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
91 typedef __float128 _Quad;
93 #define KMP_HAVE_QUAD 1
95 #elif KMP_COMPILER_MSVC
96 typedef long double _Quad;
99 #if __LDBL_MAX_EXP__ >= 16384 && KMP_COMPILER_GCC
100 typedef long double _Quad;
102 #define KMP_HAVE_QUAD 1
106 #define KMP_USE_X87CONTROL 0
108 #define KMP_END_OF_LINE "\r\n"
109 typedef char kmp_int8;
110 typedef unsigned char kmp_uint8;
111 typedef short kmp_int16;
112 typedef unsigned short kmp_uint16;
113 typedef int kmp_int32;
114 typedef unsigned int kmp_uint32;
115 #define KMP_INT32_SPEC "d"
116 #define KMP_UINT32_SPEC "u"
118 typedef __int64 kmp_int64;
119 typedef unsigned __int64 kmp_uint64;
120 #define KMP_INT64_SPEC "I64d"
121 #define KMP_UINT64_SPEC "I64u"
123 struct kmp_struct64 {
126 typedef struct kmp_struct64 kmp_int64;
127 typedef struct kmp_struct64 kmp_uint64;
130 #if KMP_ARCH_X86 && KMP_MSVC_COMPAT
131 #undef KMP_USE_X87CONTROL
132 #define KMP_USE_X87CONTROL 1
136 typedef __int64 kmp_intptr_t;
137 typedef unsigned __int64 kmp_uintptr_t;
138 #define KMP_INTPTR_SPEC "I64d"
139 #define KMP_UINTPTR_SPEC "I64u"
144 #define KMP_END_OF_LINE "\n"
145 typedef char kmp_int8;
146 typedef unsigned char kmp_uint8;
147 typedef short kmp_int16;
148 typedef unsigned short kmp_uint16;
149 typedef int kmp_int32;
150 typedef unsigned int kmp_uint32;
151 typedef long long kmp_int64;
152 typedef unsigned long long kmp_uint64;
153 #define KMP_INT32_SPEC "d"
154 #define KMP_UINT32_SPEC "u"
155 #define KMP_INT64_SPEC "lld"
156 #define KMP_UINT64_SPEC "llu"
159 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS
160 #define KMP_SIZE_T_SPEC KMP_UINT32_SPEC
161 #elif KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
162 #define KMP_SIZE_T_SPEC KMP_UINT64_SPEC
164 #error "Can't determine size_t printf format specifier."
168 #define KMP_SIZE_T_MAX (0xFFFFFFFF)
170 #define KMP_SIZE_T_MAX (0xFFFFFFFFFFFFFFFF)
173 typedef size_t kmp_size_t;
174 typedef float kmp_real32;
175 typedef double kmp_real64;
179 typedef long kmp_intptr_t;
180 typedef unsigned long kmp_uintptr_t;
181 #define KMP_INTPTR_SPEC "ld"
182 #define KMP_UINTPTR_SPEC "lu"
186 typedef kmp_int64 kmp_int;
187 typedef kmp_uint64 kmp_uint;
189 typedef kmp_int32 kmp_int;
190 typedef kmp_uint32 kmp_uint;
192 #define KMP_INT_MAX ((kmp_int32)0x7FFFFFFF)
193 #define KMP_INT_MIN ((kmp_int32)0x80000000)
197 #define CCAST(type, var) const_cast<type>(var)
198 #define RCAST(type, var) reinterpret_cast<type>(var)
202 template <
typename T>
struct traits_t {};
204 template <>
struct traits_t<signed 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 signed_t max_value = 0x7fffffff;
210 static const signed_t min_value = 0x80000000;
211 static const int type_size =
sizeof(signed_t);
214 template <>
struct traits_t<unsigned int> {
215 typedef signed int signed_t;
216 typedef unsigned int unsigned_t;
217 typedef double floating_t;
218 static char const *spec;
219 static const unsigned_t max_value = 0xffffffff;
220 static const unsigned_t min_value = 0x00000000;
221 static const int type_size =
sizeof(unsigned_t);
224 template <>
struct traits_t<signed long> {
225 typedef signed long signed_t;
226 typedef unsigned long unsigned_t;
227 typedef long double floating_t;
228 static char const *spec;
229 static const int type_size =
sizeof(signed_t);
232 template <>
struct traits_t<signed 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 signed_t max_value = 0x7fffffffffffffffLL;
238 static const signed_t min_value = 0x8000000000000000LL;
