LLVM OpenMP* Runtime Library
Loading...
Searching...
No Matches
kmp_os.h
1/*
2 * kmp_os.h -- KPTS runtime header file.
3 */
4
5//===----------------------------------------------------------------------===//
6//
7// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8// See https://llvm.org/LICENSE.txt for license information.
9// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10//
11//===----------------------------------------------------------------------===//
12
13#ifndef KMP_OS_H
14#define KMP_OS_H
15
16#include "kmp_config.h"
17#include <atomic>
18#include <stdarg.h>
19#include <stdlib.h>
20#include <string.h>
21
22#define KMP_FTN_PLAIN 1
23#define KMP_FTN_APPEND 2
24#define KMP_FTN_UPPER 3
25/*
26#define KMP_FTN_PREPEND 4
27#define KMP_FTN_UAPPEND 5
28*/
29
30#define KMP_PTR_SKIP (sizeof(void *))
31
32/* -------------------------- Compiler variations ------------------------ */
33
34#define KMP_OFF 0
35#define KMP_ON 1
36
37#define KMP_MEM_CONS_VOLATILE 0
38#define KMP_MEM_CONS_FENCE 1
39
40#ifndef KMP_MEM_CONS_MODEL
41#define KMP_MEM_CONS_MODEL KMP_MEM_CONS_VOLATILE
42#endif
43
44#ifndef __has_cpp_attribute
45#define __has_cpp_attribute(x) 0
46#endif
47
48#ifndef __has_attribute
49#define __has_attribute(x) 0
50#endif
51
52/* ------------------------- Compiler recognition ---------------------- */
53#define KMP_COMPILER_ICC 0
54#define KMP_COMPILER_GCC 0
55#define KMP_COMPILER_CLANG 0
56#define KMP_COMPILER_MSVC 0
57#define KMP_COMPILER_ICX 0
58
59#if __INTEL_CLANG_COMPILER
60#undef KMP_COMPILER_ICX
61#define KMP_COMPILER_ICX 1
62#elif defined(__INTEL_COMPILER)
63#undef KMP_COMPILER_ICC
64#define KMP_COMPILER_ICC 1
65#elif defined(__clang__)
66#undef KMP_COMPILER_CLANG
67#define KMP_COMPILER_CLANG 1
68#elif defined(__GNUC__)
69#undef KMP_COMPILER_GCC
70#define KMP_COMPILER_GCC 1
71#elif defined(_MSC_VER)
72#undef KMP_COMPILER_MSVC
73#define KMP_COMPILER_MSVC 1
74#else
75#error Unknown compiler
76#endif
77
78#if (KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_FREEBSD || KMP_OS_AIX) && \
79 !KMP_OS_WASI
80#define KMP_AFFINITY_SUPPORTED 1
81#if KMP_OS_WINDOWS && KMP_ARCH_X86_64
82#define KMP_GROUP_AFFINITY 1
83#else
84#define KMP_GROUP_AFFINITY 0
85#endif
86#else
87#define KMP_AFFINITY_SUPPORTED 0
88#define KMP_GROUP_AFFINITY 0
89#endif
90
91#if (KMP_OS_LINUX || (KMP_OS_FREEBSD && __FreeBSD_version >= 1301000))
92#define KMP_HAVE_SCHED_GETCPU 1
93#else
94#define KMP_HAVE_SCHED_GETCPU 0
95#endif
96
97/* Check for quad-precision extension. */
98#define KMP_HAVE_QUAD 0
99#if KMP_ARCH_X86 || KMP_ARCH_X86_64
100#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
101/* _Quad is already defined for icc */
102#undef KMP_HAVE_QUAD
103#define KMP_HAVE_QUAD 1
104#elif KMP_COMPILER_CLANG
105/* Clang doesn't support a software-implemented
106 128-bit extended precision type yet */
107typedef long double _Quad;
108#elif KMP_COMPILER_GCC
109/* GCC on NetBSD lacks __multc3/__divtc3 builtins needed for quad until
110 NetBSD 10.0 which ships with GCC 10.5 */
111#if (!KMP_OS_NETBSD || __GNUC__ >= 10)
112typedef __float128 _Quad;
113#undef KMP_HAVE_QUAD
114#define KMP_HAVE_QUAD 1
115#endif
116#elif KMP_COMPILER_MSVC
117typedef long double _Quad;
118#endif
119#else
120#if __LDBL_MAX_EXP__ >= 16384 && KMP_COMPILER_GCC
121typedef long double _Quad;
122#undef KMP_HAVE_QUAD
123#define KMP_HAVE_QUAD 1
124#endif
125#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
126
127#define KMP_USE_X87CONTROL 0
128#if KMP_OS_WINDOWS
129#define KMP_END_OF_LINE "\r\n"
130typedef char kmp_int8;
131typedef unsigned char kmp_uint8;
132typedef short kmp_int16;
133typedef unsigned short kmp_uint16;
134typedef int kmp_int32;
135typedef unsigned int kmp_uint32;
136#define KMP_INT32_SPEC "d"
137#define KMP_UINT32_SPEC "u"
138#ifndef KMP_STRUCT64
139typedef __int64 kmp_int64;
140typedef unsigned __int64 kmp_uint64;
141#define KMP_INT64_SPEC "I64d"
142#define KMP_UINT64_SPEC "I64u"
143#else
144struct kmp_struct64 {
145 kmp_int32 a, b;
146};
147typedef struct kmp_struct64 kmp_int64;
148typedef struct kmp_struct64 kmp_uint64;
149/* Not sure what to use for KMP_[U]INT64_SPEC here */
150#endif
151#if KMP_ARCH_X86 && KMP_MSVC_COMPAT
152#undef KMP_USE_X87CONTROL
153#define KMP_USE_X87CONTROL 1
154#endif
155#if KMP_ARCH_X86_64 || KMP_ARCH_AARCH64
156#define KMP_INTPTR 1
157typedef __int64 kmp_intptr_t;
158typedef unsigned __int64 kmp_uintptr_t;
159#define KMP_INTPTR_SPEC "I64d"
160#define KMP_UINTPTR_SPEC "I64u"
161#endif
162#endif /* KMP_OS_WINDOWS */
163
164#if KMP_OS_UNIX
165#define KMP_END_OF_LINE "\n"
166typedef char kmp_int8;
167typedef unsigned char kmp_uint8;
168typedef short kmp_int16;
169typedef unsigned short kmp_uint16;
170typedef int kmp_int32;
171typedef unsigned int kmp_uint32;
172typedef long long kmp_int64;
173typedef unsigned long long kmp_uint64;
174#define KMP_INT32_SPEC "d"
175#define KMP_UINT32_SPEC "u"
176#define KMP_INT64_SPEC "lld"
177#define KMP_UINT64_SPEC "llu"
178#endif /* KMP_OS_UNIX */
179
180#if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS || KMP_ARCH_WASM || \
181 KMP_ARCH_PPC
182#define KMP_SIZE_T_SPEC KMP_UINT32_SPEC
183#elif KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || \
184 KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 || \
185 KMP_ARCH_VE || KMP_ARCH_S390X
186#define KMP_SIZE_T_SPEC KMP_UINT64_SPEC
187#else
188#error "Can't determine size_t printf format specifier."
