16 #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
43 #ifndef __has_cpp_attribute
44 #define __has_cpp_attribute(x) 0
47 #ifndef __has_attribute
48 #define __has_attribute(x) 0
52 #define KMP_COMPILER_ICC 0
53 #define KMP_COMPILER_GCC 0
54 #define KMP_COMPILER_CLANG 0
55 #define KMP_COMPILER_MSVC 0
57 #if defined(__INTEL_COMPILER)
58 #undef KMP_COMPILER_ICC
59 #define KMP_COMPILER_ICC 1
60 #elif defined(__clang__)
61 #undef KMP_COMPILER_CLANG
62 #define KMP_COMPILER_CLANG 1
63 #elif defined(__GNUC__)
64 #undef KMP_COMPILER_GCC
65 #define KMP_COMPILER_GCC 1
66 #elif defined(_MSC_VER)
67 #undef KMP_COMPILER_MSVC
68 #define KMP_COMPILER_MSVC 1
70 #error Unknown compiler
73 #if (KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_FREEBSD)
74 #define KMP_AFFINITY_SUPPORTED 1
75 #if KMP_OS_WINDOWS && KMP_ARCH_X86_64
76 #define KMP_GROUP_AFFINITY 1
78 #define KMP_GROUP_AFFINITY 0
81 #define KMP_AFFINITY_SUPPORTED 0
82 #define KMP_GROUP_AFFINITY 0
86 #define KMP_HAVE_QUAD 0
87 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
91 #define KMP_HAVE_QUAD 1
92 #elif KMP_COMPILER_CLANG
95 typedef long double _Quad;
96 #elif KMP_COMPILER_GCC
99 typedef __float128 _Quad;
101 #define KMP_HAVE_QUAD 1
103 #elif KMP_COMPILER_MSVC
104 typedef long double _Quad;
107 #if __LDBL_MAX_EXP__ >= 16384 && KMP_COMPILER_GCC
108 typedef long double _Quad;
110 #define KMP_HAVE_QUAD 1
114 #define KMP_USE_X87CONTROL 0
116 #define KMP_END_OF_LINE "\r\n"
117 typedef char kmp_int8;
118 typedef unsigned char kmp_uint8;
119 typedef short kmp_int16;
120 typedef unsigned short kmp_uint16;
121 typedef int kmp_int32;
122 typedef unsigned int kmp_uint32;
123 #define KMP_INT32_SPEC "d"
124 #define KMP_UINT32_SPEC "u"
126 typedef __int64 kmp_int64;
127 typedef unsigned __int64 kmp_uint64;
128 #define KMP_INT64_SPEC "I64d"
129 #define KMP_UINT64_SPEC "I64u"
131 struct kmp_struct64 {
134 typedef struct kmp_struct64 kmp_int64;
135 typedef struct kmp_struct64 kmp_uint64;
138 #if KMP_ARCH_X86 && KMP_MSVC_COMPAT
139 #undef KMP_USE_X87CONTROL
140 #define KMP_USE_X87CONTROL 1
144 typedef __int64 kmp_intptr_t;
145 typedef unsigned __int64 kmp_uintptr_t;
146 #define KMP_INTPTR_SPEC "I64d"
147 #define KMP_UINTPTR_SPEC "I64u"
152 #define KMP_END_OF_LINE "\n"
153 typedef char kmp_int8;
154 typedef unsigned char kmp_uint8;
155 typedef short kmp_int16;
156 typedef unsigned short kmp_uint16;
157 typedef int kmp_int32;
158 typedef unsigned int kmp_uint32;
159 typedef long long kmp_int64;
160 typedef unsigned long long kmp_uint64;
161 #define KMP_INT32_SPEC "d"
162 #define KMP_UINT32_SPEC "u"
163 #define KMP_INT64_SPEC "lld"
164 #define KMP_UINT64_SPEC "llu"
167 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS
168 #define KMP_SIZE_T_SPEC KMP_UINT32_SPEC
169 #elif KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || \
170 KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64
171 #define KMP_SIZE_T_SPEC KMP_UINT64_SPEC
173 #error "Can't determine size_t printf format specifier."
