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 __PGI) || ((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 #if defined(EIGEN_VECTORIZE_SSE) && !defined(EIGEN_ANDROID_POSIX_MEMALIGN_WR)
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 
throw_std_bad_alloc()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 
101 /** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
102   * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
103   */
handmade_aligned_malloc(std::size_t size)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 
113 /** \internal Frees memory allocated with handmade_aligned_malloc */
handmade_aligned_free(void * ptr)114 inline void handmade_aligned_free(void *ptr)
115 {
116   if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
117 }
118 
119 /** \internal
120   * \brief Reallocates aligned memory.
121   * Since we know that our handmade version is based on std::realloc
122   * we can use std::realloc to implement efficient reallocation.
123   */
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 
147 /** \internal
148   * \brief Reallocates aligned memory.
149   * Allows reallocation with aligned ptr types. This implementation will
150   * always create a new memory chunk and copy the old data.
151   */
generic_aligned_realloc(void * ptr,size_t size,size_t old_size)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
check_that_malloc_is_allowed()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 }
is_malloc_allowed()198 inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
set_is_malloc_allowed(bool new_value)199 inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
check_that_malloc_is_allowed()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
check_that_malloc_is_allowed()205 inline void check_that_malloc_is_allowed()
206 {}
207 #endif
208 
209 /** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment.
210   * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
211   */
aligned_malloc(size_t size)212 inline void* aligned_malloc(size_t size)
213 {
214   check_that_malloc_is_allowed();
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)
232     throw_std_bad_alloc();
233 
234   return result;
235 }
236 
237 /** \internal Frees memory allocated with aligned_malloc. */
aligned_free(void * ptr)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
251     handmade_aligned_free(ptr);
252   #endif
253 }
254 
255 /**
256 * \internal
257 * \brief Reallocates an aligned block of memory.
258 * \throws std::bad_alloc on allocation failure
259 **/
aligned_realloc(void * ptr,size_t new_size,size_t old_size)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(_WIN32_WCE)) && 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) && (!defined(_WIN32_WCE))
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)
287     throw_std_bad_alloc();
288 
289   return result;
290 }
291 
292 /*****************************************************************************
293 *** Implementation of conditionally aligned functions                      ***
294 *****************************************************************************/
295 
296 /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
297   * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
298   */
conditional_aligned_malloc(size_t size)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 {
306   check_that_malloc_is_allowed();
307 
308   void *result = std::malloc(size);
309   if(!result && size)
310     throw_std_bad_alloc();
311   return result;
312 }
313 
314 /** \internal Frees memory allocated with conditional_aligned_malloc */
conditional_aligned_free(void * ptr)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 
conditional_aligned_realloc(void * ptr,size_t new_size,size_t old_size)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 
339 /** \internal Constructs the elements of an array.
340   * The \a size parameter tells on how many objects to call the constructor of T.
341   */
construct_elements_of_array(T * ptr,size_t size)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 
348 /** \internal Destructs the elements of an array.
349   * The \a size parameters tells on how many objects to call the destructor of T.
350   */
destruct_elements_of_array(T * ptr,size_t size)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>
check_size_for_overflow(size_t size)363 EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size)
364 {
365   if(size > size_t(-1) / sizeof(T))
366     throw_std_bad_alloc();
367 }
368 
369 /** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
370   * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
371   * The default constructor of T is called.
372   */
aligned_new(size_t size)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 
conditional_aligned_new(size_t size)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 
387 /** \internal Deletes objects constructed with aligned_new
388   * The \a size parameters tells on how many objects to call the destructor of T.
389   */
aligned_delete(T * ptr,size_t size)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 
396 /** \internal Deletes objects constructed with conditional_aligned_new
397   * The \a size parameters tells on how many objects to call the destructor of T.
