17 #include "kmp_config.h"
21 #define KMP_FTN_PLAIN 1
22 #define KMP_FTN_APPEND 2
23 #define KMP_FTN_UPPER 3
29 #define KMP_PTR_SKIP (sizeof(void *))
36 #define KMP_MEM_CONS_VOLATILE 0
37 #define KMP_MEM_CONS_FENCE 1
39 #ifndef KMP_MEM_CONS_MODEL
40 #define KMP_MEM_CONS_MODEL KMP_MEM_CONS_VOLATILE
44 #define KMP_COMPILER_ICC 0
45 #define KMP_COMPILER_GCC 0
46 #define KMP_COMPILER_CLANG 0
47 #define KMP_COMPILER_MSVC 0
49 #if defined(__INTEL_COMPILER)
50 #undef KMP_COMPILER_ICC
51 #define KMP_COMPILER_ICC 1
52 #elif defined(__clang__)
53 #undef KMP_COMPILER_CLANG
54 #define KMP_COMPILER_CLANG 1
55 #elif defined(__GNUC__)
56 #undef KMP_COMPILER_GCC
57 #define KMP_COMPILER_GCC 1
58 #elif defined(_MSC_VER)
59 #undef KMP_COMPILER_MSVC
60 #define KMP_COMPILER_MSVC 1
62 #error Unknown compiler
65 #if (KMP_OS_LINUX || KMP_OS_WINDOWS) && !KMP_OS_CNK
66 #define KMP_AFFINITY_SUPPORTED 1
67 #if KMP_OS_WINDOWS && KMP_ARCH_X86_64
68 #define KMP_GROUP_AFFINITY 1
70 #define KMP_GROUP_AFFINITY 0
73 #define KMP_AFFINITY_SUPPORTED 0
74 #define KMP_GROUP_AFFINITY 0
78 #define KMP_HAVE_QUAD 0
79 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
83 #define KMP_HAVE_QUAD 1
84 #elif KMP_COMPILER_CLANG
87 typedef long double _Quad;
88 #elif KMP_COMPILER_GCC
89 typedef __float128 _Quad;
91 #define KMP_HAVE_QUAD 1
92 #elif KMP_COMPILER_MSVC
93 typedef long double _Quad;
96 #if __LDBL_MAX_EXP__ >= 16384 && KMP_COMPILER_GCC
97 typedef long double _Quad;
99 #define KMP_HAVE_QUAD 1
104 typedef char kmp_int8;
105 typedef unsigned char kmp_uint8;
106 typedef short kmp_int16;
107 typedef unsigned short kmp_uint16;
108 typedef int kmp_int32;
109 typedef unsigned int kmp_uint32;
110 #define KMP_INT32_SPEC "d"
111 #define KMP_UINT32_SPEC "u"
113 typedef __int64 kmp_int64;
114 typedef unsigned __int64 kmp_uint64;
115 #define KMP_INT64_SPEC "I64d"
116 #define KMP_UINT64_SPEC "I64u"
118 struct kmp_struct64 {
121 typedef struct kmp_struct64 kmp_int64;
122 typedef struct kmp_struct64 kmp_uint64;
127 typedef __int64 kmp_intptr_t;
128 typedef unsigned __int64 kmp_uintptr_t;
129 #define KMP_INTPTR_SPEC "I64d"
130 #define KMP_UINTPTR_SPEC "I64u"
135 typedef char kmp_int8;
136 typedef unsigned char kmp_uint8;
137 typedef short kmp_int16;
138 typedef unsigned short kmp_uint16;
139 typedef int kmp_int32;
140 typedef unsigned int kmp_uint32;
141 typedef long long kmp_int64;
142 typedef unsigned long long kmp_uint64;
143 #define KMP_INT32_SPEC "d"
144 #define KMP_UINT32_SPEC "u"
145 #define KMP_INT64_SPEC "lld"
146 #define KMP_UINT64_SPEC "llu"
149 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS
150 #define KMP_SIZE_T_SPEC KMP_UINT32_SPEC
151 #elif KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
152 #define KMP_SIZE_T_SPEC KMP_UINT64_SPEC
154 #error "Can't determine size_t printf format specifier."