177 #define KMP_SIZE_T_MAX (0xFFFFFFFF)
179 #define KMP_SIZE_T_MAX (0xFFFFFFFFFFFFFFFF)
182 typedef size_t kmp_size_t;
183 typedef float kmp_real32;
184 typedef double kmp_real64;
188 typedef long kmp_intptr_t;
189 typedef unsigned long kmp_uintptr_t;
190 #define KMP_INTPTR_SPEC "ld"
191 #define KMP_UINTPTR_SPEC "lu"
195 typedef kmp_int64 kmp_int;
196 typedef kmp_uint64 kmp_uint;
198 typedef kmp_int32 kmp_int;
199 typedef kmp_uint32 kmp_uint;
201 #define KMP_INT_MAX ((kmp_int32)0x7FFFFFFF)
202 #define KMP_INT_MIN ((kmp_int32)0x80000000)
205 #if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && \
206 (KMP_OS_FREEBSD || KMP_OS_LINUX)
207 typedef va_list *kmp_va_list;
208 #define kmp_va_deref(ap) (*(ap))
209 #define kmp_va_addr_of(ap) (&(ap))
211 typedef va_list kmp_va_list;
212 #define kmp_va_deref(ap) (ap)
213 #define kmp_va_addr_of(ap) (ap)
218 #define CCAST(type, var) const_cast<type>(var)
219 #define RCAST(type, var) reinterpret_cast<type>(var)
223 template <
typename T>
struct traits_t {};
225 template <>
struct traits_t<signed int> {
226 typedef signed int signed_t;
227 typedef unsigned int unsigned_t;
228 typedef double floating_t;
229 static char const *spec;
230 static const signed_t max_value = 0x7fffffff;
231 static const signed_t min_value = 0x80000000;
232 static const int type_size =
sizeof(signed_t);
235 template <>
struct traits_t<unsigned int> {
236 typedef signed int signed_t;
237 typedef unsigned int unsigned_t;
238 typedef double floating_t;
239 static char const *spec;
240 static const unsigned_t max_value = 0xffffffff;
241 static const unsigned_t min_value = 0x00000000;
242 static const int type_size =
sizeof(unsigned_t);
245 template <>
struct traits_t<signed long> {
246 typedef signed long signed_t;
247 typedef unsigned long unsigned_t;
248 typedef long double floating_t;
249 static char const *spec;
250 static const int type_size =
sizeof(signed_t);
253 template <>
struct traits_t<signed long long> {
254 typedef signed long long signed_t;
255 typedef unsigned long long unsigned_t;
256 typedef long double floating_t;
257 static char const *spec;
258 static const signed_t max_value = 0x7fffffffffffffffLL;
259 static const signed_t min_value = 0x8000000000000000LL;
260 static const int type_size =
sizeof(signed_t);
263 template <>
struct traits_t<unsigned long long> {
264 typedef signed long long signed_t;
265 typedef unsigned long long unsigned_t;
266 typedef long double floating_t;
267 static char const *spec;
268 static const unsigned_t max_value = 0xffffffffffffffffLL;
269 static const unsigned_t min_value = 0x0000000000000000LL;
270 static const int type_size =
sizeof(unsigned_t);
274 #define CCAST(type, var) (type)(var)
275 #define RCAST(type, var) (type)(var)
278 #define KMP_EXPORT extern
280 #if __GNUC__ >= 4 && !defined(__MINGW32__)
281 #define __forceinline __inline
287 #define KMP_HAVE_MWAIT \
288 ((KMP_ARCH_X86 || KMP_ARCH_X86_64) && (KMP_OS_LINUX || KMP_OS_WINDOWS) && \
290 #define KMP_HAVE_UMWAIT \
291 ((KMP_ARCH_X86 || KMP_ARCH_X86_64) && (KMP_OS_LINUX || KMP_OS_WINDOWS) && \
297 static inline int KMP_GET_PAGE_SIZE(
void) {
300 return si.dwPageSize;
303 #define KMP_GET_PAGE_SIZE() getpagesize()
306 #define PAGE_ALIGNED(_addr) \
307 (!((size_t)_addr & (size_t)(KMP_GET_PAGE_SIZE() - 1)))
308 #define ALIGN_TO_PAGE(x) \
309 (void *)(((size_t)(x)) & ~((size_t)(KMP_GET_PAGE_SIZE() - 1)))
317 #define INTERNODE_CACHE_LINE 4096
321 #define CACHE_LINE 128
323 #if (CACHE_LINE < 64) && !defined(KMP_OS_DARWIN)
325 #warning CACHE_LINE is too small.