398   */
conditional_aligned_delete(T * ptr,size_t size)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 
conditional_aligned_realloc_new(T * pts,size_t new_size,size_t old_size)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 
conditional_aligned_new_auto(size_t size)418 template<typename T, bool Align> inline T* conditional_aligned_new_auto(size_t size)
419 {
420   if(size==0)
421     return 0; // short-cut. Also fixes Bug 884
422   check_size_for_overflow<T>(size);
423   T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
424   if(NumTraits<T>::RequireInitialization)
425     construct_elements_of_array(result, size);
426   return result;
427 }
428 
conditional_aligned_realloc_new_auto(T * pts,size_t new_size,size_t old_size)429 template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, size_t new_size, size_t old_size)
430 {
431   check_size_for_overflow<T>(new_size);
432   check_size_for_overflow<T>(old_size);
433   if(NumTraits<T>::RequireInitialization && (new_size < old_size))
434     destruct_elements_of_array(pts+new_size, old_size-new_size);
435   T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
436   if(NumTraits<T>::RequireInitialization && (new_size > old_size))
437     construct_elements_of_array(result+old_size, new_size-old_size);
438   return result;
439 }
440 
conditional_aligned_delete_auto(T * ptr,size_t size)441 template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *ptr, size_t size)
442 {
443   if(NumTraits<T>::RequireInitialization)
444     destruct_elements_of_array<T>(ptr, size);
445   conditional_aligned_free<Align>(ptr);
446 }
447 
448 /****************************************************************************/
449 
450 /** \internal Returns the index of the first element of the array that is well aligned for vectorization.
451   *
452   * \param array the address of the start of the array
453   * \param size the size of the array
454   *
455   * \note If no element of the array is well aligned, the size of the array is returned. Typically,
456   * for example with SSE, "well aligned" means 16-byte-aligned. If vectorization is disabled or if the
457   * packet size for the given scalar type is 1, then everything is considered well-aligned.
458   *
459   * \note If the scalar type is vectorizable, we rely on the following assumptions: sizeof(Scalar) is a
460   * power of 2, the packet size in bytes is also a power of 2, and is a multiple of sizeof(Scalar). On the
461   * other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
462   * example with Scalar=double on certain 32-bit platforms, see bug #79.
463   *
464   * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
465   */
466 template<typename Scalar, typename Index>
first_aligned(const Scalar * array,Index size)467 static inline Index first_aligned(const Scalar* array, Index size)
468 {
469   static const Index PacketSize = packet_traits<Scalar>::size;
470   static const Index PacketAlignedMask = PacketSize-1;
471 
472   if(PacketSize==1)
473   {
474     // Either there is no vectorization, or a packet consists of exactly 1 scalar so that all elements
475     // of the array have the same alignment.
476     return 0;
477   }
478   else if(size_t(array) & (sizeof(Scalar)-1))
479   {
480     // There is vectorization for this scalar type, but the array is not aligned to the size of a single scalar.
481     // Consequently, no element of the array is well aligned.
482     return size;
483   }
484   else
485   {
486     return std::min<Index>( (PacketSize - (Index((size_t(array)/sizeof(Scalar))) & PacketAlignedMask))
487                            & PacketAlignedMask, size);
488   }
489 }
490 
491 /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
492   */
493 template<typename Index>
first_multiple(Index size,Index base)494 inline static Index first_multiple(Index size, Index base)
495 {
496   return ((size+base-1)/base)*base;
497 }
498 
499 // std::copy is much slower than memcpy, so let's introduce a smart_copy which
500 // use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
501 template<typename T, bool UseMemcpy> struct smart_copy_helper;
502 
smart_copy(const T * start,const T * end,T * target)503 template<typename T> void smart_copy(const T* start, const T* end, T* target)
504 {
505   smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
506 }
507 
508 template<typename T> struct smart_copy_helper<T,true> {
509   static inline void run(const T* start, const T* end, T* target)
510   { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
511 };
512 
513 template<typename T> struct smart_copy_helper<T,false> {
514   static inline void run(const T* start, const T* end, T* target)
515   { std::copy(start, end, target); }
516 };
517 
518 
519 /*****************************************************************************
520 *** Implementation of runtime stack allocation (falling back to malloc)    ***
521 *****************************************************************************/
522 
523 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
524 // to the appropriate stack allocation function
525 #ifndef EIGEN_ALLOCA
526   #if (defined __linux__) || (defined __APPLE__) || (defined alloca)
527     #define EIGEN_ALLOCA alloca
528   #elif defined(_MSC_VER)
529     #define EIGEN_ALLOCA _alloca
530   #endif
531 #endif
532 
533 // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
534 // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
535 template<typename T> class aligned_stack_memory_handler
536 {
537   public:
538     /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
539      * Note that \a ptr can be 0 regardless of the other parameters.