239 static const int type_size =
sizeof(signed_t);
242 template <>
struct traits_t<unsigned long long> {
243 typedef signed long long signed_t;
244 typedef unsigned long long unsigned_t;
245 typedef long double floating_t;
246 static char const *spec;
247 static const unsigned_t max_value = 0xffffffffffffffffLL;
248 static const unsigned_t min_value = 0x0000000000000000LL;
249 static const int type_size =
sizeof(unsigned_t);
253 #define CCAST(type, var) (type)(var)
254 #define RCAST(type, var) (type)(var)
255 #endif // __cplusplus
257 #define KMP_EXPORT extern
259 #if __GNUC__ >= 4 && !defined(__MINGW32__)
260 #define __forceinline __inline
266 static inline int KMP_GET_PAGE_SIZE(
void) {
269 return si.dwPageSize;
272 #define KMP_GET_PAGE_SIZE() getpagesize()
275 #define PAGE_ALIGNED(_addr) \
276 (!((size_t)_addr & (size_t)(KMP_GET_PAGE_SIZE() - 1)))
277 #define ALIGN_TO_PAGE(x) \
278 (void *)(((size_t)(x)) & ~((size_t)(KMP_GET_PAGE_SIZE() - 1)))
284 #endif // __cplusplus
286 #define INTERNODE_CACHE_LINE 4096
290 #define CACHE_LINE 128
292 #if (CACHE_LINE < 64) && !defined(KMP_OS_DARWIN)
294 #warning CACHE_LINE is too small.
298 #define KMP_CACHE_PREFETCH(ADDR)
301 #if __cplusplus >= 201103L
302 #define KMP_NORETURN [[noreturn]]
304 #define KMP_NORETURN __declspec(noreturn)
306 #define KMP_NORETURN __attribute__((noreturn))
309 #if KMP_OS_WINDOWS && KMP_MSVC_COMPAT
310 #define KMP_ALIGN(bytes) __declspec(align(bytes))
311 #define KMP_THREAD_LOCAL __declspec(thread)
314 #define KMP_ALIGN(bytes) __attribute__((aligned(bytes)))
315 #define KMP_THREAD_LOCAL __thread
316 #define KMP_ALIAS(alias_of) __attribute__((alias(alias_of)))
319 #if KMP_HAVE_WEAK_ATTRIBUTE
320 #define KMP_WEAK_ATTRIBUTE __attribute__((weak))
322 #define KMP_WEAK_ATTRIBUTE
327 #define KMP_STR(x) _KMP_STR(x)
328 #define _KMP_STR(x) #x
331 #ifdef KMP_USE_VERSION_SYMBOLS
334 #define KMP_EXPAND_NAME(api_name) _KMP_EXPAND_NAME(api_name)
335 #define _KMP_EXPAND_NAME(api_name) __kmp_api_##api_name
336 #define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str) \
337 _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, "VERSION")
338 #define _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, default_ver) \
339 __typeof__(__kmp_api_##api_name) __kmp_api_##api_name##_##ver_num##_alias \
340 __attribute__((alias(KMP_STR(__kmp_api_##api_name)))); \
342 ".symver " KMP_STR(__kmp_api_##api_name##_##ver_num##_alias) "," KMP_STR( \
343 api_name) "@" ver_str "\n\t"); \
344 __asm__(".symver " KMP_STR(__kmp_api_##api_name) "," KMP_STR( \
345 api_name) "@@" default_ver "\n\t")
346 #else // KMP_USE_VERSION_SYMBOLS
347 #define KMP_EXPAND_NAME(api_name) api_name
348 #define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str)
349 #endif // KMP_USE_VERSION_SYMBOLS
353 #define KMP_DO_ALIGN(bytes) KMP_ALIGN(bytes)
354 #define KMP_ALIGN_CACHE KMP_ALIGN(CACHE_LINE)
355 #define KMP_ALIGN_CACHE_INTERNODE KMP_ALIGN(INTERNODE_CACHE_LINE)
358 enum kmp_mem_fence_type {
367 #if KMP_ASM_INTRINS && KMP_OS_WINDOWS
369 #if KMP_MSVC_COMPAT && !KMP_COMPILER_CLANG
370 #pragma intrinsic(InterlockedExchangeAdd)
371 #pragma intrinsic(InterlockedCompareExchange)
372 #pragma intrinsic(InterlockedExchange)
373 #pragma intrinsic(InterlockedExchange64)
378 #define KMP_TEST_THEN_INC32(p) InterlockedExchangeAdd((volatile long *)(p), 1)
379 #define KMP_TEST_THEN_INC_ACQ32(p) \
380 InterlockedExchangeAdd((volatile long *)(p), 1)