158 #define KMP_SIZE_T_MAX (0xFFFFFFFF)
160 #define KMP_SIZE_T_MAX (0xFFFFFFFFFFFFFFFF)
163 typedef size_t kmp_size_t;
164 typedef float kmp_real32;
165 typedef double kmp_real64;
169 typedef long kmp_intptr_t;
170 typedef unsigned long kmp_uintptr_t;
171 #define KMP_INTPTR_SPEC "ld"
172 #define KMP_UINTPTR_SPEC "lu"
176 typedef kmp_int64 kmp_int;
177 typedef kmp_uint64 kmp_uint;
179 typedef kmp_int32 kmp_int;
180 typedef kmp_uint32 kmp_uint;
182 #define KMP_INT_MAX ((kmp_int32)0x7FFFFFFF)
183 #define KMP_INT_MIN ((kmp_int32)0x80000000)
187 #define CCAST(type, var) const_cast<type>(var)
188 #define RCAST(type, var) reinterpret_cast<type>(var)
192 template <
typename T>
struct traits_t {};
194 template <>
struct traits_t<signed int> {
195 typedef signed int signed_t;
196 typedef unsigned int unsigned_t;
197 typedef double floating_t;
198 static char const *spec;
199 static const signed_t max_value = 0x7fffffff;
200 static const signed_t min_value = 0x80000000;
201 static const int type_size =
sizeof(signed_t);
204 template <>
struct traits_t<unsigned int> {
205 typedef signed int signed_t;
206 typedef unsigned int unsigned_t;
207 typedef double floating_t;
208 static char const *spec;
209 static const unsigned_t max_value = 0xffffffff;
210 static const unsigned_t min_value = 0x00000000;
211 static const int type_size =
sizeof(unsigned_t);
214 template <>
struct traits_t<signed long> {
215 typedef signed long signed_t;
216 typedef unsigned long unsigned_t;
217 typedef long double floating_t;
218 static char const *spec;
219 static const int type_size =
sizeof(signed_t);
222 template <>
struct traits_t<signed long long> {
223 typedef signed long long signed_t;
224 typedef unsigned long long unsigned_t;
225 typedef long double floating_t;
226 static char const *spec;
227 static const signed_t max_value = 0x7fffffffffffffffLL;
228 static const signed_t min_value = 0x8000000000000000LL;
229 static const int type_size =
sizeof(signed_t);
232 template <>
struct traits_t<unsigned long long> {
233 typedef signed long long signed_t;
234 typedef unsigned long long unsigned_t;
235 typedef long double floating_t;
236 static char const *spec;
237 static const unsigned_t max_value = 0xffffffffffffffffLL;
238 static const unsigned_t min_value = 0x0000000000000000LL;
239 static const int type_size =
sizeof(unsigned_t);
243 #define CCAST(type, var) (type)(var)
244 #define RCAST(type, var) (type)(var)
245 #endif // __cplusplus
247 #define KMP_EXPORT extern
250 #define __forceinline __inline
256 static inline int KMP_GET_PAGE_SIZE(
void) {
259 return si.dwPageSize;
262 #define KMP_GET_PAGE_SIZE() getpagesize()
265 #define PAGE_ALIGNED(_addr) \
266 (!((size_t)_addr & (size_t)(KMP_GET_PAGE_SIZE() - 1)))
267 #define ALIGN_TO_PAGE(x) \
268 (void *)(((size_t)(x)) & ~((size_t)(KMP_GET_PAGE_SIZE() - 1)))
274 #endif // __cplusplus
276 #define INTERNODE_CACHE_LINE 4096
280 #define CACHE_LINE 128
282 #if (CACHE_LINE < 64) && !defined(KMP_OS_DARWIN)
284 #warning CACHE_LINE is too small.
