Core/util/Memory.h
Go to the documentation of this file.
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 #ifndef EIGEN_MALLOC_ALREADY_ALIGNED
23 
24 // Try to determine automatically if malloc is already aligned.
25 
26 // On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
27 // http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
28 // This is true at least since glibc 2.8.
29 // This leaves the question how to detect 64-bit. According to this document,
30 // http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
31 // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
32 // quite safe, at least within the context of glibc, to equate 64-bit with LP64.
33 #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
34  && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ )
35  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
36 #else
37  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
38 #endif
39 
40 // FreeBSD 6 seems to have 16-byte aligned malloc
41 // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
42 // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
43 // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
44 #if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__)
45  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
46 #else
47  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
48 #endif
49 
50 #if defined(__APPLE__) \
51  || defined(_WIN64) \
52  || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \
53  || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
54  #define EIGEN_MALLOC_ALREADY_ALIGNED 1
55 #else
56  #define EIGEN_MALLOC_ALREADY_ALIGNED 0
57 #endif
58 
59 #endif
60 
61 // See bug 554 (http://eigen.tuxfamily.org/bz/show_bug.cgi?id=554)
62 // It seems to be unsafe to check _POSIX_ADVISORY_INFO without including unistd.h first.
63 // Currently, let's include it only on unix systems:
64 #if defined(__unix__) || defined(__unix)
65  #include <unistd.h>
66  #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
67  #define EIGEN_HAS_POSIX_MEMALIGN 1
68  #endif
69 #endif
70 
71 #ifndef EIGEN_HAS_POSIX_MEMALIGN
72  #define EIGEN_HAS_POSIX_MEMALIGN 0
73 #endif
74 
75 #ifdef EIGEN_VECTORIZE_SSE
76  #define EIGEN_HAS_MM_MALLOC 1
77 #else
78  #define EIGEN_HAS_MM_MALLOC 0
79 #endif
80 
81 namespace Eigen {
82 
83 namespace internal {
84 
85 inline void throw_std_bad_alloc()
86 {
87  #ifdef EIGEN_EXCEPTIONS
88  throw std::bad_alloc();
89  #else
90  std::size_t huge = -1;
91  new int[huge];
92  #endif
93 }
94 
95 /*****************************************************************************
96 *** Implementation of handmade aligned functions ***
97 *****************************************************************************/
98 
99 /* ----- Hand made implementations of aligned malloc/free and realloc ----- */
100 
104 inline void* handmade_aligned_malloc(std::size_t size)
105 {
106  void *original = std::malloc(size+16);
107  if (original == 0) return 0;
108  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
109  *(reinterpret_cast<void**>(aligned) - 1) = original;
110  return aligned;
111 }
112 
114 inline void handmade_aligned_free(void *ptr)
115 {
116  if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
117 }
118 
124 inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
125 {
126  if (ptr == 0) return handmade_aligned_malloc(size);
127  void *original = *(reinterpret_cast<void**>(ptr) - 1);
128  std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
129  original = std::realloc(original,size+16);
130  if (original == 0) return 0;
131  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(15))) + 16);
