {
175  eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
176 }
177 #elif defined EIGEN_RUNTIME_NO_MALLOC
178 inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
179 {
180  static bool value = true;
181  if (update == 1)
182  value = new_value;
183  return value;
184 }
185 inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
186 inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
187 inline void check_that_malloc_is_allowed()
188 {
189  eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
190 }
191 #else
192 inline void check_that_malloc_is_allowed()
193 {}
194 #endif
195 
199 inline void* aligned_malloc(size_t size)
200 {
201  check_that_malloc_is_allowed();
202 
203  void *result;
204  #if !EIGEN_ALIGN
205  result = std::malloc(size);
206  #elif EIGEN_MALLOC_ALREADY_ALIGNED
207  result = std::malloc(size);
208  #elif EIGEN_HAS_POSIX_MEMALIGN
209  if(posix_memalign(&result, 16, size)) result = 0;
210  #elif EIGEN_HAS_MM_MALLOC
211  result = _mm_malloc(size, 16);
212 #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
213  result = _aligned_malloc(size, 16);
214  #else
215  result = handmade_aligned_malloc(size);
216  #endif
217 
218  if(!result && size)
219  throw_std_bad_alloc();
220 
221  return result;
222 }
223 
225 inline void aligned_free(void *ptr)
226 {
227  #if !EIGEN_ALIGN
228  std::free(ptr);
229  #elif EIGEN_MALLOC_ALREADY_ALIGNED
230  std::free(ptr);
231  #elif EIGEN_HAS_POSIX_MEMALIGN
232  std::free(ptr);
233  #elif EIGEN_HAS_MM_MALLOC
234  _mm_free(ptr);
235  #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
236  _aligned_free(ptr);
237  #else
238  handmade_aligned_free(ptr);
239  #endif
240 }
241 
247 inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
248 {
249  EIGEN_UNUSED_VARIABLE(old_size);
250 
251  void *result;
252 #if !EIGEN_ALIGN
253  result = std::realloc(ptr,new_size);
254 #elif EIGEN_MALLOC_ALREADY_ALIGNED
255  result = std::realloc(ptr,new_size);
256 #elif EIGEN_HAS_POSIX_MEMALIGN
257  result = generic_aligned_realloc(ptr,new_size,old_size);
258 #elif EIGEN_HAS_MM_MALLOC
259  // The defined(_mm_free) is just here to verify that this MSVC version
260  // implements _mm_malloc/_mm_free based on the corresponding _aligned_
261  // functions. This may not always be the case and we just try to be safe.
262  #if defined(_MSC_VER) && defined(_mm_free)
263  result = _aligned_realloc(ptr,new_size,16);
264  #else
265  result = generic_aligned_realloc(ptr,new_size,old_size);
266  #endif
267 #elif defined(_MSC_VER)
268  result = _aligned_realloc(ptr,new_size,16);
269 #else
270  result = handmade_aligned_realloc(ptr,new_size,old_size);
271 #endif
272 
273  if (!result && new_size)
274  throw_std_bad_alloc();
275 
276  return result;
277 }
278 
279 /*****************************************************************************
280 *** Implementation of conditionally aligned functions ***
281 *****************************************************************************/
282 
286 template<bool Align> inline void* conditional_aligned_malloc(size_t size)
287 {
288  return aligned_malloc(size);
289 }
290 
291 template<> inline void* conditional_aligned_malloc<false>(size_t size)
292 {
293  check_that_malloc_is_allowed();
294 
295  void *result = std::malloc(size);
296  if(!result && size)
297  throw_std_bad_alloc();
298  return result;
299 }
300 
302 template<bool Align> inline void conditional_aligned_free(void *ptr)
303 {
304  aligned_free(ptr);
305 }
306 
307 template<> inline void conditional_aligned_free<false>(void *ptr)
308 {
309  std::free(ptr);
310 }
311 
312 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
313 {
314  return aligned_realloc(ptr, new_size, old_size);
315 }
316 
317 template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
318 {
319  return std::realloc(ptr, new_size);
320 }
321 
322 /*****************************************************************************
323 *** Construction/destruction of array elements ***
324 *****************************************************************************/
325 
329 template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
330 {
331  for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
332  return ptr;
333 }
334 
338 template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
339 {
340  // always destruct an array starting from the end.