381 #define KMP_TEST_THEN_ADD4_32(p) InterlockedExchangeAdd((volatile long *)(p), 4)
382 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
383 InterlockedExchangeAdd((volatile long *)(p), 4)
384 #define KMP_TEST_THEN_DEC32(p) InterlockedExchangeAdd((volatile long *)(p), -1)
385 #define KMP_TEST_THEN_DEC_ACQ32(p) \
386 InterlockedExchangeAdd((volatile long *)(p), -1)
387 #define KMP_TEST_THEN_ADD32(p, v) \
388 InterlockedExchangeAdd((volatile long *)(p), (v))
390 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
391 InterlockedCompareExchange((volatile long *)(p), (long)(sv), (long)(cv))
393 #define KMP_XCHG_FIXED32(p, v) \
394 InterlockedExchange((volatile long *)(p), (long)(v))
395 #define KMP_XCHG_FIXED64(p, v) \
396 InterlockedExchange64((volatile kmp_int64 *)(p), (kmp_int64)(v))
398 inline kmp_real32 KMP_XCHG_REAL32(
volatile kmp_real32 *p, kmp_real32 v) {
399 kmp_int32 tmp = InterlockedExchange((
volatile long *)p, *(
long *)&v);
400 return *(kmp_real32 *)&tmp;
404 extern kmp_int8 __kmp_test_then_add8(
volatile kmp_int8 *p, kmp_int8 v);
405 extern kmp_int8 __kmp_test_then_or8(
volatile kmp_int8 *p, kmp_int8 v);
406 extern kmp_int8 __kmp_test_then_and8(
volatile kmp_int8 *p, kmp_int8 v);
407 extern kmp_int32 __kmp_test_then_add32(
volatile kmp_int32 *p, kmp_int32 v);
408 extern kmp_uint32 __kmp_test_then_or32(
volatile kmp_uint32 *p, kmp_uint32 v);
409 extern kmp_uint32 __kmp_test_then_and32(
volatile kmp_uint32 *p, kmp_uint32 v);
410 extern kmp_int64 __kmp_test_then_add64(
volatile kmp_int64 *p, kmp_int64 v);
411 extern kmp_uint64 __kmp_test_then_or64(
volatile kmp_uint64 *p, kmp_uint64 v);
412 extern kmp_uint64 __kmp_test_then_and64(
volatile kmp_uint64 *p, kmp_uint64 v);
414 extern kmp_int8 __kmp_compare_and_store8(
volatile kmp_int8 *p, kmp_int8 cv,
416 extern kmp_int16 __kmp_compare_and_store16(
volatile kmp_int16 *p, kmp_int16 cv,
418 extern kmp_int32 __kmp_compare_and_store32(
volatile kmp_int32 *p, kmp_int32 cv,
420 extern kmp_int32 __kmp_compare_and_store64(
volatile kmp_int64 *p, kmp_int64 cv,
422 extern kmp_int8 __kmp_compare_and_store_ret8(
volatile kmp_int8 *p, kmp_int8 cv,
424 extern kmp_int16 __kmp_compare_and_store_ret16(
volatile kmp_int16 *p,
425 kmp_int16 cv, kmp_int16 sv);
426 extern kmp_int32 __kmp_compare_and_store_ret32(
volatile kmp_int32 *p,
427 kmp_int32 cv, kmp_int32 sv);
428 extern kmp_int64 __kmp_compare_and_store_ret64(
volatile kmp_int64 *p,
429 kmp_int64 cv, kmp_int64 sv);
431 extern kmp_int8 __kmp_xchg_fixed8(
volatile kmp_int8 *p, kmp_int8 v);
432 extern kmp_int16 __kmp_xchg_fixed16(
volatile kmp_int16 *p, kmp_int16 v);
433 extern kmp_int32 __kmp_xchg_fixed32(
volatile kmp_int32 *p, kmp_int32 v);
434 extern kmp_int64 __kmp_xchg_fixed64(
volatile kmp_int64 *p, kmp_int64 v);
435 extern kmp_real32 __kmp_xchg_real32(
volatile kmp_real32 *p, kmp_real32 v);
436 extern kmp_real64 __kmp_xchg_real64(
volatile kmp_real64 *p, kmp_real64 v);
440 #define KMP_TEST_THEN_INC64(p) __kmp_test_then_add64((p), 1LL)
441 #define KMP_TEST_THEN_INC_ACQ64(p) __kmp_test_then_add64((p), 1LL)
444 #define KMP_TEST_THEN_ADD4_64(p) __kmp_test_then_add64((p), 4LL)
445 #define KMP_TEST_THEN_ADD4_ACQ64(p) __kmp_test_then_add64((p), 4LL)
448 #define KMP_TEST_THEN_DEC64(p) __kmp_test_then_add64((p), -1LL)
449 #define KMP_TEST_THEN_DEC_ACQ64(p) __kmp_test_then_add64((p), -1LL)
451 #define KMP_TEST_THEN_ADD8(p, v) __kmp_test_then_add8((p), (v))
452 #define KMP_TEST_THEN_ADD64(p, v) __kmp_test_then_add64((p), (v))
454 #define KMP_TEST_THEN_OR8(p, v) __kmp_test_then_or8((p), (v))