288 #define KMP_CACHE_PREFETCH(ADDR)
291 #if __cplusplus >= 201103L
292 #define KMP_NORETURN [[noreturn]]
294 #define KMP_NORETURN __declspec(noreturn)
296 #define KMP_NORETURN __attribute__((noreturn))
300 #define KMP_ALIGN(bytes) __declspec(align(bytes))
301 #define KMP_THREAD_LOCAL __declspec(thread)
304 #define KMP_ALIGN(bytes) __attribute__((aligned(bytes)))
305 #define KMP_THREAD_LOCAL __thread
306 #define KMP_ALIAS(alias_of) __attribute__((alias(alias_of)))
309 #if KMP_HAVE_WEAK_ATTRIBUTE
310 #define KMP_WEAK_ATTRIBUTE __attribute__((weak))
312 #define KMP_WEAK_ATTRIBUTE
317 #define KMP_STR(x) _KMP_STR(x)
318 #define _KMP_STR(x) #x
321 #ifdef KMP_USE_VERSION_SYMBOLS
324 #define KMP_EXPAND_NAME(api_name) _KMP_EXPAND_NAME(api_name)
325 #define _KMP_EXPAND_NAME(api_name) __kmp_api_##api_name
326 #define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str) \
327 _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, "VERSION")
328 #define _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, default_ver) \
329 __typeof__(__kmp_api_##api_name) __kmp_api_##api_name##_##ver_num##_alias \
330 __attribute__((alias(KMP_STR(__kmp_api_##api_name)))); \
332 ".symver " KMP_STR(__kmp_api_##api_name##_##ver_num##_alias) "," KMP_STR( \
333 api_name) "@" ver_str "\n\t"); \
334 __asm__(".symver " KMP_STR(__kmp_api_##api_name) "," KMP_STR( \
335 api_name) "@@" default_ver "\n\t")
336 #else // KMP_USE_VERSION_SYMBOLS
337 #define KMP_EXPAND_NAME(api_name) api_name
338 #define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str)
339 #endif // KMP_USE_VERSION_SYMBOLS
343 #define KMP_DO_ALIGN(bytes) KMP_ALIGN(bytes)
344 #define KMP_ALIGN_CACHE KMP_ALIGN(CACHE_LINE)
345 #define KMP_ALIGN_CACHE_INTERNODE KMP_ALIGN(INTERNODE_CACHE_LINE)
348 enum kmp_mem_fence_type {
357 #if KMP_ASM_INTRINS && KMP_OS_WINDOWS
359 #pragma intrinsic(InterlockedExchangeAdd)
360 #pragma intrinsic(InterlockedCompareExchange)
361 #pragma intrinsic(InterlockedExchange)
362 #pragma intrinsic(InterlockedExchange64)
366 #define KMP_TEST_THEN_INC32(p) InterlockedExchangeAdd((volatile long *)(p), 1)
367 #define KMP_TEST_THEN_INC_ACQ32(p) \
368 InterlockedExchangeAdd((volatile long *)(p), 1)
369 #define KMP_TEST_THEN_ADD4_32(p) InterlockedExchangeAdd((volatile long *)(p), 4)
370 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
371 InterlockedExchangeAdd((volatile long *)(p), 4)
372 #define KMP_TEST_THEN_DEC32(p) InterlockedExchangeAdd((volatile long *)(p), -1)
373 #define KMP_TEST_THEN_DEC_ACQ32(p) \
374 InterlockedExchangeAdd((volatile long *)(p), -1)
375 #define KMP_TEST_THEN_ADD32(p, v) \
376 InterlockedExchangeAdd((volatile long *)(p), (v))
378 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
379 InterlockedCompareExchange((volatile long *)(p), (long)(sv), (long)(cv))
381 #define KMP_XCHG_FIXED32(p, v) \
382 InterlockedExchange((volatile long *)(p), (long)(v))
383 #define KMP_XCHG_FIXED64(p, v) \
384 InterlockedExchange64((volatile kmp_int64 *)(p), (kmp_int64)(v))
386 inline kmp_real32 KMP_XCHG_REAL32(
volatile kmp_real32 *p, kmp_real32 v) {
387 kmp_int32 tmp = InterlockedExchange((
volatile long *)p, *(
long *)&v);
388 return *(kmp_real32 *)&tmp;
392 extern kmp_int8 __kmp_test_then_add8(
volatile kmp_int8 *p, kmp_int8 v);
393 extern kmp_int8 __kmp_test_then_or8(
volatile kmp_int8 *p, kmp_int8 v);
394 extern kmp_int8 __kmp_test_then_and8(
volatile kmp_int8 *p, kmp_int8 v);
395 extern kmp_int32 __kmp_test_then_add32(
volatile kmp_int32 *p, kmp_int32 v);
396 extern kmp_uint32 __kmp_test_then_or32(
volatile kmp_uint32 *p, kmp_uint32 v);
397 extern kmp_uint32 __kmp_test_then_and32(
volatile kmp_uint32 *p, kmp_uint32 v);
398 extern kmp_int64 __kmp_test_then_add64(
volatile kmp_int64 *p, kmp_int64 v);
399 extern kmp_uint64 __kmp_test_then_or64(
volatile kmp_uint64 *p, kmp_uint64 v);
400 extern kmp_uint64 __kmp_test_then_and64(