132  void *previous_aligned = static_cast<char *>(original)+previous_offset;
133  if(aligned!=previous_aligned)
134  std::memmove(aligned, previous_aligned, size);
135 
136  *(reinterpret_cast<void**>(aligned) - 1) = original;
137  return aligned;
138 }
139 
140 /*****************************************************************************
141 *** Implementation of generic aligned realloc (when no realloc can be used)***
142 *****************************************************************************/
143 
144 void* aligned_malloc(std::size_t size);
145 void aligned_free(void *ptr);
146 
152 inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size)
153 {
154  if (ptr==0)
155  return aligned_malloc(size);
156 
157  if (size==0)
158  {
159  aligned_free(ptr);
160  return 0;
161  }
162 
163  void* newptr = aligned_malloc(size);
164  if (newptr == 0)
165  {
166  #ifdef EIGEN_HAS_ERRNO
167  errno = ENOMEM; // according to the standard
168  #endif
169  return 0;
170  }
171 
172  if (ptr != 0)
173  {
174  std::memcpy(newptr, ptr, (std::min)(size,old_size));
175  aligned_free(ptr);
176  }
177 
178  return newptr;
179 }
180 
181 /*****************************************************************************
182 *** Implementation of portable aligned versions of malloc/free/realloc ***
183 *****************************************************************************/
184 
185 #ifdef EIGEN_NO_MALLOC
186 inline void check_that_malloc_is_allowed()
187 {
188  eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
189 }
190 #elif defined EIGEN_RUNTIME_NO_MALLOC
191 inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
192 {
193  static bool value = true;
194  if (update == 1)
195  value = new_value;
196  return value;
197 }
198 inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
199 inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
200 inline void check_that_malloc_is_allowed()
201 {
202  eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
203 }
204 #else
206 {}
207 #endif
208 
212 inline void* aligned_malloc(size_t size)
213 {
215 
216  void *result;
217  #if !EIGEN_ALIGN
218  result = std::malloc(size);
219  #elif EIGEN_MALLOC_ALREADY_ALIGNED
220  result = std::malloc(size);
221  #elif EIGEN_HAS_POSIX_MEMALIGN
222  if(posix_memalign(&result, 16, size)) result = 0;
223  #elif EIGEN_HAS_MM_MALLOC
224  result = _mm_malloc(size, 16);
225  #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
226  result = _aligned_malloc(size, 16);
227  #else
228  result = handmade_aligned_malloc(size);
229  #endif
230 
231  if(!result && size)
233 
234  return result;
235 }
236 
238 inline void aligned_free(void *ptr)
239 {
240  #if !EIGEN_ALIGN
241  std::free(ptr);
242  #elif EIGEN_MALLOC_ALREADY_ALIGNED
243  std::free(ptr);
244  #elif EIGEN_HAS_POSIX_MEMALIGN
245  std::free(ptr);
246  #elif EIGEN_HAS_MM_MALLOC
247  _mm_free(ptr);
248  #elif defined(_MSC_VER) && (!defined(_WIN32_WCE))
249  _aligned_free(ptr);
250  #else
252  #endif
253 }
254 
260 inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size)
261 {
262  EIGEN_UNUSED_VARIABLE(old_size);
263 
264  void *result;
265 #if !EIGEN_ALIGN
266  result = std::realloc(ptr,new_size);
267 #elif EIGEN_MALLOC_ALREADY_ALIGNED
268  result = std::realloc(ptr,new_size);
269 #elif EIGEN_HAS_POSIX_MEMALIGN
270  result = generic_aligned_realloc(ptr,new_size,old_size);
271 #elif EIGEN_HAS_MM_MALLOC
272  // The defined(_mm_free) is just here to verify that this MSVC version
273  // implements _mm_malloc/_mm_free based on the corresponding _aligned_
274  // functions. This may not always be the case and we just try to be safe.