540      * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
541      * In this case, the buffer elements will also be destructed when this handler will be destructed.
542      * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
543      **/
544     aligned_stack_memory_handler(T* ptr, size_t size, bool dealloc)
545       : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
546     {
547       if(NumTraits<T>::RequireInitialization && m_ptr)
548         Eigen::internal::construct_elements_of_array(m_ptr, size);
549     }
550     ~aligned_stack_memory_handler()
551     {
552       if(NumTraits<T>::RequireInitialization && m_ptr)
553         Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
554       if(m_deallocate)
555         Eigen::internal::aligned_free(m_ptr);
556     }
557   protected:
558     T* m_ptr;
559     size_t m_size;
560     bool m_deallocate;
561 };
562 
563 } // end namespace internal
564 
565 /** \internal
566   * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
567   * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
568   * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
569   * The allocated buffer is automatically deleted when exiting the scope of this declaration.
570   * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
571   * Here is an example:
572   * \code
573   * {
574   *   ei_declare_aligned_stack_constructed_variable(float,data,size,0);
575   *   // use data[0] to data[size-1]
576   * }
577   * \endcode
578   * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
579   */
580 #ifdef EIGEN_ALLOCA
581 
582   #if defined(__arm__) || defined(_WIN32)
583     #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
584   #else
585     #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
586   #endif
587 
588   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
589     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
590     TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
591                : reinterpret_cast<TYPE*>( \
592                       (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
593                     : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
594     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)
595 
596 #else
597 
598   #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
599     Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
600     TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
601     Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
602 
603 #endif
604 
605 
606 /*****************************************************************************
607 *** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
608 *****************************************************************************/
609 
610 #if EIGEN_ALIGN
611   #ifdef EIGEN_EXCEPTIONS
612     #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
613       void* operator new(size_t size, const std::nothrow_t&) throw() { \
614         try { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
615         catch (...) { 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       static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \
638       void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \
639       void operator delete[](void * memory, void *ptr) throw() { return ::operator delete[](memory,ptr); } \
640       /* nothrow-new (returns zero instead of std::bad_alloc) */ \
641       EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
642       void operator delete(void *ptr, const std::nothrow_t&) throw() { \
643         Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
644       } \
645       typedef void eigen_aligned_operator_new_marker_type;
646 #else
647   #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
648 #endif
649 
650 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
651 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
652   EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
653 
654 /****************************************************************************/
655 
656 /** \class aligned_allocator
657 * \ingroup Core_Module
658 *
659 * \brief STL compatible allocator to use with with 16 byte aligned types
660 *
661 * Example:
662 * \code
663 * // Matrix4f requires 16 bytes alignment:
664 * std::map< int, Matrix4f, std::less<int>,
665 *           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
666 * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
667 * std::map< int, Vector3f > my_map_vec3;
668 * \endcode
669 *
670 * \sa \ref TopicStlContainers.