189#endif
190
191#if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_WASM || KMP_ARCH_PPC
192#define KMP_SIZE_T_MAX (0xFFFFFFFF)
193#else
194#define KMP_SIZE_T_MAX (0xFFFFFFFFFFFFFFFF)
195#endif
196
197typedef size_t kmp_size_t;
198typedef float kmp_real32;
199typedef double kmp_real64;
200
201#ifndef KMP_INTPTR
202#define KMP_INTPTR 1
203typedef long kmp_intptr_t;
204typedef unsigned long kmp_uintptr_t;
205#define KMP_INTPTR_SPEC "ld"
206#define KMP_UINTPTR_SPEC "lu"
207#endif
208
209#ifdef BUILD_I8
210typedef kmp_int64 kmp_int;
211typedef kmp_uint64 kmp_uint;
212#else
213typedef kmp_int32 kmp_int;
214typedef kmp_uint32 kmp_uint;
215#endif /* BUILD_I8 */
216#define KMP_INT_MAX ((kmp_int32)0x7FFFFFFF)
217#define KMP_INT_MIN ((kmp_int32)0x80000000)
218
219// stdarg handling
220#if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64 || KMP_ARCH_WASM) && \
221 (KMP_OS_FREEBSD || KMP_OS_LINUX || KMP_OS_WASI)
222typedef va_list *kmp_va_list;
223#define kmp_va_deref(ap) (*(ap))
224#define kmp_va_addr_of(ap) (&(ap))
225#else
226typedef va_list kmp_va_list;
227#define kmp_va_deref(ap) (ap)
228#define kmp_va_addr_of(ap) (ap)
229#endif
230
231#ifdef __cplusplus
232// macros to cast out qualifiers and to re-interpret types
233#define CCAST(type, var) const_cast<type>(var)
234#define RCAST(type, var) reinterpret_cast<type>(var)
235//-------------------------------------------------------------------------
236// template for debug prints specification ( d, u, lld, llu ), and to obtain
237// signed/unsigned flavors of a type
238template <typename T> struct traits_t {};
239// int
240template <> struct traits_t<signed int> {
241 typedef signed int signed_t;
242 typedef unsigned int unsigned_t;
243 typedef double floating_t;
244 static char const *spec;
245 static const signed_t max_value = 0x7fffffff;
246 static const signed_t min_value = 0x80000000;
247 static const int type_size = sizeof(signed_t);
248};
249// unsigned int
250template <> struct traits_t<unsigned int> {
251 typedef signed int signed_t;
252 typedef unsigned int unsigned_t;
253 typedef double floating_t;
254 static char const *spec;
255 static const unsigned_t max_value = 0xffffffff;
256 static const unsigned_t min_value = 0x00000000;
257 static const int type_size = sizeof(unsigned_t);
258};
259// long
260template <> struct traits_t<signed long> {
261 typedef signed long signed_t;
262 typedef unsigned long unsigned_t;
263 typedef long double floating_t;
264 static char const *spec;
265 static const int type_size = sizeof(signed_t);
266};
267// long long
268template <> struct traits_t<signed long long> {
269 typedef signed long long signed_t;
270 typedef unsigned long long unsigned_t;
271 typedef long double floating_t;
272 static char const *spec;
273 static const signed_t max_value = 0x7fffffffffffffffLL;
274 static const signed_t min_value = 0x8000000000000000LL;
275 static const int type_size = sizeof(signed_t);
276};
277// unsigned long long
278template <> struct traits_t<unsigned long long> {
279 typedef signed long long signed_t;
280 typedef unsigned long long unsigned_t;
281 typedef long double floating_t;
282 static char const *spec;
283 static const unsigned_t max_value = 0xffffffffffffffffLL;
284 static const unsigned_t min_value = 0x0000000000000000LL;
285 static const int type_size = sizeof(unsigned_t);
286};
287//-------------------------------------------------------------------------
288#else
289#define CCAST(type, var) (type)(var)
290#define RCAST(type, var) (type)(var)
291#endif // __cplusplus
292
293#define KMP_EXPORT extern /* export declaration in guide libraries */
294
295#if __GNUC__ >= 4 && !defined(__MINGW32__)
296#define __forceinline __inline
297#endif
298
299/* Check if the OS/arch can support user-level mwait */
300// All mwait code tests for UMWAIT first, so it should only fall back to ring3
301// MWAIT for KNL.
302#define KMP_HAVE_MWAIT \
303 ((KMP_ARCH_X86 || KMP_ARCH_X86_64) && (KMP_OS_LINUX || KMP_OS_WINDOWS) && \
304 !KMP_MIC2)
305#define KMP_HAVE_UMWAIT \
306 ((KMP_ARCH_X86 || KMP_ARCH_X86_64) && (KMP_OS_LINUX || KMP_OS_WINDOWS) && \
307 !KMP_MIC)
308
309#if KMP_OS_WINDOWS
310// Don't include everything related to NT status code, we'll do that explicitly
311#define WIN32_NO_STATUS
312#include <windows.h>
313
314static inline int KMP_GET_PAGE_SIZE(void) {
315 SYSTEM_INFO si;
316 GetSystemInfo(&si);
317 return si.dwPageSize;
318}
319#else
320#define KMP_GET_PAGE_SIZE() getpagesize()
321#endif
322
323#define PAGE_ALIGNED(_addr) \
324 (!((size_t)_addr & (size_t)(KMP_GET_PAGE_SIZE() - 1)))
325#define ALIGN_TO_PAGE(x) \
326 (void *)(((size_t)(x)) & ~((size_t)(KMP_GET_PAGE_SIZE() - 1)))
327
328/* ---------- Support for cache alignment, padding, etc. ----------------*/
329
330#ifdef __cplusplus
331extern "C" {
332#endif // __cplusplus
333
334#define INTERNODE_CACHE_LINE 4096 /* for multi-node systems */
335
336/* Define the default size of the cache line */
337#ifndef CACHE_LINE
338#define CACHE_LINE 128 /* cache line size in bytes */
339#else
340#if (CACHE_LINE < 64) && !defined(KMP_OS_DARWIN)
341// 2006-02-13: This produces too many warnings on OS X*. Disable for now
342#warning CACHE_LINE is too small.