455 #define KMP_TEST_THEN_AND8(p, v) __kmp_test_then_and8((p), (v))
456 #define KMP_TEST_THEN_OR32(p, v) __kmp_test_then_or32((p), (v))
457 #define KMP_TEST_THEN_AND32(p, v) __kmp_test_then_and32((p), (v))
458 #define KMP_TEST_THEN_OR64(p, v) __kmp_test_then_or64((p), (v))
459 #define KMP_TEST_THEN_AND64(p, v) __kmp_test_then_and64((p), (v))
461 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
462 __kmp_compare_and_store8((p), (cv), (sv))
463 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
464 __kmp_compare_and_store8((p), (cv), (sv))
465 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
466 __kmp_compare_and_store16((p), (cv), (sv))
467 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
468 __kmp_compare_and_store16((p), (cv), (sv))
469 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
470 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
472 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
473 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
475 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
476 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
478 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
479 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
483 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
484 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
487 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
488 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
492 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
493 __kmp_compare_and_store_ret8((p), (cv), (sv))
494 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
495 __kmp_compare_and_store_ret16((p), (cv), (sv))
496 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
497 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
500 #define KMP_XCHG_FIXED8(p, v) \
501 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
502 #define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
506 #define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
508 #elif (KMP_ASM_INTRINS && KMP_OS_UNIX) || !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
511 #define KMP_TEST_THEN_INC32(p) \
512 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
513 #define KMP_TEST_THEN_INC_ACQ32(p) \
514 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
515 #define KMP_TEST_THEN_INC64(p) \
516 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
517 #define KMP_TEST_THEN_INC_ACQ64(p) \
518 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
519 #define KMP_TEST_THEN_ADD4_32(p) \
520 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
521 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
522 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
523 #define KMP_TEST_THEN_ADD4_64(p) \
524 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
525 #define KMP_TEST_THEN_ADD4_ACQ64(p) \
526 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
527 #define KMP_TEST_THEN_DEC32(p) \
528 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
529 #define KMP_TEST_THEN_DEC_ACQ32(p) \
530 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
531 #define KMP_TEST_THEN_DEC64(p) \
532 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
533 #define KMP_TEST_THEN_DEC_ACQ64(p) \
534 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
535 #define KMP_TEST_THEN_ADD8(p, v) \
536 __sync_fetch_and_add((volatile kmp_int8 *)(p), (kmp_int8)(v))
537 #define KMP_TEST_THEN_ADD32(p, v) \