671 */
672 template<class T>
673 class aligned_allocator
674 {
675 public:
676     typedef size_t    size_type;
677     typedef std::ptrdiff_t difference_type;
678     typedef T*        pointer;
679     typedef const T*  const_pointer;
680     typedef T&        reference;
681     typedef const T&  const_reference;
682     typedef T         value_type;
683 
684     template<class U>
685     struct rebind
686     {
687         typedef aligned_allocator<U> other;
688     };
689 
690     pointer address( reference value ) const
691     {
692         return &value;
693     }
694 
695     const_pointer address( const_reference value ) const
696     {
697         return &value;
698     }
699 
700     aligned_allocator()
701     {
702     }
703 
704     aligned_allocator( const aligned_allocator& )
705     {
706     }
707 
708     template<class U>
709     aligned_allocator( const aligned_allocator<U>& )
710     {
711     }
712 
713     ~aligned_allocator()
714     {
715     }
716 
717     size_type max_size() const
718     {
719         return (std::numeric_limits<size_type>::max)();
720     }
721 
722     pointer allocate( size_type num, const void* hint = 0 )
723     {
724         EIGEN_UNUSED_VARIABLE(hint);
725         internal::check_size_for_overflow<T>(num);
726         return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
727     }
728 
729     void construct( pointer p, const T& value )
730     {
731         ::new( p ) T( value );
732     }
733 
734     void destroy( pointer p )
735     {
736         p->~T();
737     }
738 
739     void deallocate( pointer p, size_type /*num*/ )
740     {
741         internal::aligned_free( p );
742     }
743 
744     bool operator!=(const aligned_allocator<T>& ) const
745     { return false; }
746 
747     bool operator==(const aligned_allocator<T>& ) const
748     { return true; }
749 };
750 
751 //---------- Cache sizes ----------
752 
753 #if !defined(EIGEN_NO_CPUID)
754 #  if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
755 #    if defined(__PIC__) && defined(__i386__)
756        // Case for x86 with PIC
757 #      define EIGEN_CPUID(abcd,func,id) \
758          __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));
759 #    elif defined(__PIC__) && defined(__x86_64__)
760        // 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.
761        // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
762 #      define EIGEN_CPUID(abcd,func,id) \
763         __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));
764 #    else
765        // Case for x86_64 or x86 w/o PIC
766 #      define EIGEN_CPUID(abcd,func,id) \
767          __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
768 #    endif
769 #  elif defined(_MSC_VER)
770 #    if (_MSC_VER > 1500) && ( defined(_M_IX86) || defined(_M_X64) )
771 #      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
772 #    endif
773 #  endif
774 #endif
775 
776 namespace internal {
777 
778 #ifdef EIGEN_CPUID
779 
780 inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
781 {
782   return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
783 }
784 
785 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
786 {
787   int abcd[4];
788   l1 = l2 = l3 = 0;
789   int cache_id = 0;
790   int cache_type = 0;
791   do {
792     abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
793     EIGEN_CPUID(abcd,0x4,cache_id);
794     cache_type  = (abcd[0] & 0x0F) >> 0;
795     if(cache_type==1||cache_type==3) // data or unified cache
796     {
797       int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
798       int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
799       int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
800       int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
801       int sets        = (abcd[2]);                    // C[31:0]
802 
803       int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
804 
805       switch(cache_level)
806       {
807         case 1: l1 = cache_size; break;
808         case 2: l2 = cache_size; break;
809         case 3: l3 = cache_size; break;
810         default: break;
811       }
812     }
813     cache_id++;
814   } while(cache_type>0 && cache_id<16);
815 }
816 
817 inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
818 {
819   int abcd[4];
820   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
821   l1 = l2 = l3 = 0;
822   EIGEN_CPUID(abcd,0x00000002,0);
823   unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
824   bool check_for_p2_core2 = false;
825   for(int i=0; i<14; ++i)
826   {
827     switch(bytes[i])
828     {
829       case 0x0A: l1 = 8; break;   // 0Ah   data L1 cache, 8 KB, 2 ways, 32 byte lines
830       case 0x0C: l1 = 16; break;  // 0Ch   data L1 cache, 16 KB, 4 ways, 32 byte lines
831       case 0x0E: l1 = 24; break;  // 0Eh   data L1 cache, 24 KB, 6 ways, 64 byte lines
832       case 0x10: l1 = 16; break;  // 10h   data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
833       case 0x15: l1 = 16; break;  // 15h   code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
834       case 0x2C: l1 = 32; break;  // 2Ch   data L1 cache, 32 KB, 8 ways, 64 byte lines
835       case 0x30: l1 = 32; break;  // 30h   code L1 cache, 32 KB, 8 ways, 64 byte lines
836       case 0x60: l1 = 16; break;  // 60h   data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
837       case 0x66: l1 = 8; break;   // 66h   data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
838       case 0x67: l1 = 16; break;  // 67h   data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
839       case 0x68: l1 = 32; break;  // 68h   data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
840       case 0x1A: l2 = 96; break;   // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
841       case 0x22: l3 = 512; break;   // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
842       case 0x23: l3 = 1024; break;   // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
843       case 0x25: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
844       case 0x29: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