343#endif
344#endif /* CACHE_LINE */
345
346#define KMP_CACHE_PREFETCH(ADDR) /* nothing */
347
348// Define attribute that indicates that the fall through from the previous
349// case label is intentional and should not be diagnosed by a compiler
350// Code from libcxx/include/__config
351// Use a function like macro to imply that it must be followed by a semicolon
352#if __cplusplus > 201402L && __has_cpp_attribute(fallthrough)
353#define KMP_FALLTHROUGH() [[fallthrough]]
354// icc cannot properly tell this attribute is absent so force off
355#elif KMP_COMPILER_ICC
356#define KMP_FALLTHROUGH() ((void)0)
357#elif __has_cpp_attribute(clang::fallthrough)
358#define KMP_FALLTHROUGH() [[clang::fallthrough]]
359#elif __has_attribute(fallthrough) || __GNUC__ >= 7
360#define KMP_FALLTHROUGH() __attribute__((__fallthrough__))
361#else
362#define KMP_FALLTHROUGH() ((void)0)
363#endif
364
365#if KMP_HAVE_ATTRIBUTE_WAITPKG
366#define KMP_ATTRIBUTE_TARGET_WAITPKG __attribute__((target("waitpkg")))
367#else
368#define KMP_ATTRIBUTE_TARGET_WAITPKG /* Nothing */
369#endif
370
371#if KMP_HAVE_ATTRIBUTE_RTM
372#define KMP_ATTRIBUTE_TARGET_RTM __attribute__((target("rtm")))
373#else
374#define KMP_ATTRIBUTE_TARGET_RTM /* Nothing */
375#endif
376
377// Define attribute that indicates a function does not return
378#if __cplusplus >= 201103L
379#define KMP_NORETURN [[noreturn]]
380#elif KMP_OS_WINDOWS
381#define KMP_NORETURN __declspec(noreturn)
382#else
383#define KMP_NORETURN __attribute__((noreturn))
384#endif
385
386#if KMP_OS_WINDOWS && KMP_MSVC_COMPAT
387#define KMP_ALIGN(bytes) __declspec(align(bytes))
388#define KMP_THREAD_LOCAL __declspec(thread)
389#define KMP_ALIAS /* Nothing */
390#else
391#define KMP_ALIGN(bytes) __attribute__((aligned(bytes)))
392#define KMP_THREAD_LOCAL __thread
393#define KMP_ALIAS(alias_of) __attribute__((alias(alias_of)))
394#endif
395
396#if KMP_HAVE_WEAK_ATTRIBUTE && !KMP_DYNAMIC_LIB
397#define KMP_WEAK_ATTRIBUTE_EXTERNAL __attribute__((weak))
398#else
399#define KMP_WEAK_ATTRIBUTE_EXTERNAL /* Nothing */
400#endif
401
402#if KMP_HAVE_WEAK_ATTRIBUTE
403#define KMP_WEAK_ATTRIBUTE_INTERNAL __attribute__((weak))
404#else
405#define KMP_WEAK_ATTRIBUTE_INTERNAL /* Nothing */
406#endif
407
408// Define KMP_VERSION_SYMBOL and KMP_EXPAND_NAME
409#ifndef KMP_STR
410#define KMP_STR(x) _KMP_STR(x)
411#define _KMP_STR(x) #x
412#endif
413
414#ifdef KMP_USE_VERSION_SYMBOLS
415// If using versioned symbols, KMP_EXPAND_NAME prepends
416// __kmp_api_ to the real API name
417#define KMP_EXPAND_NAME(api_name) _KMP_EXPAND_NAME(api_name)
418#define _KMP_EXPAND_NAME(api_name) __kmp_api_##api_name
419#define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str) \
420 _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, "VERSION")
421#define _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, default_ver) \
422 __typeof__(__kmp_api_##api_name) __kmp_api_##api_name##_##ver_num##_alias \
423 __attribute__((alias(KMP_STR(__kmp_api_##api_name)))); \
424 __asm__( \
425 ".symver " KMP_STR(__kmp_api_##api_name##_##ver_num##_alias) "," KMP_STR( \
426 api_name) "@" ver_str "\n\t"); \
427 __asm__(".symver " KMP_STR(__kmp_api_##api_name) "," KMP_STR( \
428 api_name) "@@" default_ver "\n\t")
429
430#define KMP_VERSION_OMPC_SYMBOL(apic_name, api_name, ver_num, ver_str) \
431 _KMP_VERSION_OMPC_SYMBOL(apic_name, api_name, ver_num, ver_str, "VERSION")
432#define _KMP_VERSION_OMPC_SYMBOL(apic_name, api_name, ver_num, ver_str, \
433 default_ver) \
434 __typeof__(__kmp_api_##apic_name) __kmp_api_##apic_name##_##ver_num##_alias \
435 __attribute__((alias(KMP_STR(__kmp_api_##apic_name)))); \
436 __asm__(".symver " KMP_STR(__kmp_api_##apic_name) "," KMP_STR( \
437 apic_name) "@@" default_ver "\n\t"); \
438 __asm__( \
439 ".symver " KMP_STR(__kmp_api_##apic_name##_##ver_num##_alias) "," KMP_STR( \
440 api_name) "@" ver_str "\n\t")
441
442#else // KMP_USE_VERSION_SYMBOLS
443#define KMP_EXPAND_NAME(api_name) api_name
444#define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str) /* Nothing */
445#define KMP_VERSION_OMPC_SYMBOL(apic_name, api_name, ver_num, \
446 ver_str) /* Nothing */
447#endif // KMP_USE_VERSION_SYMBOLS
448
449/* Temporary note: if performance testing of this passes, we can remove
450 all references to KMP_DO_ALIGN and replace with KMP_ALIGN. */
451#define KMP_DO_ALIGN(bytes) KMP_ALIGN(bytes)
452#define KMP_ALIGN_CACHE KMP_ALIGN(CACHE_LINE)
453#define KMP_ALIGN_CACHE_INTERNODE KMP_ALIGN(INTERNODE_CACHE_LINE)
454
455/* General purpose fence types for memory operations */
456enum kmp_mem_fence_type {
457 kmp_no_fence, /* No memory fence */
458 kmp_acquire_fence, /* Acquire (read) memory fence */
459 kmp_release_fence, /* Release (write) memory fence */
460 kmp_full_fence /* Full (read+write) memory fence */
461};
462
463// Synchronization primitives
464
465#if KMP_ASM_INTRINS && KMP_OS_WINDOWS && !((KMP_ARCH_AARCH64 || KMP_ARCH_ARM) && (KMP_COMPILER_CLANG || KMP_COMPILER_GCC))
466
467#if KMP_MSVC_COMPAT && !KMP_COMPILER_CLANG
468#pragma intrinsic(InterlockedExchangeAdd)
469#pragma intrinsic(InterlockedCompareExchange)
470#pragma intrinsic(InterlockedExchange)
471#if !(KMP_COMPILER_ICX && KMP_32_BIT_ARCH)
472#pragma intrinsic(InterlockedExchange64)
473#endif
474#endif
475
476// Using InterlockedIncrement / InterlockedDecrement causes a library loading
477// ordering problem, so we use InterlockedExchangeAdd instead.