329 #define KMP_CACHE_PREFETCH(ADDR)
335 #if __cplusplus > 201402L && __has_cpp_attribute(fallthrough)
336 #define KMP_FALLTHROUGH() [[fallthrough]]
337 #elif __has_cpp_attribute(clang::fallthrough)
338 #define KMP_FALLTHROUGH() [[clang::fallthrough]]
339 #elif __has_attribute(fallthrough) || __GNUC__ >= 7
340 #define KMP_FALLTHROUGH() __attribute__((__fallthrough__))
342 #define KMP_FALLTHROUGH() ((void)0)
345 #if KMP_HAVE_ATTRIBUTE_WAITPKG
346 #define KMP_ATTRIBUTE_TARGET_WAITPKG __attribute__((target("waitpkg")))
348 #define KMP_ATTRIBUTE_TARGET_WAITPKG
351 #if KMP_HAVE_ATTRIBUTE_RTM
352 #define KMP_ATTRIBUTE_TARGET_RTM __attribute__((target("rtm")))
354 #define KMP_ATTRIBUTE_TARGET_RTM
358 #if __cplusplus >= 201103L
359 #define KMP_NORETURN [[noreturn]]
361 #define KMP_NORETURN __declspec(noreturn)
363 #define KMP_NORETURN __attribute__((noreturn))
366 #if KMP_OS_WINDOWS && KMP_MSVC_COMPAT
367 #define KMP_ALIGN(bytes) __declspec(align(bytes))
368 #define KMP_THREAD_LOCAL __declspec(thread)
371 #define KMP_ALIGN(bytes) __attribute__((aligned(bytes)))
372 #define KMP_THREAD_LOCAL __thread
373 #define KMP_ALIAS(alias_of) __attribute__((alias(alias_of)))
376 #if KMP_HAVE_WEAK_ATTRIBUTE && !KMP_DYNAMIC_LIB
377 #define KMP_WEAK_ATTRIBUTE_EXTERNAL __attribute__((weak))
379 #define KMP_WEAK_ATTRIBUTE_EXTERNAL
382 #if KMP_HAVE_WEAK_ATTRIBUTE
383 #define KMP_WEAK_ATTRIBUTE_INTERNAL __attribute__((weak))
385 #define KMP_WEAK_ATTRIBUTE_INTERNAL
390 #define KMP_STR(x) _KMP_STR(x)
391 #define _KMP_STR(x) #x
394 #ifdef KMP_USE_VERSION_SYMBOLS
397 #define KMP_EXPAND_NAME(api_name) _KMP_EXPAND_NAME(api_name)
398 #define _KMP_EXPAND_NAME(api_name) __kmp_api_##api_name
399 #define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str) \
400 _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, "VERSION")
401 #define _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, default_ver) \
402 __typeof__(__kmp_api_##api_name) __kmp_api_##api_name##_##ver_num##_alias \
403 __attribute__((alias(KMP_STR(__kmp_api_##api_name)))); \
405 ".symver " KMP_STR(__kmp_api_##api_name##_##ver_num##_alias) "," KMP_STR( \
406 api_name) "@" ver_str "\n\t"); \
407 __asm__(".symver " KMP_STR(__kmp_api_##api_name) "," KMP_STR( \
408 api_name) "@@" default_ver "\n\t")
410 #define KMP_EXPAND_NAME(api_name) api_name
411 #define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str)
416 #define KMP_DO_ALIGN(bytes) KMP_ALIGN(bytes)
417 #define KMP_ALIGN_CACHE KMP_ALIGN(CACHE_LINE)
418 #define KMP_ALIGN_CACHE_INTERNODE KMP_ALIGN(INTERNODE_CACHE_LINE)
421 enum kmp_mem_fence_type {
430 #if KMP_ASM_INTRINS && KMP_OS_WINDOWS
432 #if KMP_MSVC_COMPAT && !KMP_COMPILER_CLANG
433 #pragma intrinsic(InterlockedExchangeAdd)
434 #pragma intrinsic(InterlockedCompareExchange)
435 #pragma intrinsic(InterlockedExchange)
436 #pragma intrinsic(InterlockedExchange64)
441 #define KMP_TEST_THEN_INC32(p) InterlockedExchangeAdd((volatile long *)(p), 1)
442 #define KMP_TEST_THEN_INC_ACQ32(p) \
443 InterlockedExchangeAdd((volatile long *)(p), 1)
444 #define KMP_TEST_THEN_ADD4_32(p) InterlockedExchangeAdd((volatile long *)(p), 4)
445 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
446 InterlockedExchangeAdd((volatile long *)(p), 4)
447 #define KMP_TEST_THEN_DEC32(p) InterlockedExchangeAdd((volatile long *)(p), -1)
448 #define KMP_TEST_THEN_DEC_ACQ32(p) \
449 InterlockedExchangeAdd((volatile long *)(p), -1)
450 #define KMP_TEST_THEN_ADD32(p, v) \
451 InterlockedExchangeAdd((volatile long *)(p), (v))
453 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
454 InterlockedCompareExchange((volatile long *)(p), (long)(sv), (long)(cv))
456 #define KMP_XCHG_FIXED32(p, v) \
457 InterlockedExchange((volatile long *)(p), (long)(v))
458 #define KMP_XCHG_FIXED64(p, v) \
459 InterlockedExchange64((volatile kmp_int64 *)(p), (kmp_int64)(v))
461 inline kmp_real32 KMP_XCHG_REAL32(
volatile kmp_real32 *p, kmp_real32 v) {
462 kmp_int32 tmp = InterlockedExchange((
volatile long *)p, *(
long *)&v);
463 return *(kmp_real32 *)&tmp;
467 extern kmp_int8 __kmp_test_then_add8(
volatile kmp_int8 *p, kmp_int8 v);
468 extern kmp_int8 __kmp_test_then_or8(
volatile kmp_int8 *p, kmp_int8 v);
469 extern kmp_int8 __kmp_test_then_and8(
volatile kmp_int8 *p, kmp_int8 v);
470 extern kmp_int32 __kmp_test_then_add32(
volatile kmp_int32 *p, kmp_int32 v);
471 extern kmp_uint32 __kmp_test_then_or32(