538 __sync_fetch_and_add((volatile kmp_int32 *)(p), (kmp_int32)(v))
539 #define KMP_TEST_THEN_ADD64(p, v) \
540 __sync_fetch_and_add((volatile kmp_int64 *)(p), (kmp_int64)(v))
542 #define KMP_TEST_THEN_OR8(p, v) \
543 __sync_fetch_and_or((volatile kmp_int8 *)(p), (kmp_int8)(v))
544 #define KMP_TEST_THEN_AND8(p, v) \
545 __sync_fetch_and_and((volatile kmp_int8 *)(p), (kmp_int8)(v))
546 #define KMP_TEST_THEN_OR32(p, v) \
547 __sync_fetch_and_or((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
548 #define KMP_TEST_THEN_AND32(p, v) \
549 __sync_fetch_and_and((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
550 #define KMP_TEST_THEN_OR64(p, v) \
551 __sync_fetch_and_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
552 #define KMP_TEST_THEN_AND64(p, v) \
553 __sync_fetch_and_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
555 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
556 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
558 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
559 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
561 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
562 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
564 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
565 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
567 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
568 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
570 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
571 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
573 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
574 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
576 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
577 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
579 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
580 __sync_bool_compare_and_swap((void *volatile *)(p), (void *)(cv), \
583 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
584 __sync_val_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
586 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
587 __sync_val_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
589 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
590 __sync_val_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
592 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
593 __sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
596 #define KMP_XCHG_FIXED8(p, v) \
597 __sync_lock_test_and_set((volatile kmp_uint8 *)(p), (kmp_uint8)(v))
598 #define KMP_XCHG_FIXED16(p, v) \
599 __sync_lock_test_and_set((volatile kmp_uint16 *)(p), (kmp_uint16)(v))
600 #define KMP_XCHG_FIXED32(p, v) \
601 __sync_lock_test_and_set((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
602 #define KMP_XCHG_FIXED64(p, v) \
603 __sync_lock_test_and_set((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
605 inline kmp_real32 KMP_XCHG_REAL32(
volatile kmp_real32 *p, kmp_real32 v) {
607 __sync_lock_test_and_set((
volatile kmp_uint32 *)(p), *(kmp_uint32 *)&v);
608 return *(kmp_real32 *)&tmp;
611 inline kmp_real64 KMP_XCHG_REAL64(
volatile kmp_real64 *p, kmp_real64 v) {
613 __sync_lock_test_and_set((
volatile kmp_uint64 *)(p), *(kmp_uint64 *)&v);
614 return *(kmp_real64 *)&tmp;
619 extern kmp_int8 __kmp_test_then_add8(
volatile kmp_int8 *p, kmp_int8 v);
620 extern kmp_int8 __kmp_test_then_or8(
volatile kmp_int8 *p, kmp_int8 v);
621 extern kmp_int8 __kmp_test_then_and8(
volatile kmp_int8 *p, kmp_int8 v);
622 extern kmp_int32 __kmp_test_then_add32(