341  if(ptr)
342  while(size) ptr[--size].~T();
343 }
344 
345 /*****************************************************************************
346 *** Implementation of aligned new/delete-like functions ***
347 *****************************************************************************/
348 
349 template<typename T>
350 EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
351 {
352  if(size > size_t(-1) / sizeof(T))
353  throw_std_bad_alloc();
354 }
355 
360 template<typename T> inline T* aligned_new(size_t size)
361 {
362  check_size_for_overflow<T>(size);
363  T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
364  return construct_elements_of_array(result, size);
365 }
366 
367 template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
368 {
369  check_size_for_overflow<T>(size);
370  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
371  return construct_elements_of_array(result, size);
372 }
373 
377 template<typename T> inline void aligned_delete(T *ptr, size_t size)
378 {
379  destruct_elements_of_array<T>(ptr, size);
380  aligned_free(ptr);
381 }
382 
386 template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
387 {
388  destruct_elements_of_array<T>(ptr, size);
389  conditional_aligned_free<Align>(ptr);
390 }
391 
392 template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
393 {
394  check_size_for_overflow<T>(new_size);
395  check_size_for_overflow<T>(old_size);
396  if(new_size < old_size)
397  destruct_elements_of_array(pts+new_size, old_size-new_size);
398  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
399  if(new_size > old_size)
400  construct_elements_of_array(result+old_size, new_size-old_size);
401  return result;
402 }
403 
404 
405 template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
406 {
407  check_size_for_overflow<T>(size);
408  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
409  if(NumTraits<T>::RequireInitialization)
410  construct_elements_of_array(result, size);
411  return result;
412 }
413 
414 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
415 {
416  check_size_for_overflow<T>(new_size);
417  check_size_for_overflow<T>(old_size);
418  if(NumTraits<T>::RequireInitialization && (new_size < old_size))
419  destruct_elements_of_array(pts+new_size, old_size-new_size);
420  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
421  if(NumTraits<T>::RequireInitialization && (new_size > old_size))
422  construct_elements_of_array(result+old_size, new_size-old_size);
423  return result;
424 }
425 
426 template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
427 {
428  if(NumTraits<T>::RequireInitialization)
429  destruct_elements_of_array<T>(ptr, size);
430  conditional_aligned_free<Align>(ptr);
431 }
432 
433 /****************************************************************************/
434 
451 template<typename Scalar, typename Index>
452 static inline Index first_aligned(const Scalar* array, Index size)
453 {
454  typedef typename packet_traits<Scalar>::type Packet;
455  enum { PacketSize = packet_traits<Scalar>::size,
456  PacketAlignedMask = PacketSize-1
457  };
458 
459  if(PacketSize==1)
460  {
461  // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
462  // of the array have the same alignment.
463  return 0;
464  }
465  else if(size_t(array) & (sizeof(Scalar)-1))
466  {
467  // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
468  // Consequently, no element of the array is well aligned.
469  return size;
470  }
471  else
472  {
473  return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
474  & PacketAlignedMask, size);
475  }
476 }
477 
478 
479 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
480 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
481 template<typename T, bool UseMemcpy> struct smart_copy_helper;
482 
483 template<typename T> void smart_copy(const T* start, const T* end, T* target)
484 {
485  smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
486 }
487 
488 template<typename T> struct smart_copy_helper<T,true> {
489  static inline void run(const T* start, const T* end, T* target)
490  { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
491 };
492 
493 template<typename T> struct smart_copy_helper<T,false> {
494  static inline void run(const T* start, const T* end, T* target)
495  { std::copy(start, end, target); }
496 };
497 
498 
499 /*****************************************************************************
500 *** Implementation of runtime stack allocation (falling back to malloc) ***
501 *****************************************************************************/
502 
503 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
504 // to the appropriate stack allocation function
505 #ifndef EIGEN_ALLOCA
506  #if (defined __linux__)
507  #define EIGEN_ALLOCA alloca
508  #elif defined(_MSC_VER)
509  #define EIGEN_ALLOCA _alloca
510  #endif
511 #endif
512 
513 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
514 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
515 template<typename T> class aligned_stack_memory_handler
516 {
517  public:
518  /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
519  * Note that \a ptr can be 0 regardless of the other parameters.
520  * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
521  * In this case, the buffer elements will also be destructed when this handler will be destructed.