volatile kmp_uint32 *p, kmp_uint32 v);
472 extern kmp_uint32 __kmp_test_then_and32(
volatile kmp_uint32 *p, kmp_uint32 v);
473 extern kmp_int64 __kmp_test_then_add64(
volatile kmp_int64 *p, kmp_int64 v);
474 extern kmp_uint64 __kmp_test_then_or64(
volatile kmp_uint64 *p, kmp_uint64 v);
475 extern kmp_uint64 __kmp_test_then_and64(
volatile kmp_uint64 *p, kmp_uint64 v);
477 extern kmp_int8 __kmp_compare_and_store8(
volatile kmp_int8 *p, kmp_int8 cv,
479 extern kmp_int16 __kmp_compare_and_store16(
volatile kmp_int16 *p, kmp_int16 cv,
481 extern kmp_int32 __kmp_compare_and_store32(
volatile kmp_int32 *p, kmp_int32 cv,
483 extern kmp_int32 __kmp_compare_and_store64(
volatile kmp_int64 *p, kmp_int64 cv,
485 extern kmp_int8 __kmp_compare_and_store_ret8(
volatile kmp_int8 *p, kmp_int8 cv,
487 extern kmp_int16 __kmp_compare_and_store_ret16(
volatile kmp_int16 *p,
488 kmp_int16 cv, kmp_int16 sv);
489 extern kmp_int32 __kmp_compare_and_store_ret32(
volatile kmp_int32 *p,
490 kmp_int32 cv, kmp_int32 sv);
491 extern kmp_int64 __kmp_compare_and_store_ret64(
volatile kmp_int64 *p,
492 kmp_int64 cv, kmp_int64 sv);
494 extern kmp_int8 __kmp_xchg_fixed8(
volatile kmp_int8 *p, kmp_int8 v);
495 extern kmp_int16 __kmp_xchg_fixed16(
volatile kmp_int16 *p, kmp_int16 v);
496 extern kmp_int32 __kmp_xchg_fixed32(
volatile kmp_int32 *p, kmp_int32 v);
497 extern kmp_int64 __kmp_xchg_fixed64(
volatile kmp_int64 *p, kmp_int64 v);
498 extern kmp_real32 __kmp_xchg_real32(
volatile kmp_real32 *p, kmp_real32 v);
499 extern kmp_real64 __kmp_xchg_real64(
volatile kmp_real64 *p, kmp_real64 v);
503 #define KMP_TEST_THEN_INC64(p) __kmp_test_then_add64((p), 1LL)
504 #define KMP_TEST_THEN_INC_ACQ64(p) __kmp_test_then_add64((p), 1LL)
507 #define KMP_TEST_THEN_ADD4_64(p) __kmp_test_then_add64((p), 4LL)
508 #define KMP_TEST_THEN_ADD4_ACQ64(p) __kmp_test_then_add64((p), 4LL)
511 #define KMP_TEST_THEN_DEC64(p) __kmp_test_then_add64((p), -1LL)
512 #define KMP_TEST_THEN_DEC_ACQ64(p) __kmp_test_then_add64((p), -1LL)
514 #define KMP_TEST_THEN_ADD8(p, v) __kmp_test_then_add8((p), (v))
515 #define KMP_TEST_THEN_ADD64(p, v) __kmp_test_then_add64((p), (v))
517 #define KMP_TEST_THEN_OR8(p, v) __kmp_test_then_or8((p), (v))
518 #define KMP_TEST_THEN_AND8(p, v) __kmp_test_then_and8((p), (v))
519 #define KMP_TEST_THEN_OR32(p, v) __kmp_test_then_or32((p), (v))
520 #define KMP_TEST_THEN_AND32(p, v) __kmp_test_then_and32((p), (v))
521 #define KMP_TEST_THEN_OR64(p, v) __kmp_test_then_or64((p), (v))
522 #define KMP_TEST_THEN_AND64(p, v) __kmp_test_then_and64((p), (v))
524 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
525 __kmp_compare_and_store8((p), (cv), (sv))
526 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
527 __kmp_compare_and_store8((p), (cv), (sv))
528 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
529 __kmp_compare_and_store16((p), (cv), (sv))
530 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
531 __kmp_compare_and_store16((p), (cv), (sv))
532 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
533 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
535 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
536 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
538 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
539 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
541 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
542 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
546 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
547 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
550 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
551 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
555 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
556 __kmp_compare_and_store_ret8((p), (cv), (sv))
557 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
558 __kmp_compare_and_store_ret16((p), (cv), (sv))
559 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
560 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
563 #define KMP_XCHG_FIXED8(p, v) \
564 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
565 #define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
569 #define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
571 #elif (KMP_ASM_INTRINS && KMP_OS_UNIX) || !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
574 #define KMP_TEST_THEN_INC32(p) \