275  #if defined(_MSC_VER) && defined(_mm_free)
276  result = _aligned_realloc(ptr,new_size,16);
277  #else
278  result = generic_aligned_realloc(ptr,new_size,old_size);
279  #endif
280 #elif defined(_MSC_VER)
281  result = _aligned_realloc(ptr,new_size,16);
282 #else
283  result = handmade_aligned_realloc(ptr,new_size,old_size);
284 #endif
285 
286  if (!result && new_size)
288 
289  return result;
290 }
291 
292 /*****************************************************************************
293 *** Implementation of conditionally aligned functions ***
294 *****************************************************************************/
295 
299 template<bool Align> inline void* conditional_aligned_malloc(size_t size)
300 {
301  return aligned_malloc(size);
302 }
303 
304 template<> inline void* conditional_aligned_malloc<false>(size_t size)
305 {
307 
308  void *result = std::malloc(size);
309  if(!result && size)
311  return result;
312 }
313 
315 template<bool Align> inline void conditional_aligned_free(void *ptr)
316 {
317  aligned_free(ptr);
318 }
319 
320 template<> inline void conditional_aligned_free<false>(void *ptr)
321 {
322  std::free(ptr);
323 }
324 
325 template<bool Align> inline void* conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size)
326 {
327  return aligned_realloc(ptr, new_size, old_size);
328 }
329 
330 template<> inline void* conditional_aligned_realloc<false>(void* ptr, size_t new_size, size_t)
331 {
332  return std::realloc(ptr, new_size);
333 }
334 
335 /*****************************************************************************
336 *** Construction/destruction of array elements ***
337 *****************************************************************************/
338 
342 template<typename T> inline T* construct_elements_of_array(T *ptr, size_t size)
343 {
344  for (size_t i=0; i < size; ++i) ::new (ptr + i) T;
345  return ptr;
346 }
347 
351 template<typename T> inline void destruct_elements_of_array(T *ptr, size_t size)
352 {
353  // always destruct an array starting from the end.
354  if(ptr)
355  while(size) ptr[--size].~T();
356 }
357 
358 /*****************************************************************************
359 *** Implementation of aligned new/delete-like functions ***
360 *****************************************************************************/
361 
362 template<typename T>
364 {
365  if(size > size_t(-1) / sizeof(T))
367 }
368 
373 template<typename T> inline T* aligned_new(size_t size)
374 {
375  check_size_for_overflow<T>(size);
376  T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
377  return construct_elements_of_array(result, size);
378 }
379 
380 template<typename T, bool Align> inline T* conditional_aligned_new(size_t size)
381 {
382  check_size_for_overflow<T>(size);
383  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
384  return construct_elements_of_array(result, size);
385 }
386 
390 template<typename T> inline void aligned_delete(T *ptr, size_t size)
391 {
392  destruct_elements_of_array<T>(ptr, size);
393  aligned_free(ptr);
394 }
395 
399 template<typename T, bool Align> inline void conditional_aligned_delete(T *ptr, size_t size)
400 {
401  destruct_elements_of_array<T>(ptr, size);
402  conditional_aligned_free<Align>(ptr);
403 }
404 
405 template<typename T, bool Align> inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size)
406 {
407  check_size_for_overflow<T>(new_size);
408  check_size_for_overflow<T>(old_size);
409  if(new_size < old_size)
410  destruct_elements_of_array(pts+new_size, old_size-new_size);
411  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
412  if(new_size > old_size)
413  construct_elements_of_array(result+old_size, new_size-old_size);
414  return result;
415 }
416 
417 
418 template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
419 {
420  check_size_for_overflow<T>(size);
421  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
423  construct_elements_of_array(result, size);
424  return result;
425 }
426 
427 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
428 {
429  check_size_for_overflow<T>(new_size);
430  check_size_for_overflow<T>(old_size);
431  if(NumTraits<T>::RequireInitialization && (new_size < old_size))
432  destruct_elements_of_array(pts+new_size, old_size-new_size);
433  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
434  if(NumTraits<T>::RequireInitialization && (new_size > old_size))
435  construct_elements_of_array(result+old_size, new_size-old_size);
436  return result;
437 }
438 
439 template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
440 {
442  destruct_elements_of_array<T>(ptr, size);
443  conditional_aligned_free<Align>(ptr);
444 }
445 
446 /****************************************************************************/
447 
464 template<typename Scalar, typename Index>
465 static inline Index first_aligned(const Scalar* array, Index size)
466 {
467  enum { PacketSize = packet_traits<Scalar>::size,
468  PacketAlignedMask = PacketSize-1
469  };
470 
471  if(PacketSize==1)
472  {
473  // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
474  // of the array have the same alignment.
475  return 0;
476  }
477  else if(size_t(array) & (sizeof(Scalar)-1))
478  {
479  // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
480  // Consequently, no element of the array is well aligned.