845       case 0x39: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
846       case 0x3A: l2 = 192; break;   // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
847       case 0x3B: l2 = 128; break;   // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
848       case 0x3C: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
849       case 0x3D: l2 = 384; break;   // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
850       case 0x3E: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
851       case 0x40: l2 = 0; break;   // no integrated L2 cache (P6 core) or L3 cache (P4 core)
852       case 0x41: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
853       case 0x42: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
854       case 0x43: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
855       case 0x44: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
856       case 0x45: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
857       case 0x46: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
858       case 0x47: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
859       case 0x48: l2 = 3072; break;   // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
860       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
861       case 0x4A: l3 = 6144; break;   // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
862       case 0x4B: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
863       case 0x4C: l3 = 12288; break;   // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
864       case 0x4D: l3 = 16384; break;   // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
865       case 0x4E: l2 = 6144; break;   // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
866       case 0x78: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
867       case 0x79: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
868       case 0x7A: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
869       case 0x7B: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
870       case 0x7C: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
871       case 0x7D: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
872       case 0x7E: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
873       case 0x7F: l2 = 512; break;   // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
874       case 0x80: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
875       case 0x81: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
876       case 0x82: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
877       case 0x83: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
878       case 0x84: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
879       case 0x85: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
880       case 0x86: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
881       case 0x87: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
882       case 0x88: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
883       case 0x89: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
884       case 0x8A: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
885       case 0x8D: l3 = 3072; break;   // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
886 
887       default: break;
888     }
889   }
890   if(check_for_p2_core2 && l2 == l3)
891     l3 = 0;
892   l1 *= 1024;
893   l2 *= 1024;
894   l3 *= 1024;
895 }
896 
897 inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
898 {
899   if(max_std_funcs>=4)
900     queryCacheSizes_intel_direct(l1,l2,l3);
901   else
902     queryCacheSizes_intel_codes(l1,l2,l3);
903 }
904 
905 inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
906 {
907   int abcd[4];
908   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
909   EIGEN_CPUID(abcd,0x80000005,0);
910   l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
911   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
912   EIGEN_CPUID(abcd,0x80000006,0);
913   l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
914   l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
915 }
916 #endif
917 
918 /** \internal
919  * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
920 inline void queryCacheSizes(int& l1, int& l2, int& l3)
921 {
922   #ifdef EIGEN_CPUID
923   int abcd[4];
924   const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
925   const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
926   const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
927 
928   // identify the CPU vendor
929   EIGEN_CPUID(abcd,0x0,0);
930   int max_std_funcs = abcd[1];
931   if(cpuid_is_vendor(abcd,GenuineIntel))
932     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
933   else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
934     queryCacheSizes_amd(l1,l2,l3);
935   else
936     // by default let's use Intel's API
937     queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
938 
939   // here is the list of other vendors:
940 //   ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
941 //   ||cpuid_is_vendor(abcd,"CyrixInstead")
942 //   ||cpuid_is_vendor(abcd,"CentaurHauls")
943 //   ||cpuid_is_vendor(abcd,"GenuineTMx86")
944 //   ||cpuid_is_vendor(abcd,"TransmetaCPU")
945 //   ||cpuid_is_vendor(abcd,"RiseRiseRise")
946 //   ||cpuid_is_vendor(abcd,"Geode by NSC")
947 //   ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
948 //   ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
949 //   ||cpuid_is_vendor(abcd,"NexGenDriven")
950   #else
951   l1 = l2 = l3 = -1;
952   #endif
953 }
954 
955 /** \internal
956  * \returns the size in Bytes of the L1 data cache */
957 inline int queryL1CacheSize()
958 {
959   int l1(-1), l2, l3;
960   queryCacheSizes(l1,l2,l3);
961   return l1;
962 }
963 
964 /** \internal
965  * \returns the size in Bytes of the L2 or L3 cache if this later is present */
966 inline int queryTopLevelCacheSize()
967 {
968   int l1, l2(-1), l3(-1);
969   queryCacheSizes(l1,l2,l3);
970   return (std::max)(l2,l3);
971 }
972 
973 } // end namespace internal
974 
975 } // end namespace Eigen
976 
977 #endif // EIGEN_MEMORY_H
978