478#define KMP_TEST_THEN_INC32(p) InterlockedExchangeAdd((volatile long *)(p), 1)
479#define KMP_TEST_THEN_INC_ACQ32(p) \
480 InterlockedExchangeAdd((volatile long *)(p), 1)
481#define KMP_TEST_THEN_ADD4_32(p) InterlockedExchangeAdd((volatile long *)(p), 4)
482#define KMP_TEST_THEN_ADD4_ACQ32(p) \
483 InterlockedExchangeAdd((volatile long *)(p), 4)
484#define KMP_TEST_THEN_DEC32(p) InterlockedExchangeAdd((volatile long *)(p), -1)
485#define KMP_TEST_THEN_DEC_ACQ32(p) \
486 InterlockedExchangeAdd((volatile long *)(p), -1)
487#define KMP_TEST_THEN_ADD32(p, v) \
488 InterlockedExchangeAdd((volatile long *)(p), (v))
489
490#define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
491 InterlockedCompareExchange((volatile long *)(p), (long)(sv), (long)(cv))
492
493#define KMP_XCHG_FIXED32(p, v) \
494 InterlockedExchange((volatile long *)(p), (long)(v))
495#define KMP_XCHG_FIXED64(p, v) \
496 InterlockedExchange64((volatile kmp_int64 *)(p), (kmp_int64)(v))
497
498inline kmp_real32 KMP_XCHG_REAL32(volatile kmp_real32 *p, kmp_real32 v) {
499 kmp_int32 tmp = InterlockedExchange((volatile long *)p, *(long *)&v);
500 return *(kmp_real32 *)&tmp;
501}
502
503#define KMP_TEST_THEN_OR8(p, v) __kmp_test_then_or8((p), (v))
504#define KMP_TEST_THEN_AND8(p, v) __kmp_test_then_and8((p), (v))
505#define KMP_TEST_THEN_OR32(p, v) __kmp_test_then_or32((p), (v))
506#define KMP_TEST_THEN_AND32(p, v) __kmp_test_then_and32((p), (v))
507#define KMP_TEST_THEN_OR64(p, v) __kmp_test_then_or64((p), (v))
508#define KMP_TEST_THEN_AND64(p, v) __kmp_test_then_and64((p), (v))
509
510extern kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 v);
511extern kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 v);
512extern kmp_int32 __kmp_test_then_add32(volatile kmp_int32 *p, kmp_int32 v);
513extern kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 v);
514extern kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 v);
515extern kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 v);
516extern kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 v);
517extern kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 v);
518
519#if KMP_ARCH_AARCH64 && KMP_COMPILER_MSVC && !KMP_COMPILER_CLANG
520#define KMP_TEST_THEN_INC64(p) _InterlockedExchangeAdd64((p), 1LL)
521#define KMP_TEST_THEN_INC_ACQ64(p) _InterlockedExchangeAdd64_acq((p), 1LL)
522#define KMP_TEST_THEN_ADD4_64(p) _InterlockedExchangeAdd64((p), 4LL)
523// #define KMP_TEST_THEN_ADD4_ACQ64(p) _InterlockedExchangeAdd64_acq((p), 4LL)
524// #define KMP_TEST_THEN_DEC64(p) _InterlockedExchangeAdd64((p), -1LL)
525// #define KMP_TEST_THEN_DEC_ACQ64(p) _InterlockedExchangeAdd64_acq((p), -1LL)
526// #define KMP_TEST_THEN_ADD8(p, v) _InterlockedExchangeAdd8((p), (v))
527#define KMP_TEST_THEN_ADD64(p, v) _InterlockedExchangeAdd64((p), (v))
528
529#define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
530 __kmp_compare_and_store_acq8((p), (cv), (sv))
531#define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
532 __kmp_compare_and_store_rel8((p), (cv), (sv))
533#define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
534 __kmp_compare_and_store_acq16((p), (cv), (sv))
535/*
536#define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
537 __kmp_compare_and_store_rel16((p), (cv), (sv))
538*/
539#define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
540 __kmp_compare_and_store_acq32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
541 (kmp_int32)(sv))
542#define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
543 __kmp_compare_and_store_rel32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
544 (kmp_int32)(sv))
545#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
546 __kmp_compare_and_store_acq64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
547 (kmp_int64)(sv))
548#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
549 __kmp_compare_and_store_rel64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
550 (kmp_int64)(sv))
551#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
552 __kmp_compare_and_store_ptr((void *volatile *)(p), (void *)(cv), (void *)(sv))
553
554// KMP_COMPARE_AND_STORE expects this order: pointer, compare, exchange
555// _InterlockedCompareExchange expects this order: pointer, exchange, compare
556// KMP_COMPARE_AND_STORE also returns a bool indicating a successful write. A
557// write is successful if the return value of _InterlockedCompareExchange is the
558// same as the compare value.