481  return size;
482  }
483  else
484  {
485  return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
486  & PacketAlignedMask, size);
487  }
488 }
489 
492 template<typename Index>
493 inline static Index first_multiple(Index size, Index base)
494 {
495  return ((size+base-1)/base)*base;
496 }
497 
498 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
499 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
500 template<typename T, bool UseMemcpy> struct smart_copy_helper;
501 
502 template<typename T> void smart_copy(const T* start, const T* end, T* target)
503 {
505 }
506 
507 template<typename T> struct smart_copy_helper<T,true> {
508  static inline void run(const T* start, const T* end, T* target)
509  { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
510 };
511 
512 template<typename T> struct smart_copy_helper<T,false> {
513  static inline void run(const T* start, const T* end, T* target)
514  { std::copy(start, end, target); }
515 };
516 
517 
518 /*****************************************************************************
519 *** Implementation of runtime stack allocation (falling back to malloc) ***
520 *****************************************************************************/
521 
522 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
523 // to the appropriate stack allocation function
524 #ifndef EIGEN_ALLOCA
525  #if (defined __linux__)
526  #define EIGEN_ALLOCA alloca
527  #elif defined(_MSC_VER)
528  #define EIGEN_ALLOCA _alloca
529  #endif
530 #endif
531 
532 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
533 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
534 template<typename T> class aligned_stack_memory_handler
535 {
536  public:
537  /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
538  * Note that \a ptr can be 0 regardless of the other parameters.
539  * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
540  * In this case, the buffer elements will also be destructed when this handler will be destructed.
541  * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
542  **/
543  aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
544  : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
545  {
548  }
550  {
552  Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
553  if(m_deallocate)
555  }
556  protected:
558  size_t m_size;
560 };
561 
562 } // end namespace internal
563 
579 #ifdef EIGEN_ALLOCA
580 
581  #ifdef __arm__
582  #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
583  #else
584  #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
585  #endif
586 
587  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
588  Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
589  TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
590  : reinterpret_cast<TYPE*>( \
591  (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
592  : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) ); \
593  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)
594 
595 #else
596 
597  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
598  Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
599  TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE)); \
600  Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
601 
602 #endif
603 
604 
605 /*****************************************************************************
606 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF] ***
607 *****************************************************************************/
608 
609 #if EIGEN_ALIGN
610  #ifdef EIGEN_EXCEPTIONS
611  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
612  void* operator new(size_t size, const std::nothrow_t&) throw() { \
613  try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
614  catch (...) { return 0; } \
615  return 0; \
616  }
617  #else
618  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
619  void* operator new(size_t size, const std::nothrow_t&) throw() { \
620  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
621  }
622  #endif
623 
624  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
625  void *operator new(size_t size) { \
626  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
627  } \
628  void *operator new[](size_t size) { \
629  return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
630  } \
631  void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
632  void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
633  /* in-place new and delete. since (at least afaik) there is no actual */ \
634  /* memory allocated we can safely let the default implementation handle */ \
635  /* this particular case. */ \
636  static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \
637  void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
638  /* nothrow-new (returns zero instead of std::bad_alloc) */ \
639  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
640  void operator delete(void *ptr, const std::nothrow_t&) throw() { \
641  Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
642  } \
643  typedef void eigen_aligned_operator_new_marker_type;
644 #else
645  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
646 #endif
647 