575 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
576 #define KMP_TEST_THEN_INC_ACQ32(p) \
577 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
579 #define KMP_TEST_THEN_INC64(p) \
580 __atomic_fetch_add((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
581 #define KMP_TEST_THEN_INC_ACQ64(p) \
582 __atomic_fetch_add((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
584 #define KMP_TEST_THEN_INC64(p) \
585 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
586 #define KMP_TEST_THEN_INC_ACQ64(p) \
587 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
589 #define KMP_TEST_THEN_ADD4_32(p) \
590 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
591 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
592 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
594 #define KMP_TEST_THEN_ADD4_64(p) \
595 __atomic_fetch_add((volatile kmp_int64 *)(p), 4LL, __ATOMIC_SEQ_CST)
596 #define KMP_TEST_THEN_ADD4_ACQ64(p) \
597 __atomic_fetch_add((volatile kmp_int64 *)(p), 4LL, __ATOMIC_SEQ_CST)
598 #define KMP_TEST_THEN_DEC64(p) \
599 __atomic_fetch_sub((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
600 #define KMP_TEST_THEN_DEC_ACQ64(p) \
601 __atomic_fetch_sub((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
603 #define KMP_TEST_THEN_ADD4_64(p) \
604 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
605 #define KMP_TEST_THEN_ADD4_ACQ64(p) \
606 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
607 #define KMP_TEST_THEN_DEC64(p) \
608 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
609 #define KMP_TEST_THEN_DEC_ACQ64(p) \
610 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
612 #define KMP_TEST_THEN_DEC32(p) \
613 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
614 #define KMP_TEST_THEN_DEC_ACQ32(p) \
615 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
616 #define KMP_TEST_THEN_ADD8(p, v) \
617 __sync_fetch_and_add((volatile kmp_int8 *)(p), (kmp_int8)(v))
618 #define KMP_TEST_THEN_ADD32(p, v) \
619 __sync_fetch_and_add((volatile kmp_int32 *)(p), (kmp_int32)(v))
621 #define KMP_TEST_THEN_ADD64(p, v) \
622 __atomic_fetch_add((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
625 #define KMP_TEST_THEN_ADD64(p, v) \
626 __sync_fetch_and_add((volatile kmp_int64 *)(p), (kmp_int64)(v))
629 #define KMP_TEST_THEN_OR8(p, v) \
630 __sync_fetch_and_or((volatile kmp_int8 *)(p), (kmp_int8)(v))
631 #define KMP_TEST_THEN_AND8(p, v) \
632 __sync_fetch_and_and((volatile kmp_int8 *)(p), (kmp_int8)(v))
633 #define KMP_TEST_THEN_OR32(p, v) \
634 __sync_fetch_and_or((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
635 #define KMP_TEST_THEN_AND32(p, v) \
636 __sync_fetch_and_and((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
638 #define KMP_TEST_THEN_OR64(p, v) \
639 __atomic_fetch_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
641 #define KMP_TEST_THEN_AND64(p, v) \
642 __atomic_fetch_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
645 #define KMP_TEST_THEN_OR64(p, v) \
646 __sync_fetch_and_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
647 #define KMP_TEST_THEN_AND64(p, v) \
648 __sync_fetch_and_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
651 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
652 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
654 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
655 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
657 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
658 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
660 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
661 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
663 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
664 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
666 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
667 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
669 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
670 __sync_bool_compare_and_swap((void *volatile *)(p), (void *)(cv), \
673 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
674 __sync_val_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
676 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
677 __sync_val_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
679 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
680 __sync_val_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