volatile kmp_uint64 *p, kmp_uint64 v);
402 extern kmp_int8 __kmp_compare_and_store8(
volatile kmp_int8 *p, kmp_int8 cv,
404 extern kmp_int16 __kmp_compare_and_store16(
volatile kmp_int16 *p, kmp_int16 cv,
406 extern kmp_int32 __kmp_compare_and_store32(
volatile kmp_int32 *p, kmp_int32 cv,
408 extern kmp_int32 __kmp_compare_and_store64(
volatile kmp_int64 *p, kmp_int64 cv,
410 extern kmp_int8 __kmp_compare_and_store_ret8(
volatile kmp_int8 *p, kmp_int8 cv,
412 extern kmp_int16 __kmp_compare_and_store_ret16(
volatile kmp_int16 *p,
413 kmp_int16 cv, kmp_int16 sv);
414 extern kmp_int32 __kmp_compare_and_store_ret32(
volatile kmp_int32 *p,
415 kmp_int32 cv, kmp_int32 sv);
416 extern kmp_int64 __kmp_compare_and_store_ret64(
volatile kmp_int64 *p,
417 kmp_int64 cv, kmp_int64 sv);
419 extern kmp_int8 __kmp_xchg_fixed8(
volatile kmp_int8 *p, kmp_int8 v);
420 extern kmp_int16 __kmp_xchg_fixed16(
volatile kmp_int16 *p, kmp_int16 v);
421 extern kmp_int32 __kmp_xchg_fixed32(
volatile kmp_int32 *p, kmp_int32 v);
422 extern kmp_int64 __kmp_xchg_fixed64(
volatile kmp_int64 *p, kmp_int64 v);
423 extern kmp_real32 __kmp_xchg_real32(
volatile kmp_real32 *p, kmp_real32 v);
424 extern kmp_real64 __kmp_xchg_real64(
volatile kmp_real64 *p, kmp_real64 v);
428 #define KMP_TEST_THEN_INC64(p) __kmp_test_then_add64((p), 1LL)
429 #define KMP_TEST_THEN_INC_ACQ64(p) __kmp_test_then_add64((p), 1LL)
432 #define KMP_TEST_THEN_ADD4_64(p) __kmp_test_then_add64((p), 4LL)
433 #define KMP_TEST_THEN_ADD4_ACQ64(p) __kmp_test_then_add64((p), 4LL)
436 #define KMP_TEST_THEN_DEC64(p) __kmp_test_then_add64((p), -1LL)
437 #define KMP_TEST_THEN_DEC_ACQ64(p) __kmp_test_then_add64((p), -1LL)
439 #define KMP_TEST_THEN_ADD8(p, v) __kmp_test_then_add8((p), (v))
440 #define KMP_TEST_THEN_ADD64(p, v) __kmp_test_then_add64((p), (v))
442 #define KMP_TEST_THEN_OR8(p, v) __kmp_test_then_or8((p), (v))
443 #define KMP_TEST_THEN_AND8(p, v) __kmp_test_then_and8((p), (v))
444 #define KMP_TEST_THEN_OR32(p, v) __kmp_test_then_or32((p), (v))
445 #define KMP_TEST_THEN_AND32(p, v) __kmp_test_then_and32((p), (v))
446 #define KMP_TEST_THEN_OR64(p, v) __kmp_test_then_or64((p), (v))
447 #define KMP_TEST_THEN_AND64(p, v) __kmp_test_then_and64((p), (v))
449 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
450 __kmp_compare_and_store8((p), (cv), (sv))
451 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
452 __kmp_compare_and_store8((p), (cv), (sv))
453 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
454 __kmp_compare_and_store16((p), (cv), (sv))
455 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
456 __kmp_compare_and_store16((p), (cv), (sv))
457 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
458 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
460 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
461 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
463 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
464 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
466 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
467 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
471 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
472 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
475 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
476 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
480 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
481 __kmp_compare_and_store_ret8((p), (cv), (sv))
482 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
483 __kmp_compare_and_store_ret16((p), (cv), (sv))
484 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
485 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
488 #define KMP_XCHG_FIXED8(p, v) \
489 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
490 #define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