648 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
649 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
650  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
651 
652 /****************************************************************************/
653 
670 template<class T>
672 {
673 public:
674  typedef size_t size_type;
675  typedef std::ptrdiff_t difference_type;
676  typedef T* pointer;
677  typedef const T* const_pointer;
678  typedef T& reference;
679  typedef const T& const_reference;
680  typedef T value_type;
681 
682  template<class U>
683  struct rebind
684  {
686  };
687 
688  pointer address( reference value ) const
689  {
690  return &value;
691  }
692 
693  const_pointer address( const_reference value ) const
694  {
695  return &value;
696  }
697 
699  {
700  }
701 
703  {
704  }
705 
706  template<class U>
708  {
709  }
710 
712  {
713  }
714 
715  size_type max_size() const
716  {
717  return (std::numeric_limits<size_type>::max)();
718  }
719 
720  pointer allocate( size_type num, const void* hint = 0 )
721  {
722  EIGEN_UNUSED_VARIABLE(hint);
723  internal::check_size_for_overflow<T>(num);
724  return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
725  }
726 
727  void construct( pointer p, const T& value )
728  {
729  ::new( p ) T( value );
730  }
731 
732  // Support for c++11
733 #if (__cplusplus >= 201103L)
734  template<typename... Args>
735  void construct(pointer p, Args&&... args)
736  {
737  ::new(p) T(std::forward<Args>(args)...);
738  }
739 #endif
740 
741  void destroy( pointer p )
742  {
743  p->~T();
744  }
745 
746  void deallocate( pointer p, size_type /*num*/ )
747  {
749  }
750 
751  bool operator!=(const aligned_allocator<T>& ) const
752  { return false; }
753 
754  bool operator==(const aligned_allocator<T>& ) const
755  { return true; }
756 };
757 
758 //---------- Cache sizes ----------
759 
760 #if !defined(EIGEN_NO_CPUID)
761 # if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
762 # if defined(__PIC__) && defined(__i386__)
763  // Case for x86 with PIC
764 # define EIGEN_CPUID(abcd,func,id) \
765  __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));
766 # elif defined(__PIC__) && defined(__x86_64__)
767  // 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.
768  // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
769 # define EIGEN_CPUID(abcd,func,id) \
770  __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));
771 # else
772  // Case for x86_64 or x86 w/o PIC
773 # define EIGEN_CPUID(abcd,func,id) \
774  __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
775 # endif
776 # elif defined(_MSC_VER)
777 # if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) )
778 # define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
779 # endif
780 # endif
781 #endif
782 
783 namespace internal {
784 
785 #ifdef EIGEN_CPUID
786 
787 inline bool cpuid_is_vendor(int abcd[4], const char* vendor)
788 {
789  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];
790 }
791 
792 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
793 {
794  int abcd[4];
795  l1 = l2 = l3 = 0;
796  int cache_id = 0;
797  int cache_type = 0;
798  do {
799  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
800  EIGEN_CPUID(abcd,0x4,cache_id);
801  cache_type = (abcd[0] & 0x0F) >> 0;
802  if(cache_type==1||cache_type==3) // data or unified cache
803  {
804  int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
805  int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
806  int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
807  int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
808  int sets = (abcd[2]); // C[31:0]
809 
810  int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
811 
812  switch(cache_level)
813  {
814  case 1: l1 = cache_size; break;
815  case 2: l2 = cache_size; break;
816  case 3: l3 = cache_size; break;
817  default: break;
818  }
819  }
820  cache_id++;
821  } while(cache_type>0 && cache_id<16);
822 }
823 
824 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
825 {
826  int abcd[4];
827  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
828  l1 = l2 = l3 = 0;
829  EIGEN_CPUID(abcd,0x00000002,0);
830  unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
831  bool check_for_p2_core2 = false;
832  for(int i=0; i<14; ++i)
833  {
834  switch(bytes[i])
835  {
836  case 0x0A: l1 = 8; break; // 0Ah data L1 cache, 8 KB, 2 ways, 32 byte lines
837  case 0x0C: l1 = 16; break; // 0Ch data L1 cache, 16 KB, 4 ways, 32 byte lines
838  case 0x0E: l1 = 24; break; // 0Eh data L1 cache, 24 KB, 6 ways, 64 byte lines
839  case 0x10: l1 = 16; break; // 10h data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
840  case 0x15: l1 = 16; break; // 15h code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
841  case 0x2C: l1 = 32; break; // 2Ch data L1 cache, 32 KB, 8 ways, 64 byte lines
842  case 0x30: l1 = 32; break; // 30h code L1 cache, 32 KB, 8 ways, 64 byte lines
843  case 0x60: l1 = 16; break; // 60h data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
844  case 0x66: l1 = 8; break; // 66h data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