683 static inline bool mips_sync_bool_compare_and_swap(
volatile kmp_uint64 *p,
686 return __atomic_compare_exchange(p, &cv, &sv,
false, __ATOMIC_SEQ_CST,
689 static inline bool mips_sync_val_compare_and_swap(
volatile kmp_uint64 *p,
692 __atomic_compare_exchange(p, &cv, &sv,
false, __ATOMIC_SEQ_CST,
696 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
697 mips_sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), \
698 (kmp_uint64)(cv), (kmp_uint64)(sv))
699 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
700 mips_sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), \
701 (kmp_uint64)(cv), (kmp_uint64)(sv))
702 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
703 mips_sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
706 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
707 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
709 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
710 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
712 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
713 __sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
717 #define KMP_XCHG_FIXED8(p, v) \
718 __sync_lock_test_and_set((volatile kmp_uint8 *)(p), (kmp_uint8)(v))
719 #define KMP_XCHG_FIXED16(p, v) \
720 __sync_lock_test_and_set((volatile kmp_uint16 *)(p), (kmp_uint16)(v))
721 #define KMP_XCHG_FIXED32(p, v) \
722 __sync_lock_test_and_set((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
723 #define KMP_XCHG_FIXED64(p, v) \
724 __sync_lock_test_and_set((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
726 inline kmp_real32 KMP_XCHG_REAL32(
volatile kmp_real32 *p, kmp_real32 v) {
728 __sync_lock_test_and_set((
volatile kmp_uint32 *)(p), *(kmp_uint32 *)&v);
729 return *(kmp_real32 *)&tmp;
732 inline kmp_real64 KMP_XCHG_REAL64(
volatile kmp_real64 *p, kmp_real64 v) {
734 __sync_lock_test_and_set((
volatile kmp_uint64 *)(p), *(kmp_uint64 *)&v);
735 return *(kmp_real64 *)&tmp;
740 extern kmp_int8 __kmp_test_then_add8(
volatile kmp_int8 *p, kmp_int8 v);
741 extern kmp_int8 __kmp_test_then_or8(
volatile kmp_int8 *p, kmp_int8 v);
742 extern kmp_int8 __kmp_test_then_and8(
volatile kmp_int8 *p, kmp_int8 v);
743 extern kmp_int32 __kmp_test_then_add32(
volatile kmp_int32 *p, kmp_int32 v);
744 extern kmp_uint32 __kmp_test_then_or32(
volatile kmp_uint32 *p, kmp_uint32 v);
745 extern kmp_uint32 __kmp_test_then_and32(
volatile kmp_uint32 *p, kmp_uint32 v);
746 extern kmp_int64 __kmp_test_then_add64(
volatile kmp_int64 *p, kmp_int64 v);
747 extern kmp_uint64 __kmp_test_then_or64(
volatile kmp_uint64 *p, kmp_uint64 v);
748 extern kmp_uint64 __kmp_test_then_and64(
volatile kmp_uint64 *p, kmp_uint64 v);
750 extern kmp_int8 __kmp_compare_and_store8(
volatile kmp_int8 *p, kmp_int8 cv,
752 extern kmp_int16 __kmp_compare_and_store16(
volatile kmp_int16 *p, kmp_int16 cv,
754 extern kmp_int32 __kmp_compare_and_store32(
volatile kmp_int32 *p, kmp_int32 cv,
756 extern kmp_int32 __kmp_compare_and_store64(
volatile kmp_int64 *p, kmp_int64 cv,
758 extern kmp_int8 __kmp_compare_and_store_ret8(
volatile kmp_int8 *p, kmp_int8 cv,
760 extern kmp_int16 __kmp_compare_and_store_ret16(
volatile kmp_int16 *p,
761 kmp_int16 cv, kmp_int16 sv);
762 extern kmp_int32 __kmp_compare_and_store_ret32(
volatile kmp_int32 *p,
763 kmp_int32 cv, kmp_int32 sv);
764 extern kmp_int64 __kmp_compare_and_store_ret64(
volatile kmp_int64 *p,
765 kmp_int64 cv, kmp_int64 sv);
767 extern kmp_int8 __kmp_xchg_fixed8(
volatile kmp_int8 *p, kmp_int8 v);
768 extern kmp_int16 __kmp_xchg_fixed16(
volatile kmp_int16 *p, kmp_int16 v);
769 extern kmp_int32 __kmp_xchg_fixed32(
volatile kmp_int32 *p, kmp_int32 v);
770 extern kmp_int64 __kmp_xchg_fixed64(
volatile kmp_int64 *p, kmp_int64 v);
771 extern kmp_real32 __kmp_xchg_real32(
volatile kmp_real32 *p, kmp_real32 v);
772 extern kmp_real64 __kmp_xchg_real64(
volatile kmp_real64 *p, kmp_real64 v);
774 #define KMP_TEST_THEN_INC32(p) \
775 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
776 #define KMP_TEST_THEN_INC_ACQ32(p) \
777 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
778 #define KMP_TEST_THEN_INC64(p) \
779 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
780 #define KMP_TEST_THEN_INC_ACQ64(p) \
781 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
782 #define KMP_TEST_THEN_ADD4_32(p) \
783 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