494 #define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
496 #elif (KMP_ASM_INTRINS && KMP_OS_UNIX) || !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
499 #define KMP_TEST_THEN_INC32(p) \
500 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
501 #define KMP_TEST_THEN_INC_ACQ32(p) \
502 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
503 #define KMP_TEST_THEN_INC64(p) \
504 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
505 #define KMP_TEST_THEN_INC_ACQ64(p) \
506 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
507 #define KMP_TEST_THEN_ADD4_32(p) \
508 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
509 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
510 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
511 #define KMP_TEST_THEN_ADD4_64(p) \
512 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
513 #define KMP_TEST_THEN_ADD4_ACQ64(p) \
514 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
515 #define KMP_TEST_THEN_DEC32(p) \
516 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
517 #define KMP_TEST_THEN_DEC_ACQ32(p) \
518 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
519 #define KMP_TEST_THEN_DEC64(p) \
520 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
521 #define KMP_TEST_THEN_DEC_ACQ64(p) \
522 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
523 #define KMP_TEST_THEN_ADD8(p, v) \
524 __sync_fetch_and_add((volatile kmp_int8 *)(p), (kmp_int8)(v))
525 #define KMP_TEST_THEN_ADD32(p, v) \
526 __sync_fetch_and_add((volatile kmp_int32 *)(p), (kmp_int32)(v))
527 #define KMP_TEST_THEN_ADD64(p, v) \
528 __sync_fetch_and_add((volatile kmp_int64 *)(p), (kmp_int64)(v))
530 #define KMP_TEST_THEN_OR8(p, v) \
531 __sync_fetch_and_or((volatile kmp_int8 *)(p), (kmp_int8)(v))
532 #define KMP_TEST_THEN_AND8(p, v) \
533 __sync_fetch_and_and((volatile kmp_int8 *)(p), (kmp_int8)(v))
534 #define KMP_TEST_THEN_OR32(p, v) \
535 __sync_fetch_and_or((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
536 #define KMP_TEST_THEN_AND32(p, v) \
537 __sync_fetch_and_and((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
538 #define KMP_TEST_THEN_OR64(p, v) \
539 __sync_fetch_and_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
540 #define KMP_TEST_THEN_AND64(p, v) \
541 __sync_fetch_and_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
543 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
544 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
546 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
547 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
549 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
550 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
552 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
553 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
555 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
556 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
558 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
559 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
561 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
562 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
564 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
565 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
567 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
568 __sync_bool_compare_and_swap((void *volatile *)(p), (void *)(cv), \
571 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
572 __sync_val_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
574 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
575 __sync_val_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
577 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
578 __sync_val_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
580 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
581 __sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