845  case 0x67: l1 = 16; break; // 67h data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
846  case 0x68: l1 = 32; break; // 68h data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
847  case 0x1A: l2 = 96; break; // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
848  case 0x22: l3 = 512; break; // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
849  case 0x23: l3 = 1024; break; // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
850  case 0x25: l3 = 2048; break; // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
851  case 0x29: l3 = 4096; break; // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
852  case 0x39: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
853  case 0x3A: l2 = 192; break; // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
854  case 0x3B: l2 = 128; break; // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
855  case 0x3C: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
856  case 0x3D: l2 = 384; break; // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
857  case 0x3E: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
858  case 0x40: l2 = 0; break; // no integrated L2 cache (P6 core) or L3 cache (P4 core)
859  case 0x41: l2 = 128; break; // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
860  case 0x42: l2 = 256; break; // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
861  case 0x43: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
862  case 0x44: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
863  case 0x45: l2 = 2048; break; // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
864  case 0x46: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
865  case 0x47: l3 = 8192; break; // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
866  case 0x48: l2 = 3072; break; // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
867  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
868  case 0x4A: l3 = 6144; break; // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
869  case 0x4B: l3 = 8192; break; // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
870  case 0x4C: l3 = 12288; break; // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
871  case 0x4D: l3 = 16384; break; // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
872  case 0x4E: l2 = 6144; break; // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
873  case 0x78: l2 = 1024; break; // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
874  case 0x79: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
875  case 0x7A: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
876  case 0x7B: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
877  case 0x7C: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
878  case 0x7D: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
879  case 0x7E: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
880  case 0x7F: l2 = 512; break; // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
881  case 0x80: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
882  case 0x81: l2 = 128; break; // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
883  case 0x82: l2 = 256; break; // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
884  case 0x83: l2 = 512; break; // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
885  case 0x84: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
886  case 0x85: l2 = 2048; break; // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
887  case 0x86: l2 = 512; break; // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
888  case 0x87: l2 = 1024; break; // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
889  case 0x88: l3 = 2048; break; // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
890  case 0x89: l3 = 4096; break; // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
891  case 0x8A: l3 = 8192; break; // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
892  case 0x8D: l3 = 3072; break; // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
893 
894  default: break;
895  }
896  }
897  if(check_for_p2_core2 && l2 == l3)
898  l3 = 0;
899  l1 *= 1024;
900  l2 *= 1024;
901  l3 *= 1024;
902 }
903 
904 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
905 {
906  if(max_std_funcs>=4)
907  queryCacheSizes_intel_direct(l1,l2,l3);
908  else
909  queryCacheSizes_intel_codes(l1,l2,l3);
910 }
911 
912 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
913 {
914  int abcd[4];
915  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
916  EIGEN_CPUID(abcd,0x80000005,0);
917  l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
918  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
919  EIGEN_CPUID(abcd,0x80000006,0);
920  l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
921  l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
922 }
923 #endif
924 
927 inline void queryCacheSizes(int& l1, int& l2, int& l3)
928 {
929  #ifdef EIGEN_CPUID
930  int abcd[4];
931 
932  // identify the CPU vendor
933  EIGEN_CPUID(abcd,0x0,0);
934  int max_std_funcs = abcd[1];
935  if(cpuid_is_vendor(abcd,"GenuineIntel"))
936  queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
937  else if(cpuid_is_vendor(abcd,"AuthenticAMD") || cpuid_is_vendor(abcd,"AMDisbetter!"))