784 #define KMP_TEST_THEN_ADD4_ACQ32(p) \
785 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
786 #define KMP_TEST_THEN_ADD4_64(p) \
787 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
788 #define KMP_TEST_THEN_ADD4_ACQ64(p) \
789 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
790 #define KMP_TEST_THEN_DEC32(p) \
791 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
792 #define KMP_TEST_THEN_DEC_ACQ32(p) \
793 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
794 #define KMP_TEST_THEN_DEC64(p) \
795 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
796 #define KMP_TEST_THEN_DEC_ACQ64(p) \
797 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
798 #define KMP_TEST_THEN_ADD8(p, v) \
799 __kmp_test_then_add8((volatile kmp_int8 *)(p), (kmp_int8)(v))
800 #define KMP_TEST_THEN_ADD32(p, v) \
801 __kmp_test_then_add32((volatile kmp_int32 *)(p), (kmp_int32)(v))
802 #define KMP_TEST_THEN_ADD64(p, v) \
803 __kmp_test_then_add64((volatile kmp_int64 *)(p), (kmp_int64)(v))
805 #define KMP_TEST_THEN_OR8(p, v) \
806 __kmp_test_then_or8((volatile kmp_int8 *)(p), (kmp_int8)(v))
807 #define KMP_TEST_THEN_AND8(p, v) \
808 __kmp_test_then_and8((volatile kmp_int8 *)(p), (kmp_int8)(v))
809 #define KMP_TEST_THEN_OR32(p, v) \
810 __kmp_test_then_or32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
811 #define KMP_TEST_THEN_AND32(p, v) \
812 __kmp_test_then_and32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
813 #define KMP_TEST_THEN_OR64(p, v) \
814 __kmp_test_then_or64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
815 #define KMP_TEST_THEN_AND64(p, v) \
816 __kmp_test_then_and64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
818 #define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
819 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
821 #define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
822 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
824 #define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
825 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
827 #define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
828 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
830 #define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
831 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
833 #define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
834 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
836 #define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
837 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
839 #define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
840 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
844 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
845 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
848 #define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
849 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
853 #define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
854 __kmp_compare_and_store_ret8((p), (cv), (sv))
855 #define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
856 __kmp_compare_and_store_ret16((p), (cv), (sv))
857 #define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
858 __kmp_compare_and_store_ret32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
860 #define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
861 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
864 #define KMP_XCHG_FIXED8(p, v) \
865 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
866 #define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
867 #define KMP_XCHG_FIXED32(p, v) __kmp_xchg_fixed32((p), (v));
868 #define KMP_XCHG_FIXED64(p, v) __kmp_xchg_fixed64((p), (v));
869 #define KMP_XCHG_REAL32(p, v) __kmp_xchg_real32((p), (v));
870 #define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
878 #define KMP_MB() asm("nop")
879 #define KMP_IMB() asm("nop")
886 #if KMP_ARCH_PPC64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS || \
887 KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64
888 #define KMP_MB() __sync_synchronize()
900 #define KMP_ST_REL32(A, D) (*(A) = (D))
904 #define KMP_ST_REL64(A, D) (*(A) = (D))
908 #define KMP_LD_ACQ32(A) (*(A))
912 #define KMP_LD_ACQ64(A) (*(A))
928 #define TCW_1(a, b) (a) = (b)
930 #define TCW_4(a, b) (a) = (b)
931 #define TCI_4(a) (++(a))
932 #define TCD_4(a) (--(a))