559inline kmp_int8 __kmp_compare_and_store_acq8(volatile kmp_int8 *p, kmp_int8 cv,
560 kmp_int8 sv) {
561 return _InterlockedCompareExchange8_acq(p, sv, cv) == cv;
562}
563
564inline kmp_int8 __kmp_compare_and_store_rel8(volatile kmp_int8 *p, kmp_int8 cv,
565 kmp_int8 sv) {
566 return _InterlockedCompareExchange8_rel(p, sv, cv) == cv;
567}
568
569inline kmp_int16 __kmp_compare_and_store_acq16(volatile kmp_int16 *p,
570 kmp_int16 cv, kmp_int16 sv) {
571 return _InterlockedCompareExchange16_acq(p, sv, cv) == cv;
572}
573
574inline kmp_int16 __kmp_compare_and_store_rel16(volatile kmp_int16 *p,
575 kmp_int16 cv, kmp_int16 sv) {
576 return _InterlockedCompareExchange16_rel(p, sv, cv) == cv;
577}
578
579inline kmp_int32 __kmp_compare_and_store_acq32(volatile kmp_int32 *p,
580 kmp_int32 cv, kmp_int32 sv) {
581 return _InterlockedCompareExchange_acq((volatile long *)p, sv, cv) == cv;
582}
583
584inline kmp_int32 __kmp_compare_and_store_rel32(volatile kmp_int32 *p,
585 kmp_int32 cv, kmp_int32 sv) {
586 return _InterlockedCompareExchange_rel((volatile long *)p, sv, cv) == cv;
587}
588
589inline kmp_int32 __kmp_compare_and_store_acq64(volatile kmp_int64 *p,
590 kmp_int64 cv, kmp_int64 sv) {
591 return _InterlockedCompareExchange64_acq(p, sv, cv) == cv;
592}
593
594inline kmp_int32 __kmp_compare_and_store_rel64(volatile kmp_int64 *p,
595 kmp_int64 cv, kmp_int64 sv) {
596 return _InterlockedCompareExchange64_rel(p, sv, cv) == cv;
597}
598
599inline kmp_int32 __kmp_compare_and_store_ptr(void *volatile *p, void *cv,
600 void *sv) {
601 return _InterlockedCompareExchangePointer(p, sv, cv) == cv;
602}
603
604// The _RET versions return the value instead of a bool
605
606#define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
607 _InterlockedCompareExchange8((p), (sv), (cv))
608#define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
609 _InterlockedCompareExchange16((p), (sv), (cv))
610
611#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
612 _InterlockedCompareExchange64((volatile kmp_int64 *)(p), (kmp_int64)(sv), \
613 (kmp_int64)(cv))
614
615
616#define KMP_XCHG_FIXED8(p, v) \
617 _InterlockedExchange8((volatile kmp_int8 *)(p), (kmp_int8)(v));
618#define KMP_XCHG_FIXED16(p, v) _InterlockedExchange16((p), (v));
619#define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
620
621inline kmp_real64 __kmp_xchg_real64(volatile kmp_real64 *p, kmp_real64 v) {
622 kmp_int64 tmp = _InterlockedExchange64((volatile kmp_int64 *)p, *(kmp_int64
623 *)&v); return *(kmp_real64 *)&tmp;
624}
625
626#else // !KMP_ARCH_AARCH64
627
628// Routines that we still need to implement in assembly.
629extern kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 v);
630
631extern kmp_int8 __kmp_compare_and_store8(volatile kmp_int8 *p, kmp_int8 cv,
632 kmp_int8 sv);
633extern kmp_int16 __kmp_compare_and_store16(volatile kmp_int16 *p, kmp_int16 cv,
634 kmp_int16 sv);
635extern kmp_int32 __kmp_compare_and_store32(volatile kmp_int32 *p, kmp_int32 cv,
636 kmp_int32 sv);
637extern kmp_int32 __kmp_compare_and_store64(volatile kmp_int64 *p, kmp_int64 cv,
638 kmp_int64 sv);
639extern kmp_int8 __kmp_compare_and_store_ret8(volatile kmp_int8 *p, kmp_int8 cv,
640 kmp_int8 sv);
641extern kmp_int16 __kmp_compare_and_store_ret16(volatile kmp_int16 *p,
642 kmp_int16 cv, kmp_int16 sv);
643extern kmp_int32 __kmp_compare_and_store_ret32(volatile kmp_int32 *p,
644 kmp_int32 cv, kmp_int32 sv);
645extern kmp_int64 __kmp_compare_and_store_ret64(volatile kmp_int64 *p,
646 kmp_int64 cv, kmp_int64 sv);
647
648extern kmp_int8 __kmp_xchg_fixed8(volatile kmp_int8 *p, kmp_int8 v);
649extern kmp_int16 __kmp_xchg_fixed16(volatile kmp_int16 *p, kmp_int16 v);
650extern kmp_int32 __kmp_xchg_fixed32(volatile kmp_int32 *p, kmp_int32 v);
651extern kmp_int64 __kmp_xchg_fixed64(volatile kmp_int64 *p, kmp_int64 v);
652extern kmp_real32 __kmp_xchg_real32(volatile kmp_real32 *p, kmp_real32 v);
653extern kmp_real64 __kmp_xchg_real64(volatile kmp_real64 *p, kmp_real64 v);
654
655//#define KMP_TEST_THEN_INC32(p) __kmp_test_then_add32((p), 1)
656//#define KMP_TEST_THEN_INC_ACQ32(p) __kmp_test_then_add32((p), 1)
657#define KMP_TEST_THEN_INC64(p) __kmp_test_then_add64((p), 1LL)
658#define KMP_TEST_THEN_INC_ACQ64(p) __kmp_test_then_add64((p), 1LL)
659//#define KMP_TEST_THEN_ADD4_32(p) __kmp_test_then_add32((p), 4)
660//#define KMP_TEST_THEN_ADD4_ACQ32(p) __kmp_test_then_add32((p), 4)
661#define KMP_TEST_THEN_ADD4_64(p) __kmp_test_then_add64((p), 4LL)
662#define KMP_TEST_THEN_ADD4_ACQ64(p) __kmp_test_then_add64((p), 4LL)
663//#define KMP_TEST_THEN_DEC32(p) __kmp_test_then_add32((p), -1)
664//#define KMP_TEST_THEN_DEC_ACQ32(p) __kmp_test_then_add32((p), -1)
665#define KMP_TEST_THEN_DEC64(p) __kmp_test_then_add64((p), -1LL)
666#define KMP_TEST_THEN_DEC_ACQ64(p) __kmp_test_then_add64((p), -1LL)
667//#define KMP_TEST_THEN_ADD32(p, v) __kmp_test_then_add32((p), (v))
668#define KMP_TEST_THEN_ADD8(p, v) __kmp_test_then_add8((p), (v))
669#define KMP_TEST_THEN_ADD64(p, v) __kmp_test_then_add64((p), (v))
670
671
672#define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
673 __kmp_compare_and_store8((p), (cv), (sv))
674#define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
675 __kmp_compare_and_store8((p), (cv), (sv))
676#define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
677 __kmp_compare_and_store16((p), (cv), (sv))
678#define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
679 __kmp_compare_and_store16((p), (cv), (sv))
680#define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
681 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
682 (kmp_int32)(sv))
683#define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
684 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
685 (kmp_int32)(sv))
686#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
687 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
688 (kmp_int64)(sv))
689#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
690 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
691 (kmp_int64)(sv))
692
693#if KMP_ARCH_X86
694#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
695 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
696 (kmp_int32)(sv))
697#else /* 64 bit pointers */