584 #define KMP_XCHG_FIXED8(p, v) \
585 __sync_lock_test_and_set((volatile kmp_uint8 *)(p), (kmp_uint8)(v))
586 #define KMP_XCHG_FIXED16(p, v) \
587 __sync_lock_test_and_set((volatile kmp_uint16 *)(p), (kmp_uint16)(v))
588 #define KMP_XCHG_FIXED32(p, v) \
589 __sync_lock_test_and_set((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
590 #define KMP_XCHG_FIXED64(p, v) \
591 __sync_lock_test_and_set((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
593 inline kmp_real32 KMP_XCHG_REAL32(
volatile kmp_real32 *p, kmp_real32 v) {
595 __sync_lock_test_and_set((
volatile kmp_uint32 *)(p), *(kmp_uint32 *)&v);
596 return *(kmp_real32 *)&tmp;
599 inline kmp_real64 KMP_XCHG_REAL64(
volatile kmp_real64 *p, kmp_real64 v) {
601 __sync_lock_test_and_set((
volatile kmp_uint64 *)(p), *(kmp_uint64 *)&v);
602 return *(kmp_real64 *)&tmp;
607 extern kmp_int8 __kmp_test_then_add8(
volatile kmp_int8 *p, kmp_int8 v);
608 extern kmp_int8 __kmp_test_then_or8(
volatile kmp_int8 *p, kmp_int8 v);
609 extern kmp_int8 __kmp_test_then_and8(
volatile kmp_int8 *p, kmp_int8 v);
610 extern kmp_int32 __kmp_test_then_add32(
volatile kmp_int32 *p, kmp_int32 v);
611 extern kmp_uint32 __kmp_test_then_or32(
volatile kmp_uint32 *p, kmp_uint32 v);
612 extern kmp_uint32 __kmp_test_then_and32(
volatile kmp_uint32 *p, kmp_uint32 v);
613 extern kmp_int64 __kmp_test_then_add64(
volatile kmp_int64 *p, kmp_int64 v);
614 extern kmp_uint64 __kmp_test_then_or64(
volatile kmp_uint64 *p, kmp_uint64 v);
615 extern kmp_uint64 __kmp_test_then_and64(
volatile kmp_uint64 *p, kmp_uint64 v);
617 extern kmp_int8 __kmp_compare_and_store8(
volatile kmp_int8 *p, kmp_int8 cv,
619 extern kmp_int16 __kmp_compare_and_store16(
volatile kmp_int16 *p, kmp_int16 cv,
621 extern kmp_int32 __kmp_compare_and_store32(
volatile kmp_int32 *p, kmp_int32 cv,
623 extern kmp_int32 __kmp_compare_and_store64(
volatile kmp_int64 *p, kmp_int64 cv,
625 extern kmp_int8 __kmp_compare_and_store_ret8(
volatile kmp_int8 *p, kmp_int8 cv,
627 extern kmp_int16 __kmp_compare_and_store_ret16(
volatile kmp_int16 *p,
628 kmp_int16 cv, kmp_int16 sv);
629 extern kmp_int32 __kmp_compare_and_store_ret32(
volatile kmp_int32 *p,
630 kmp_int32 cv, kmp_int32 sv);
631 extern kmp_int64 __kmp_compare_and_store_ret64(
volatile kmp_int64 *p,
632 kmp_int64 cv, kmp_int64 sv);
634 extern kmp_int8 __kmp_xchg_fixed8(
volatile kmp_int8 *p, kmp_int8 v);
635 extern kmp_int16 __kmp_xchg_fixed16(
volatile kmp_int16 *p, kmp_int16 v);
636 extern kmp_int32 __kmp_xchg_fixed32(
volatile kmp_int32 *p, kmp_int32 v);
637 extern kmp_int64 __kmp_xchg_fixed64(
volatile kmp_int64 *p, kmp_int64 v);
638 extern kmp_real32 __kmp_xchg_real32(
volatile kmp_real32 *p, kmp_real32 v);
639 extern kmp_real64 __kmp_xchg_real64(
volatile kmp_real64 *p, kmp_real64 v);
641 #define KMP_TEST_THEN_INC32(p) \
642 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
643 #define KMP_TEST_THEN_INC_ACQ32(p) \
644 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
645 #define KMP_TEST_THEN_INC64(p) \
646 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
647 #define KMP_TEST_THEN_INC_ACQ64(p) \
648 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
649 #define KMP_TEST_THEN_ADD4_32(p) \
650 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
651 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
652 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
653 #define KMP_TEST_THEN_ADD4_64(p) \
654 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
655 #define KMP_TEST_THEN_ADD4_ACQ64(p) \
656 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
657 #define KMP_TEST_THEN_DEC32(p) \
658 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
659 #define KMP_TEST_THEN_DEC_ACQ32(p) \
660 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
661 #define KMP_TEST_THEN_DEC64(p) \
662 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