934 #define TCW_8(a, b) (a) = (b)
935 #define TCI_8(a) (++(a))
936 #define TCD_8(a) (--(a))
937 #define TCR_SYNC_4(a) (a)
938 #define TCW_SYNC_4(a, b) (a) = (b)
939 #define TCX_SYNC_4(a, b, c) \
940 KMP_COMPARE_AND_STORE_REL32((volatile kmp_int32 *)(volatile void *)&(a), \
941 (kmp_int32)(b), (kmp_int32)(c))
942 #define TCR_SYNC_8(a) (a)
943 #define TCW_SYNC_8(a, b) (a) = (b)
944 #define TCX_SYNC_8(a, b, c) \
945 KMP_COMPARE_AND_STORE_REL64((volatile kmp_int64 *)(volatile void *)&(a), \
946 (kmp_int64)(b), (kmp_int64)(c))
948 #if KMP_ARCH_X86 || KMP_ARCH_MIPS
950 #define TCR_PTR(a) ((void *)TCR_4(a))
951 #define TCW_PTR(a, b) TCW_4((a), (b))
952 #define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_4(a))
953 #define TCW_SYNC_PTR(a, b) TCW_SYNC_4((a), (b))
954 #define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_4((a), (b), (c)))
958 #define TCR_PTR(a) ((void *)TCR_8(a))
959 #define TCW_PTR(a, b) TCW_8((a), (b))
960 #define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_8(a))
961 #define TCW_SYNC_PTR(a, b) TCW_SYNC_8((a), (b))
962 #define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_8((a), (b), (c)))
970 #define FTN_TRUE TRUE
974 #define FTN_FALSE FALSE
977 typedef void (*microtask_t)(
int *gtid,
int *npr, ...);
979 #ifdef USE_VOLATILE_CAST
980 #define VOLATILE_CAST(x) (volatile x)
982 #define VOLATILE_CAST(x) (x)
985 #define KMP_WAIT __kmp_wait_4
986 #define KMP_WAIT_PTR __kmp_wait_4_ptr
987 #define KMP_EQ __kmp_eq_4
988 #define KMP_NEQ __kmp_neq_4
989 #define KMP_LT __kmp_lt_4
990 #define KMP_GE __kmp_ge_4
991 #define KMP_LE __kmp_le_4
995 #if (KMP_ARCH_X86_64 || KMP_ARCH_PPC64) && KMP_OS_LINUX
996 #define STATIC_EFI2_WORKAROUND
998 #define STATIC_EFI2_WORKAROUND static
1002 #ifndef KMP_USE_BGET
1003 #define KMP_USE_BGET 1
1007 #ifndef USE_CMPXCHG_FIX
1008 #define USE_CMPXCHG_FIX 1
1012 #define KMP_USE_DYNAMIC_LOCK 1
1016 #if KMP_USE_DYNAMIC_LOCK
1018 #define KMP_USE_TSX (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !KMP_COMPILER_MSVC
1019 #ifdef KMP_USE_ADAPTIVE_LOCKS
1020 #undef KMP_USE_ADAPTIVE_LOCKS
1022 #define KMP_USE_ADAPTIVE_LOCKS KMP_USE_TSX
1026 #if KMP_STATS_ENABLED
1027 #define KMP_HAVE_TICK_TIME \
1028 (KMP_OS_LINUX && (KMP_MIC || KMP_ARCH_X86 || KMP_ARCH_X86_64))
1032 enum kmp_warnings_level {
1033 kmp_warnings_off = 0,
1035 kmp_warnings_explicit = 6,
1036 kmp_warnings_verbose
1044 #include "kmp_safe_c_api.h"
1047 #define KMP_ATOMIC_LD(p, order) (p)->load(std::memory_order_##order)
1048 #define KMP_ATOMIC_OP(op, p, v, order) (p)->op(v, std::memory_order_##order)
1051 #define KMP_ATOMIC_LD_ACQ(p) KMP_ATOMIC_LD(p, acquire)
1052 #define KMP_ATOMIC_LD_RLX(p) KMP_ATOMIC_LD(p, relaxed)
1053 #define KMP_ATOMIC_ST_REL(p, v) KMP_ATOMIC_OP(store, p, v, release)
1054 #define KMP_ATOMIC_ST_RLX(p, v) KMP_ATOMIC_OP(store, p, v, relaxed)
1057 #define KMP_ATOMIC_ADD(p, v) KMP_ATOMIC_OP(fetch_add, p, v, acq_rel)
1058 #define KMP_ATOMIC_SUB(p, v) KMP_ATOMIC_OP(fetch_sub, p, v, acq_rel)
1059 #define KMP_ATOMIC_AND(p, v) KMP_ATOMIC_OP(fetch_and, p, v, acq_rel)
1060 #define KMP_ATOMIC_OR(p, v) KMP_ATOMIC_OP(fetch_or, p, v, acq_rel)
1061 #define KMP_ATOMIC_INC(p) KMP_ATOMIC_OP(fetch_add, p, 1, acq_rel)
1062 #define KMP_ATOMIC_DEC(p) KMP_ATOMIC_OP(fetch_sub, p, 1, acq_rel)
1063 #define KMP_ATOMIC_ADD_RLX(p, v) KMP_ATOMIC_OP(fetch_add, p, v, relaxed)
1064 #define KMP_ATOMIC_INC_RLX(p) KMP_ATOMIC_OP(fetch_add, p, 1, relaxed)
1067 template <
typename T>
1068 bool __kmp_atomic_compare_store(std::atomic<T> *p, T expected, T desired) {
1069 return p->compare_exchange_strong(
1070 expected, desired, std::memory_order_acq_rel, std::memory_order_relaxed);
1073 template <
typename T>
1074 bool __kmp_atomic_compare_store_acq(std::atomic<T> *p, T expected, T desired) {
1075 return p->compare_exchange_strong(
1076 expected, desired, std::memory_order_acquire, std::memory_order_relaxed);
1079 template <
typename T>
1080 bool __kmp_atomic_compare_store_rel(std::atomic<T> *p, T expected, T desired) {
1081 return p->compare_exchange_strong(
1082 expected, desired, std::memory_order_release, std::memory_order_relaxed);
1087 extern void *__kmp_lookup_symbol(
const char *name);
1088 #define KMP_DLSYM(name) __kmp_lookup_symbol(name)
1089 #define KMP_DLSYM_NEXT(name) nullptr
1091 #define KMP_DLSYM(name) dlsym(RTLD_DEFAULT, name)
1092 #define KMP_DLSYM_NEXT(name) dlsym(RTLD_NEXT, name)