volatile kmp_int32 *p, kmp_int32 v);
623 extern kmp_uint32 __kmp_test_then_or32(
volatile kmp_uint32 *p, kmp_uint32 v);
624 extern kmp_uint32 __kmp_test_then_and32(
volatile kmp_uint32 *p, kmp_uint32 v);
625 extern kmp_int64 __kmp_test_then_add64(
volatile kmp_int64 *p, kmp_int64 v);
626 extern kmp_uint64 __kmp_test_then_or64(
volatile kmp_uint64 *p, kmp_uint64 v);
627 extern kmp_uint64 __kmp_test_then_and64(
volatile kmp_uint64 *p, kmp_uint64 v);
629 extern kmp_int8 __kmp_compare_and_store8(
volatile kmp_int8 *p, kmp_int8 cv,
631 extern kmp_int16 __kmp_compare_and_store16(
volatile kmp_int16 *p, kmp_int16 cv,
633 extern kmp_int32 __kmp_compare_and_store32(
volatile kmp_int32 *p, kmp_int32 cv,
635 extern kmp_int32 __kmp_compare_and_store64(
volatile kmp_int64 *p, kmp_int64 cv,
637 extern kmp_int8 __kmp_compare_and_store_ret8(
volatile kmp_int8 *p, kmp_int8 cv,
639 extern kmp_int16 __kmp_compare_and_store_ret16(
volatile kmp_int16 *p,
640 kmp_int16 cv, kmp_int16 sv);
641 extern kmp_int32 __kmp_compare_and_store_ret32(
volatile kmp_int32 *p,
642 kmp_int32 cv, kmp_int32 sv);
643 extern kmp_int64 __kmp_compare_and_store_ret64(
volatile kmp_int64 *p,
644 kmp_int64 cv, kmp_int64 sv);
646 extern kmp_int8 __kmp_xchg_fixed8(
volatile kmp_int8 *p, kmp_int8 v);
647 extern kmp_int16 __kmp_xchg_fixed16(
volatile kmp_int16 *p, kmp_int16 v);
648 extern kmp_int32 __kmp_xchg_fixed32(
volatile kmp_int32 *p, kmp_int32 v);
649 extern kmp_int64 __kmp_xchg_fixed64(
volatile kmp_int64 *p, kmp_int64 v);
650 extern kmp_real32 __kmp_xchg_real32(
volatile kmp_real32 *p, kmp_real32 v);
651 extern kmp_real64 __kmp_xchg_real64(
volatile kmp_real64 *p, kmp_real64 v);
653 #define KMP_TEST_THEN_INC32(p) \
654 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
655 #define KMP_TEST_THEN_INC_ACQ32(p) \
656 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
657 #define KMP_TEST_THEN_INC64(p) \
658 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
659 #define KMP_TEST_THEN_INC_ACQ64(p) \
660 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
661 #define KMP_TEST_THEN_ADD4_32(p) \
662 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
663 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
664 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
665 #define KMP_TEST_THEN_ADD4_64(p) \
666 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
667 #define KMP_TEST_THEN_ADD4_ACQ64(p) \
668 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
669 #define KMP_TEST_THEN_DEC32(p) \
670 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
671 #define KMP_TEST_THEN_DEC_ACQ32(p) \
672 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
673 #define KMP_TEST_THEN_DEC64(p) \
674 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
675 #define KMP_TEST_THEN_DEC_ACQ64(p) \
676 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
677 #define KMP_TEST_THEN_ADD8(p, v) \
678 __kmp_test_then_add8((volatile kmp_int8 *)(p), (kmp_int8)(v))
679 #define KMP_TEST_THEN_ADD32(p, v) \
680 __kmp_test_then_add32((volatile kmp_int32 *)(p), (kmp_int32)(v))
681 #define KMP_TEST_THEN_ADD64(p, v) \
682 __kmp_test_then_add64((volatile kmp_int64 *)(p), (kmp_int64)(v))
684 #define KMP_TEST_THEN_OR8(p, v) \
685 __kmp_test_then_or8((volatile kmp_int8 *)(p), (kmp_int8)(v))
686 #define KMP_TEST_THEN_AND8(p, v) \
687 __kmp_test_then_and8((volatile kmp_int8 *)(p), (kmp_int8)(v))
688 #define KMP_TEST_THEN_OR32(p, v) \
689 __kmp_test_then_or32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