938  queryCacheSizes_amd(l1,l2,l3);
939  else
940  // by default let's use Intel's API
941  queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
942 
943  // here is the list of other vendors:
944 // ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
945 // ||cpuid_is_vendor(abcd,"CyrixInstead")
946 // ||cpuid_is_vendor(abcd,"CentaurHauls")
947 // ||cpuid_is_vendor(abcd,"GenuineTMx86")
948 // ||cpuid_is_vendor(abcd,"TransmetaCPU")
949 // ||cpuid_is_vendor(abcd,"RiseRiseRise")
950 // ||cpuid_is_vendor(abcd,"Geode by NSC")
951 // ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
952 // ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
953 // ||cpuid_is_vendor(abcd,"NexGenDriven")
954  #else
955  l1 = l2 = l3 = -1;
956  #endif
957 }
958 
961 inline int queryL1CacheSize()
962 {
963  int l1(-1), l2, l3;
964  queryCacheSizes(l1,l2,l3);
965  return l1;
966 }
967 
971 {
972  int l1, l2(-1), l3(-1);
973  queryCacheSizes(l1,l2,l3);
974  return (std::max)(l2,l3);
975 }
976 
977 } // end namespace internal
978 
979 } // end namespace Eigen
980 
981 #endif // EIGEN_MEMORY_H
static void run(const T *start, const T *end, T *target)
void * conditional_aligned_malloc< false >(size_t size)
void * handmade_aligned_realloc(void *ptr, std::size_t size, std::size_t=0)
bool operator!=(const aligned_allocator< T > &) const
static Index first_multiple(Index size, Index base)
void conditional_aligned_free(void *ptr)
USING_NAMESPACE_ACADO typedef TaylorVariable< Interval > T
T * conditional_aligned_realloc_new_auto(T *pts, size_t new_size, size_t old_size)
void * aligned_realloc(void *ptr, size_t new_size, size_t old_size)
#define EIGEN_UNUSED_VARIABLE(var)
iterative scaling algorithm to equilibrate rows and column norms in matrices
Definition: matrix.hpp:471
Holds information about the various numeric (i.e. scalar) types allowed by Eigen. ...
Definition: NumTraits.h:88
void * generic_aligned_realloc(void *ptr, size_t size, size_t old_size)
pointer address(reference value) const
void aligned_free(void *ptr)
T * conditional_aligned_new_auto(size_t size)
aligned_allocator(const aligned_allocator &)
const_pointer address(const_reference value) const
void conditional_aligned_free< false >(void *ptr)
void * handmade_aligned_malloc(std::size_t size)
T * conditional_aligned_realloc_new(T *pts, size_t new_size, size_t old_size)
T * conditional_aligned_new(size_t size)
size_type max_size() const
void throw_std_bad_alloc()
STL compatible allocator to use with with 16 byte aligned types.
void construct(pointer p, const T &value)
void * conditional_aligned_realloc< false >(void *ptr, size_t new_size, size_t)
void queryCacheSizes(int &l1, int &l2, int &l3)
pointer allocate(size_type num, const void *hint=0)
void smart_copy(const T *start, const T *end, T *target)
T * construct_elements_of_array(T *ptr, size_t size)
void conditional_aligned_delete_auto(T *ptr, size_t size)
void aligned_delete(T *ptr, size_t size)
void deallocate(pointer p, size_type)
void conditional_aligned_delete(T *ptr, size_t size)
T * aligned_new(size_t size)
std::ptrdiff_t difference_type
void check_that_malloc_is_allowed()
bool operator==(const aligned_allocator< T > &) const
void * conditional_aligned_malloc(size_t size)
#define EIGEN_ALWAYS_INLINE
static void run(const T *start, const T *end, T *target)
void handmade_aligned_free(void *ptr)
aligned_stack_memory_handler(T *ptr, size_t size, bool dealloc)
static Derived::Index first_aligned(const Derived &m)
#define eigen_assert(x)
void * conditional_aligned_realloc(void *ptr, size_t new_size, size_t old_size)
void destruct_elements_of_array(T *ptr, size_t size)
aligned_allocator(const aligned_allocator< U > &)
void * aligned_malloc(std::size_t size)
EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)


acado
Author(s): Milan Vukov, Rien Quirynen
autogenerated on Mon Jun 10 2019 12:34:53