698#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
699 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
700 (kmp_int64)(sv))
701#endif /* KMP_ARCH_X86 */
702
703#define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
704 __kmp_compare_and_store_ret8((p), (cv), (sv))
705#define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
706 __kmp_compare_and_store_ret16((p), (cv), (sv))
707#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
708 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
709 (kmp_int64)(sv))
710
711#define KMP_XCHG_FIXED8(p, v) \
712 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
713#define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
714//#define KMP_XCHG_FIXED32(p, v) __kmp_xchg_fixed32((p), (v));
715//#define KMP_XCHG_FIXED64(p, v) __kmp_xchg_fixed64((p), (v));
716//#define KMP_XCHG_REAL32(p, v) __kmp_xchg_real32((p), (v));
717#define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
718#endif
719
720#elif (KMP_ASM_INTRINS && KMP_OS_UNIX) || !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
721
722/* cast p to correct type so that proper intrinsic will be used */
723#define KMP_TEST_THEN_INC32(p) \
724 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
725#define KMP_TEST_THEN_INC_ACQ32(p) \
726 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
727#if KMP_ARCH_MIPS
728#define KMP_TEST_THEN_INC64(p) \
729 __atomic_fetch_add((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
730#define KMP_TEST_THEN_INC_ACQ64(p) \
731 __atomic_fetch_add((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
732#else
733#define KMP_TEST_THEN_INC64(p) \
734 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
735#define KMP_TEST_THEN_INC_ACQ64(p) \
736 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
737#endif
738#define KMP_TEST_THEN_ADD4_32(p) \
739 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
740#define KMP_TEST_THEN_ADD4_ACQ32(p) \
741 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
742#if KMP_ARCH_MIPS
743#define KMP_TEST_THEN_ADD4_64(p) \
744 __atomic_fetch_add((volatile kmp_int64 *)(p), 4LL, __ATOMIC_SEQ_CST)
745#define KMP_TEST_THEN_ADD4_ACQ64(p) \
746 __atomic_fetch_add((volatile kmp_int64 *)(p), 4LL, __ATOMIC_SEQ_CST)
747#define KMP_TEST_THEN_DEC64(p) \
748 __atomic_fetch_sub((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
749#define KMP_TEST_THEN_DEC_ACQ64(p) \
750 __atomic_fetch_sub((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
751#else
752#define KMP_TEST_THEN_ADD4_64(p) \
753 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
754#define KMP_TEST_THEN_ADD4_ACQ64(p) \
755 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
756#define KMP_TEST_THEN_DEC64(p) \
757 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
758#define KMP_TEST_THEN_DEC_ACQ64(p) \
759 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
760#endif
761#define KMP_TEST_THEN_DEC32(p) \
762 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
763#define KMP_TEST_THEN_DEC_ACQ32(p) \
764 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
765#define KMP_TEST_THEN_ADD8(p, v) \
766 __sync_fetch_and_add((volatile kmp_int8 *)(p), (kmp_int8)(v))
767#define KMP_TEST_THEN_ADD32(p, v) \
768 __sync_fetch_and_add((volatile kmp_int32 *)(p), (kmp_int32)(v))
769#if KMP_ARCH_MIPS
770#define KMP_TEST_THEN_ADD64(p, v) \
771 __atomic_fetch_add((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
772 __ATOMIC_SEQ_CST)
773#else
774#define KMP_TEST_THEN_ADD64(p, v) \
775 __sync_fetch_and_add((volatile kmp_int64 *)(p), (kmp_int64)(v))
776#endif
777
778#define KMP_TEST_THEN_OR8(p, v) \
779 __sync_fetch_and_or((volatile kmp_int8 *)(p), (kmp_int8)(v))
780#define KMP_TEST_THEN_AND8(p, v) \
781 __sync_fetch_and_and((volatile kmp_int8 *)(p), (kmp_int8)(v))
782#define KMP_TEST_THEN_OR32(p, v) \
783 __sync_fetch_and_or((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
784#define KMP_TEST_THEN_AND32(p, v) \
785 __sync_fetch_and_and((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
786#if KMP_ARCH_MIPS
787#define KMP_TEST_THEN_OR64(p, v) \
788 __atomic_fetch_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
789 __ATOMIC_SEQ_CST)
790#define KMP_TEST_THEN_AND64(p, v) \
791 __atomic_fetch_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
792 __ATOMIC_SEQ_CST)
793#else
794#define KMP_TEST_THEN_OR64(p, v) \
795 __sync_fetch_and_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
796#define KMP_TEST_THEN_AND64(p, v) \
797 __sync_fetch_and_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
798#endif
799
800#define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
801 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
802 (kmp_uint8)(sv))
803#define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
804 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
805 (kmp_uint8)(sv))
806#define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
807 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
808 (kmp_uint16)(sv))
809#define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
810 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
811 (kmp_uint16)(sv))
812#define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
813 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
814 (kmp_uint32)(sv))
815#define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
816 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
817 (kmp_uint32)(sv))
818#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
819 __sync_bool_compare_and_swap((void *volatile *)(p), (void *)(cv), \
820 (void *)(sv))
821
822#define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
823 __sync_val_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
824 (kmp_uint8)(sv))
825#define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
826 __sync_val_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
827 (kmp_uint16)(sv))
828#define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
829 __sync_val_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
830 (kmp_uint32)(sv))
831#if KMP_ARCH_MIPS
832static inline bool mips_sync_bool_compare_and_swap(volatile kmp_uint64 *p,
833 kmp_uint64 cv,
834 kmp_uint64 sv) {