522  * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
523  **/
524  aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
525  : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
526  {
527  if(NumTraits<T>::RequireInitialization && m_ptr)
528  Eigen::internal::construct_elements_of_array(m_ptr, size);
529  }
530  ~aligned_stack_memory_handler()
531  {
532  if(NumTraits<T>::RequireInitialization && m_ptr)
533  Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
534  if(m_deallocate)
535  Eigen::internal::aligned_free(m_ptr);
536  }
537  protected:
538  T* m_ptr;
539  size_t m_size;
540  bool m_deallocate;
541 };
542 
543 } // end namespace internal
544 
560 #ifdef EIGEN_ALLOCA
561 
562  #ifdef __arm__
563  #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
564  #else
565  #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
566  #endif
567 
568  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
569  Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
570  TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
571  : reinterpret_cast<TYPE*>( \
572  (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
573  : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \
574  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)
575 
576 #else
577 
578  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
579  Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
580  TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \
581  Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
582 
583 #endif
584 
585 
586 /*****************************************************************************
587 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] ***
588 *****************************************************************************/
589 
590 #if EIGEN_ALIGN
591  #ifdef EIGEN_EXCEPTIONS
592  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
593  void* operator new(size_t size, const std::nothrow_t&) throw() { \
594  try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
595  catch (...) { return 0; } \
596  return 0; \
597  }
598  #else
599  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
600  void* operator new(size_t size, const std::nothrow_t&) throw() { \
601  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
602  }
603  #endif
604 
605  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
606  void *operator new(size_t size) { \
607  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
608  } \
609  void *operator new[](size_t size) { \
610  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
611  } \
612  void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
613  void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
614  /* in-place new and delete. since (at least afaik) there is no actual */ \
615  /* memory allocated we can safely let the default implementation handle */ \
616  /* this particular case. */ \
617  static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
618  void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
619  /* nothrow-new (returns zero instead of std::bad_alloc) */ \
620  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
621  void operator delete(void *ptr, const std::nothrow_t&) throw() { \
622  Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
623  } \
624  typedef void eigen_aligned_operator_new_marker_type;
625 #else
626  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
627 #endif
628 
629 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
630 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
631  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
632 
633 /****************************************************************************/
634 
651 template<class T>
653 {
654 public:
655  typedef size_t size_type;
656  typedef std::ptrdiff_t difference_type;
657  typedef T* pointer;
658  typedef const T* const_pointer;
659  typedef T& reference;
660  typedef const T& const_reference;
661  typedef T value_type;
662 
663  template<class U>
664  struct rebind
665  {
666  typedef aligned_allocator<U> other;
667  };
668 
669  pointer address( reference value ) const
670  {
671  return &value;
672  }
673 
674  const_pointer address( const_reference value ) const
675  {
676  return &value;
677  }
678 
680  {
681  }
682 
684  {
685  }
686 
687  template<class U>
689  {
690  }
691 
693  {
694  }
695 
696  size_type max_size() const
697  {
698  return (std::numeric_limits<size_type>::max)();
699  }
700 
701  pointer allocate( size_type num, const void* hint = 0 )
702  {
703  EIGEN_UNUSED_VARIABLE(hint);
704  internal::check_size_for_overflow<T>(num);
705  return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
706  }
707 
708  void construct( pointer p, const T& value )
709  {
710  ::new( p ) T( value );
711  }
712 
713  // Support for c++11
714 #if (__cplusplus >= 201103L)
715  template<typename... Args>
716  void construct(pointer p, Args&&... args)
717  {
718  ::new(p) T(std::forward<Args>(args)...);
719  }
720 #endif
721 
722  void destroy( pointer p )
723  {
724  p->~T();
725  }
726 
727  void deallocate( pointer p, size_type /*num*/ )
728  {
729  internal::aligned_free( p );
730  }
731 
732  bool operator!=(const aligned_allocator<T>& ) const
733  { return false; }
734 
735  bool operator==(const aligned_allocator<T>& ) const
736  { return true; }
737 };
738 
739 //---------- Cache sizes ----------
740 
741 #if !defined(EIGEN_NO_CPUID)
742 # if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
743 # if defined(__PIC__) && defined(__i386__)
744  // Case for x86 with PIC
745 # define EIGEN_CPUID(abcd,func,id) \
746  __asm__ __volatile__ ("xchgl %%ebx, %%esi;cpuid; xchgl %%ebx,%%esi": "=a" (abcd[0]), "=S" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id));
747 # else
748  // Case for x86_64 or x86 w/o PIC
749 # define EIGEN_CPUID(abcd,func,id) \
750  __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id) );
751 # endif
752 # elif defined(_MSC_VER)
753 # if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) )
754 # define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
755 # endif
756 # endif
757 #endif
758 
759 namespace internal {
760 
761 #ifdef EIGEN_CPUID
762 
763 inline bool cpuid_is_vendor(int abcd[4], const char* vendor)
764 {
765  return abcd[1]==(reinterpret_cast<const int*>(vendor))[0] && abcd[3]==(reinterpret_cast<const int*>(vendor))[1] && abcd[2]==(reinterpret_cast<const int*>(vendor))[2];
766 }
767 
768 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
769 {
770  int abcd[4];
771  l1 = l2 = l3 = 0;
772  int cache_id = 0;
773  int cache_type = 0;
774  do {
775  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
776  EIGEN_CPUID(abcd,0x4,cache_id);
777  cache_type = (abcd[0] & 0x0F) >> 0;
778  if(cache_type==1||cache_type==3) // data or unified cache
779  {
780  int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