690 #define KMP_TEST_THEN_AND32(p, v) \
691 __kmp_test_then_and32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
692 #define KMP_TEST_THEN_OR64(p, v) \
693 __kmp_test_then_or64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
694 #define KMP_TEST_THEN_AND64(p, v) \
695 __kmp_test_then_and64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
697 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
698 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
700 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
701 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
703 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
704 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
706 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
707 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
709 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
710 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
712 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
713 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
715 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
716 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
718 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
719 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
723 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
724 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
727 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
728 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
732 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
733 __kmp_compare_and_store_ret8((p), (cv), (sv))
734 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
735 __kmp_compare_and_store_ret16((p), (cv), (sv))
736 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
737 __kmp_compare_and_store_ret32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
739 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
740 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
743 #define KMP_XCHG_FIXED8(p, v) \
744 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
745 #define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
746 #define KMP_XCHG_FIXED32(p, v) __kmp_xchg_fixed32((p), (v));
747 #define KMP_XCHG_FIXED64(p, v) __kmp_xchg_fixed64((p), (v));
748 #define KMP_XCHG_REAL32(p, v) __kmp_xchg_real32((p), (v));
749 #define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
757 #define KMP_MB() asm("nop")
758 #define KMP_IMB() asm("nop")
765 #if KMP_ARCH_PPC64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS || \
767 #define KMP_MB() __sync_synchronize()
779 #define KMP_ST_REL32(A, D) (*(A) = (D))
783 #define KMP_ST_REL64(A, D) (*(A) = (D))
787 #define KMP_LD_ACQ32(A) (*(A))
791 #define KMP_LD_ACQ64(A) (*(A))
807 #define TCW_1(a, b) (a) = (b)
809 #define TCW_4(a, b) (a) = (b)
810 #define TCI_4(a) (++(a))
811 #define TCD_4(a) (--(a))
813 #define TCW_8(a, b) (a) = (b)
814 #define TCI_8(a) (++(a))
815 #define TCD_8(a) (--(a))
816 #define TCR_SYNC_4(a) (a)
817 #define TCW_SYNC_4(a, b) (a) = (b)
818 #define TCX_SYNC_4(a, b, c) \
819 KMP_COMPARE_AND_STORE_REL32((volatile kmp_int32 *)(volatile void *)&(a), \
820 (kmp_int32)(b), (kmp_int32)(c))
821 #define TCR_SYNC_8(a) (a)
822 #define TCW_SYNC_8(a, b) (a) = (b)
823 #define TCX_SYNC_8(a, b, c) \
824 KMP_COMPARE_AND_STORE_REL64((volatile kmp_int64 *)(volatile void *)&(a), \
825 (kmp_int64)(b), (kmp_int64)(c))
827 #if KMP_ARCH_X86 || KMP_ARCH_MIPS