835 return __atomic_compare_exchange(p, &cv, &sv, false, __ATOMIC_SEQ_CST,
836 __ATOMIC_SEQ_CST);
837}
838static inline bool mips_sync_val_compare_and_swap(volatile kmp_uint64 *p,
839 kmp_uint64 cv,
840 kmp_uint64 sv) {
841 __atomic_compare_exchange(p, &cv, &sv, false, __ATOMIC_SEQ_CST,
842 __ATOMIC_SEQ_CST);
843 return cv;
844}
845#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
846 mips_sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), \
847 (kmp_uint64)(cv), (kmp_uint64)(sv))
848#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
849 mips_sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), \
850 (kmp_uint64)(cv), (kmp_uint64)(sv))
851#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
852 mips_sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
853 (kmp_uint64)(sv))
854#else
855#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
856 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
857 (kmp_uint64)(sv))
858#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
859 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
860 (kmp_uint64)(sv))
861#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
862 __sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
863 (kmp_uint64)(sv))
864#endif
865
866#if KMP_OS_DARWIN && defined(__INTEL_COMPILER) && __INTEL_COMPILER >= 1800
867#define KMP_XCHG_FIXED8(p, v) \
868 __atomic_exchange_1((volatile kmp_uint8 *)(p), (kmp_uint8)(v), \
869 __ATOMIC_SEQ_CST)
870#else
871#define KMP_XCHG_FIXED8(p, v) \
872 __sync_lock_test_and_set((volatile kmp_uint8 *)(p), (kmp_uint8)(v))
873#endif
874#define KMP_XCHG_FIXED16(p, v) \
875 __sync_lock_test_and_set((volatile kmp_uint16 *)(p), (kmp_uint16)(v))
876#define KMP_XCHG_FIXED32(p, v) \
877 __sync_lock_test_and_set((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
878#define KMP_XCHG_FIXED64(p, v) \
879 __sync_lock_test_and_set((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
880
881inline kmp_real32 KMP_XCHG_REAL32(volatile kmp_real32 *p, kmp_real32 v) {
882 volatile kmp_uint32 *up;
883 kmp_uint32 uv;
884 memcpy(&up, &p, sizeof(up));
885 memcpy(&uv, &v, sizeof(uv));
886 kmp_int32 tmp = __sync_lock_test_and_set(up, uv);
887 kmp_real32 ftmp;
888 memcpy(&ftmp, &tmp, sizeof(tmp));
889 return ftmp;
890}
891
892inline kmp_real64 KMP_XCHG_REAL64(volatile kmp_real64 *p, kmp_real64 v) {
893 volatile kmp_uint64 *up;
894 kmp_uint64 uv;
895 memcpy(&up, &p, sizeof(up));
896 memcpy(&uv, &v, sizeof(uv));
897 kmp_int64 tmp = __sync_lock_test_and_set(up, uv);
898 kmp_real64 dtmp;
899 memcpy(&dtmp, &tmp, sizeof(tmp));
900 return dtmp;
901}
902
903#else
904
905extern kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 v);
906extern kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 v);
907extern kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 v);
908extern kmp_int32 __kmp_test_then_add32(volatile kmp_int32 *p, kmp_int32 v);
909extern kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 v);
910extern kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 v);
911extern kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 v);
912extern kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 v);
913extern kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 v);
914
915extern kmp_int8 __kmp_compare_and_store8(volatile kmp_int8 *p, kmp_int8 cv,
916 kmp_int8 sv);
917extern kmp_int16 __kmp_compare_and_store16(volatile kmp_int16 *p, kmp_int16 cv,
918 kmp_int16 sv);
919extern kmp_int32 __kmp_compare_and_store32(volatile kmp_int32 *p, kmp_int32 cv,
920 kmp_int32 sv);
921extern kmp_int32 __kmp_compare_and_store64(volatile kmp_int64 *p, kmp_int64 cv,
922 kmp_int64 sv);
923extern kmp_int8 __kmp_compare_and_store_ret8(volatile kmp_int8 *p, kmp_int8 cv,
924 kmp_int8 sv);
925extern kmp_int16 __kmp_compare_and_store_ret16(volatile kmp_int16 *p,
926 kmp_int16 cv, kmp_int16 sv);
927extern kmp_int32 __kmp_compare_and_store_ret32(volatile kmp_int32 *p,
928 kmp_int32 cv, kmp_int32 sv);
929extern kmp_int64 __kmp_compare_and_store_ret64(volatile kmp_int64 *p,
930 kmp_int64 cv, kmp_int64 sv);
931
932extern kmp_int8 __kmp_xchg_fixed8(volatile kmp_int8 *p, kmp_int8 v);
933extern kmp_int16 __kmp_xchg_fixed16(volatile kmp_int16 *p, kmp_int16 v);
934extern kmp_int32 __kmp_xchg_fixed32(volatile kmp_int32 *p, kmp_int32 v);
935extern kmp_int64 __kmp_xchg_fixed64(volatile kmp_int64 *p, kmp_int64 v);
936extern kmp_real32 __kmp_xchg_real32(volatile kmp_real32 *p, kmp_real32 v);
937extern kmp_real64 __kmp_xchg_real64(volatile kmp_real64 *p, kmp_real64 v);
938
939#define KMP_TEST_THEN_INC32(p) \
940 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
941#define KMP_TEST_THEN_INC_ACQ32(p) \
942 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
943#define KMP_TEST_THEN_INC64(p) \
944 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
945#define KMP_TEST_THEN_INC_ACQ64(p) \
946 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
947#define KMP_TEST_THEN_ADD4_32(p) \
948 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
949#define KMP_TEST_THEN_ADD4_ACQ32(p) \
950 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
951#define KMP_TEST_THEN_ADD4_64(p) \
952 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
953#define KMP_TEST_THEN_ADD4_ACQ64(p) \
954 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
955#define KMP_TEST_THEN_DEC32(p) \
956 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
957#define KMP_TEST_THEN_DEC_ACQ32(p) \
958 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
959#define KMP_TEST_THEN_DEC64(p) \
960 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
961#define KMP_TEST_THEN_DEC_ACQ64(p) \
962 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
963#define KMP_TEST_THEN_ADD8(p, v) \
964 __kmp_test_then_add8((volatile kmp_int8 *)(p), (kmp_int8)(v))
965#define KMP_TEST_THEN_ADD32(p, v) \
966 __kmp_test_then_add32((volatile kmp_int32 *)(p), (kmp_int32)(v))
967#define KMP_TEST_THEN_ADD64(p, v) \
968 __kmp_test_then_add64((volatile kmp_int64 *)(p), (kmp_int64)(v))
969
970#define KMP_TEST_THEN_OR8(p, v) \
971 __kmp_test_then_or8((volatile kmp_int8 *)(p), (kmp_int8)(v))
972#define KMP_TEST_THEN_AND8(p, v) \
973 __kmp_test_then_and8((volatile kmp_int8 *)(p), (kmp_int8)(v))