663 #define KMP_TEST_THEN_DEC_ACQ64(p) \
664 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
665 #define KMP_TEST_THEN_ADD8(p, v) \
666 __kmp_test_then_add8((volatile kmp_int8 *)(p), (kmp_int8)(v))
667 #define KMP_TEST_THEN_ADD32(p, v) \
668 __kmp_test_then_add32((volatile kmp_int32 *)(p), (kmp_int32)(v))
669 #define KMP_TEST_THEN_ADD64(p, v) \
670 __kmp_test_then_add64((volatile kmp_int64 *)(p), (kmp_int64)(v))
672 #define KMP_TEST_THEN_OR8(p, v) \
673 __kmp_test_then_or8((volatile kmp_int8 *)(p), (kmp_int8)(v))
674 #define KMP_TEST_THEN_AND8(p, v) \
675 __kmp_test_then_and8((volatile kmp_int8 *)(p), (kmp_int8)(v))
676 #define KMP_TEST_THEN_OR32(p, v) \
677 __kmp_test_then_or32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
678 #define KMP_TEST_THEN_AND32(p, v) \
679 __kmp_test_then_and32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
680 #define KMP_TEST_THEN_OR64(p, v) \
681 __kmp_test_then_or64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
682 #define KMP_TEST_THEN_AND64(p, v) \
683 __kmp_test_then_and64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
685 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
686 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
688 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
689 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
691 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
692 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
694 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
695 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
697 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
698 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
700 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
701 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
703 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
704 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
706 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
707 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
711 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
712 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
715 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
716 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
720 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
721 __kmp_compare_and_store_ret8((p), (cv), (sv))
722 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
723 __kmp_compare_and_store_ret16((p), (cv), (sv))
724 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
725 __kmp_compare_and_store_ret32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
727 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
728 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
731 #define KMP_XCHG_FIXED8(p, v) \
732 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
733 #define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
734 #define KMP_XCHG_FIXED32(p, v) __kmp_xchg_fixed32((p), (v));
735 #define KMP_XCHG_FIXED64(p, v) __kmp_xchg_fixed64((p), (v));
736 #define KMP_XCHG_REAL32(p, v) __kmp_xchg_real32((p), (v));
737 #define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
745 #define KMP_MB() asm("nop")
746 #define KMP_IMB() asm("nop")
753 #if KMP_ARCH_PPC64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS || \
755 #define KMP_MB() __sync_synchronize()
767 #define KMP_ST_REL32(A, D) (*(A) = (D))
771 #define KMP_ST_REL64(A, D) (*(A) = (D))
775 #define KMP_LD_ACQ32(A) (*(A))
779 #define KMP_LD_ACQ64(A) (*(A))
795 #define TCW_1(a, b) (a) = (b)
797 #define TCW_4(a, b) (a) = (b)
798 #define TCI_4(a) (++(a))
799 #define TCD_4(a) (--(a))
801 #define TCW_8(a, b) (a) = (b)
802 #define TCI_8(a) (++(a))
803 #define TCD_8(a) (--(a))
804 #define TCR_SYNC_4(a) (a)
805 #define TCW_SYNC_4(a, b) (a) = (b)
806 #define TCX_SYNC_4(a, b, c) \