829 #define TCR_PTR(a) ((void *)TCR_4(a))
830 #define TCW_PTR(a, b) TCW_4((a), (b))
831 #define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_4(a))
832 #define TCW_SYNC_PTR(a, b) TCW_SYNC_4((a), (b))
833 #define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_4((a), (b), (c)))
837 #define TCR_PTR(a) ((void *)TCR_8(a))
838 #define TCW_PTR(a, b) TCW_8((a), (b))
839 #define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_8(a))
840 #define TCW_SYNC_PTR(a, b) TCW_SYNC_8((a), (b))
841 #define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_8((a), (b), (c)))
849 #define FTN_TRUE TRUE
853 #define FTN_FALSE FALSE
856 typedef void (*microtask_t)(
int *gtid,
int *npr, ...);
858 #ifdef USE_VOLATILE_CAST
859 #define VOLATILE_CAST(x) (volatile x)
861 #define VOLATILE_CAST(x) (x)
864 #define KMP_WAIT_YIELD __kmp_wait_yield_4
865 #define KMP_WAIT_YIELD_PTR __kmp_wait_yield_4_ptr
866 #define KMP_EQ __kmp_eq_4
867 #define KMP_NEQ __kmp_neq_4
868 #define KMP_LT __kmp_lt_4
869 #define KMP_GE __kmp_ge_4
870 #define KMP_LE __kmp_le_4
874 #if (KMP_ARCH_X86_64 || KMP_ARCH_PPC64) && KMP_OS_LINUX
875 #define STATIC_EFI2_WORKAROUND
877 #define STATIC_EFI2_WORKAROUND static
882 #define KMP_USE_BGET 1
886 #ifndef USE_CMPXCHG_FIX
887 #define USE_CMPXCHG_FIX 1
892 #define KMP_USE_DYNAMIC_LOCK 1
897 #if KMP_USE_DYNAMIC_LOCK
899 #define KMP_USE_TSX (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !KMP_COMPILER_MSVC
900 #ifdef KMP_USE_ADAPTIVE_LOCKS
901 #undef KMP_USE_ADAPTIVE_LOCKS
903 #define KMP_USE_ADAPTIVE_LOCKS KMP_USE_TSX
907 #if KMP_STATS_ENABLED
908 #define KMP_HAVE_TICK_TIME \
909 (KMP_OS_LINUX && (KMP_MIC || KMP_ARCH_X86 || KMP_ARCH_X86_64))
913 enum kmp_warnings_level {
914 kmp_warnings_off = 0,
916 kmp_warnings_explicit = 6,
922 #endif // __cplusplus
925 #define KMP_ATOMIC_LD(p, order) (p)->load(std::memory_order_##order)
926 #define KMP_ATOMIC_OP(op, p, v, order) (p)->op(v, std::memory_order_##order)
929 #define KMP_ATOMIC_LD_ACQ(p) KMP_ATOMIC_LD(p, acquire)
930 #define KMP_ATOMIC_LD_RLX(p) KMP_ATOMIC_LD(p, relaxed)
931 #define KMP_ATOMIC_ST_REL(p, v) KMP_ATOMIC_OP(store, p, v, release)
932 #define KMP_ATOMIC_ST_RLX(p, v) KMP_ATOMIC_OP(store, p, v, relaxed)
935 #define KMP_ATOMIC_ADD(p, v) KMP_ATOMIC_OP(fetch_add, p, v, acq_rel)
936 #define KMP_ATOMIC_SUB(p, v) KMP_ATOMIC_OP(fetch_sub, p, v, acq_rel)
937 #define KMP_ATOMIC_AND(p, v) KMP_ATOMIC_OP(fetch_and, p, v, acq_rel)
938 #define KMP_ATOMIC_OR(p, v) KMP_ATOMIC_OP(fetch_or, p, v, acq_rel)
939 #define KMP_ATOMIC_INC(p) KMP_ATOMIC_OP(fetch_add, p, 1, acq_rel)
940 #define KMP_ATOMIC_DEC(p) KMP_ATOMIC_OP(fetch_sub, p, 1, acq_rel)
941 #define KMP_ATOMIC_ADD_RLX(p, v) KMP_ATOMIC_OP(fetch_add, p, v, relaxed)
942 #define KMP_ATOMIC_INC_RLX(p) KMP_ATOMIC_OP(fetch_add, p, 1, relaxed)
945 template <
typename T>
946 bool __kmp_atomic_compare_store(std::atomic<T> *p, T expected, T desired) {
947 return p->compare_exchange_strong(
948 expected, desired, std::memory_order_acq_rel, std::memory_order_relaxed);
951 template <
typename T>
952 bool __kmp_atomic_compare_store_acq(std::atomic<T> *p, T expected, T desired) {
953 return p->compare_exchange_strong(
954 expected, desired, std::memory_order_acquire, std::memory_order_relaxed);
957 template <
typename T>
958 bool __kmp_atomic_compare_store_rel(std::atomic<T> *p, T expected, T desired) {
959 return p->compare_exchange_strong(
960 expected, desired, std::memory_order_release, std::memory_order_relaxed);
965 #include "kmp_safe_c_api.h"