781  int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
782  int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
783  int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
784  int sets = (abcd[2]); // C[31:0]
785 
786  int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
787 
788  switch(cache_level)
789  {
790  case 1: l1 = cache_size; break;
791  case 2: l2 = cache_size; break;
792  case 3: l3 = cache_size; break;
793  default: break;
794  }
795  }
796  cache_id++;
797  } while(cache_type>0 && cache_id<16);
798 }
799 
800 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
801 {
802  int abcd[4];
803  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
804  l1 = l2 = l3 = 0;
805  EIGEN_CPUID(abcd,0x00000002,0);
806  unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
807  bool check_for_p2_core2 = false;
808  for(int i=0; i<14; ++i)
809  {
810  switch(bytes[i])
811  {
812  case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines
813  case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines
814  case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines
815  case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
816  case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
817  case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines
818  case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines
819  case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
820  case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
821  case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
822  case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
823  case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
824  case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
825  case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
826  case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
827  case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
828  case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
829  case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
830  case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
831  case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
832  case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
833  case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
834  case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core)
835  case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
836  case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
837  case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
838  case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
839  case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
840  case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
841  case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
842  case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
843  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
844  case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
845  case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
846  case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
847  case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
848  case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
849  case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
850  case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
851  case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
852  case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
853  case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
854  case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
855  case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
856  case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
857  case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
858  case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
859  case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
860  case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
861  case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
862  case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
863  case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
864  case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
865  case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
866  case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
867  case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
868  case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
869 
870  default: break;
871  }
872  }
873  if(check_for_p2_core2 && l2 == l3)
874  l3 = 0;
875  l1 *= 1024;
876  l2 *= 1024;
877  l3 *= 1024;
878 }
879 
880 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
881 {
882  if(max_std_funcs>=4)
883  queryCacheSizes_intel_direct(l1,l2,l3);
884  else
885  queryCacheSizes_intel_codes(l1,l2,l3);
886 }
887 
888 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
889 {
890  int abcd[4];
891  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
892  EIGEN_CPUID(abcd,0x80000005,0);
893  l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
894  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
895  EIGEN_CPUID(abcd,0x80000006,0);
896  l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
897  l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
898 }
899 #endif
900 
903 inline void queryCacheSizes(int& l1, int& l2, int& l3)
904 {
905  #ifdef EIGEN_CPUID
906  int abcd[4];
907 
908  // identify the CPU vendor
909  EIGEN_CPUID(abcd,0x0,0);
910  int max_std_funcs = abcd[1];
911  if(cpuid_is_vendor(abcd,"GenuineIntel"))
912  queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
913  else if(cpuid_is_vendor(abcd,"AuthenticAMD") || cpuid_is_vendor(abcd,"AMDisbetter!"))
914  queryCacheSizes_amd(l1,l2,l3);
915  else
916  // by default let's use Intel's API
917  queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
918 
919  // here is the list of other vendors:
920 // ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
921 // ||cpuid_is_vendor(abcd,"CyrixInstead")
922 // ||cpuid_is_vendor(abcd,"CentaurHauls")
923 // ||cpuid_is_vendor(abcd,"GenuineTMx86")
924 // ||cpuid_is_vendor(abcd,"TransmetaCPU")
925 // ||cpuid_is_vendor(abcd,"RiseRiseRise")
926 // ||cpuid_is_vendor(abcd,"Geode by NSC")
927 // ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
928 // ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
929 // ||cpuid_is_vendor(abcd,"NexGenDriven")
930  #else
931  l1 = l2 = l3 = -1;
932  #endif
933 }
934 
937 inline int queryL1CacheSize()
938 {
939  int l1(-1), l2, l3;
940  queryCacheSizes(l1,l2,l3);
941  return l1;
942 }
943 
946 inline int queryTopLevelCacheSize()
947 {
948  int l1, l2(-1), l3(-1);
949  queryCacheSizes(l1,l2,l3);
950  return (std::max)(l2,l3);
951 }
952 
953 } // end namespace internal
954 
955 } // end namespace Eigen
956 
957 #endif // EIGEN_MEMORY_H