807 KMP_COMPARE_AND_STORE_REL32((volatile kmp_int32 *)(volatile void *)&(a), \
808 (kmp_int32)(b), (kmp_int32)(c))
809 #define TCR_SYNC_8(a) (a)
810 #define TCW_SYNC_8(a, b) (a) = (b)
811 #define TCX_SYNC_8(a, b, c) \
812 KMP_COMPARE_AND_STORE_REL64((volatile kmp_int64 *)(volatile void *)&(a), \
813 (kmp_int64)(b), (kmp_int64)(c))
815 #if KMP_ARCH_X86 || KMP_ARCH_MIPS
817 #define TCR_PTR(a) ((void *)TCR_4(a))
818 #define TCW_PTR(a, b) TCW_4((a), (b))
819 #define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_4(a))
820 #define TCW_SYNC_PTR(a, b) TCW_SYNC_4((a), (b))
821 #define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_4((a), (b), (c)))
825 #define TCR_PTR(a) ((void *)TCR_8(a))
826 #define TCW_PTR(a, b) TCW_8((a), (b))
827 #define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_8(a))
828 #define TCW_SYNC_PTR(a, b) TCW_SYNC_8((a), (b))
829 #define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_8((a), (b), (c)))
837 #define FTN_TRUE TRUE
841 #define FTN_FALSE FALSE
844 typedef void (*microtask_t)(
int *gtid,
int *npr, ...);
846 #ifdef USE_VOLATILE_CAST
847 #define VOLATILE_CAST(x) (volatile x)
849 #define VOLATILE_CAST(x) (x)
852 #define KMP_WAIT_YIELD __kmp_wait_yield_4
853 #define KMP_WAIT_YIELD_PTR __kmp_wait_yield_4_ptr
854 #define KMP_EQ __kmp_eq_4
855 #define KMP_NEQ __kmp_neq_4
856 #define KMP_LT __kmp_lt_4
857 #define KMP_GE __kmp_ge_4
858 #define KMP_LE __kmp_le_4
862 #if (KMP_ARCH_X86_64 || KMP_ARCH_PPC64) && KMP_OS_LINUX
863 #define STATIC_EFI2_WORKAROUND
865 #define STATIC_EFI2_WORKAROUND static
870 #define KMP_USE_BGET 1
874 #ifndef USE_CMPXCHG_FIX
875 #define USE_CMPXCHG_FIX 1
880 #define KMP_USE_DYNAMIC_LOCK 1
885 #if KMP_USE_DYNAMIC_LOCK
887 #define KMP_USE_TSX (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !KMP_COMPILER_MSVC
888 #ifdef KMP_USE_ADAPTIVE_LOCKS
889 #undef KMP_USE_ADAPTIVE_LOCKS
891 #define KMP_USE_ADAPTIVE_LOCKS KMP_USE_TSX
895 #if KMP_STATS_ENABLED
896 #define KMP_HAVE_TICK_TIME \
897 (KMP_OS_LINUX && (KMP_MIC || KMP_ARCH_X86 || KMP_ARCH_X86_64))
901 enum kmp_warnings_level {
902 kmp_warnings_off = 0,
904 kmp_warnings_explicit = 6,
910 #endif // __cplusplus
913 #define KMP_ATOMIC_LD(p, order) (p)->load(std::memory_order_##order)
914 #define KMP_ATOMIC_OP(op, p, v, order) (p)->op(v, std::memory_order_##order)
917 #define KMP_ATOMIC_LD_ACQ(p) KMP_ATOMIC_LD(p, acquire)
918 #define KMP_ATOMIC_LD_RLX(p) KMP_ATOMIC_LD(p, relaxed)
919 #define KMP_ATOMIC_ST_REL(p, v) KMP_ATOMIC_OP(store, p, v, release)
920 #define KMP_ATOMIC_ST_RLX(p, v) KMP_ATOMIC_OP(store, p, v, relaxed)
923 #define KMP_ATOMIC_ADD(p, v) KMP_ATOMIC_OP(fetch_add, p, v, acq_rel)
924 #define KMP_ATOMIC_SUB(p, v) KMP_ATOMIC_OP(fetch_sub, p, v, acq_rel)
925 #define KMP_ATOMIC_AND(p, v) KMP_ATOMIC_OP(fetch_and, p, v, acq_rel)
926 #define KMP_ATOMIC_OR(p, v) KMP_ATOMIC_OP(fetch_or, p, v, acq_rel)
927 #define KMP_ATOMIC_INC(p) KMP_ATOMIC_OP(fetch_add, p, 1, acq_rel)
928 #define KMP_ATOMIC_DEC(p) KMP_ATOMIC_OP(fetch_sub, p, 1, acq_rel)
929 #define KMP_ATOMIC_ADD_RLX(p, v) KMP_ATOMIC_OP(fetch_add, p, v, relaxed)
930 #define KMP_ATOMIC_INC_RLX(p) KMP_ATOMIC_OP(fetch_add, p, 1, relaxed)
933 template <
typename T>
934 bool __kmp_atomic_compare_store(std::atomic<T> *p, T expected, T desired) {
935 return p->compare_exchange_strong(
936 expected, desired, std::memory_order_acq_rel, std::memory_order_relaxed);
939 template <
typename T>
940 bool __kmp_atomic_compare_store_acq(std::atomic<T> *p, T expected, T desired) {
941 return p->compare_exchange_strong(
942 expected, desired, std::memory_order_acquire, std::memory_order_relaxed);
945 template <
typename T>
946 bool __kmp_atomic_compare_store_rel(std::atomic<T> *p, T expected, T desired) {
947 return p->compare_exchange_strong(
948 expected, desired, std::memory_order_release, std::memory_order_relaxed);
953 #include "kmp_safe_c_api.h"