974#define KMP_TEST_THEN_OR32(p, v) \
975 __kmp_test_then_or32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
976#define KMP_TEST_THEN_AND32(p, v) \
977 __kmp_test_then_and32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
978#define KMP_TEST_THEN_OR64(p, v) \
979 __kmp_test_then_or64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
980#define KMP_TEST_THEN_AND64(p, v) \
981 __kmp_test_then_and64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
982
983#define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
984 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
985 (kmp_int8)(sv))
986#define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
987 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
988 (kmp_int8)(sv))
989#define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
990 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
991 (kmp_int16)(sv))
992#define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
993 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
994 (kmp_int16)(sv))
995#define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
996 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
997 (kmp_int32)(sv))
998#define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
999 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
1000 (kmp_int32)(sv))
1001#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
1002 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
1003 (kmp_int64)(sv))
1004#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
1005 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
1006 (kmp_int64)(sv))
1007
1008#if KMP_ARCH_X86
1009#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
1010 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
1011 (kmp_int32)(sv))
1012#else /* 64 bit pointers */
1013#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
1014 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
1015 (kmp_int64)(sv))
1016#endif /* KMP_ARCH_X86 */
1017
1018#define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
1019 __kmp_compare_and_store_ret8((p), (cv), (sv))
1020#define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
1021 __kmp_compare_and_store_ret16((p), (cv), (sv))
1022#define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
1023 __kmp_compare_and_store_ret32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
1024 (kmp_int32)(sv))
1025#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
1026 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
1027 (kmp_int64)(sv))
1028
1029#define KMP_XCHG_FIXED8(p, v) \
1030 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
1031#define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
1032#define KMP_XCHG_FIXED32(p, v) __kmp_xchg_fixed32((p), (v));
1033#define KMP_XCHG_FIXED64(p, v) __kmp_xchg_fixed64((p), (v));
1034#define KMP_XCHG_REAL32(p, v) __kmp_xchg_real32((p), (v));
1035#define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
1036
1037#endif /* KMP_ASM_INTRINS */
1038
1039/* ------------- relaxed consistency memory model stuff ------------------ */
1040
1041#if KMP_OS_WINDOWS
1042#ifdef __ABSOFT_WIN
1043#define KMP_MB() asm("nop")
1044#define KMP_IMB() asm("nop")
1045#else
1046#define KMP_MB() /* _asm{ nop } */
1047#define KMP_IMB() /* _asm{ nop } */
1048#endif
1049#endif /* KMP_OS_WINDOWS */
1050
1051#if KMP_ARCH_PPC64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS || \
1052 KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 || \
1053 KMP_ARCH_VE || KMP_ARCH_S390X || KMP_ARCH_PPC
1054#if KMP_OS_WINDOWS
1055#undef KMP_MB
1056#define KMP_MB() std::atomic_thread_fence(std::memory_order_seq_cst)
1057#else /* !KMP_OS_WINDOWS */
1058#define KMP_MB() __sync_synchronize()
1059#endif
1060#endif
1061
1062#ifndef KMP_MB
1063#define KMP_MB() /* nothing to do */
1064#endif
1065
1066#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1067#if KMP_MIC
1068// fence-style instructions do not exist, but lock; xaddl $0,(%rsp) can be used.
1069// We shouldn't need it, though, since the ABI rules require that
1070// * If the compiler generates NGO stores it also generates the fence
1071// * If users hand-code NGO stores they should insert the fence
1072// therefore no incomplete unordered stores should be visible.
1073#define KMP_MFENCE() /* Nothing */
1074#define KMP_SFENCE() /* Nothing */
1075#else
1076#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1077#define KMP_MFENCE_() _mm_mfence()
1078#define KMP_SFENCE_() _mm_sfence()
1079#elif KMP_COMPILER_MSVC
1080#define KMP_MFENCE_() MemoryBarrier()
1081#define KMP_SFENCE_() MemoryBarrier()
1082#else
1083#define KMP_MFENCE_() __sync_synchronize()
1084#define KMP_SFENCE_() __sync_synchronize()
1085#endif
1086#define KMP_MFENCE() \
1087 if (UNLIKELY(!__kmp_cpuinfo.initialized)) { \
1088 __kmp_query_cpuid(&__kmp_cpuinfo); \
1089 } \
1090 if (__kmp_cpuinfo.flags.sse2) { \
1091 KMP_MFENCE_(); \
1092 }
1093#define KMP_SFENCE() KMP_SFENCE_()
1094#endif
1095#else
1096#define KMP_MFENCE() KMP_MB()
1097#define KMP_SFENCE() KMP_MB()
1098#endif
1099
1100#ifndef KMP_IMB
1101#define KMP_IMB() /* nothing to do */
1102#endif
1103
1104#ifndef KMP_ST_REL32
1105#define KMP_ST_REL32(A, D) (*(A) = (D))
1106#endif
1107
1108#ifndef KMP_ST_REL64
1109#define KMP_ST_REL64(A, D) (*(A) = (D))
1110#endif
1111
1112#ifndef KMP_LD_ACQ32
1113#define KMP_LD_ACQ32(A) (*(A))
1114#endif
1115
1116#ifndef KMP_LD_ACQ64
1117#define KMP_LD_ACQ64(A) (*(A))
1118#endif
1119
1120/* ------------------------------------------------------------------------ */
1121// FIXME - maybe this should this be
1122//
1123// #define TCR_4(a) (*(volatile kmp_int32 *)(&a))
1124// #define TCW_4(a,b) (a) = (*(volatile kmp_int32 *)&(b))
1125//
1126// #define TCR_8(a) (*(volatile kmp_int64 *)(a))
1127// #define TCW_8(a,b) (a) = (*(volatile kmp_int64 *)(&b))
1128//
1129// I'm fairly certain this is the correct thing to do, but I'm afraid
1130// of performance regressions.
1131
1132#define TCR_1(a) (a)
1133#define TCW_1(a, b) (a) = (b)
1134#define TCR_4(a) (a)
1135#define TCW_4(a, b) (a) = (b)
1136#define TCI_4(a) (++(a))
1137#define TCD_4(a) (--(a))
1138#define TCR_8(a) (a)
1139#define TCW_8(a, b) (a) = (b)
1140#define TCI_8(a) (++(a))
1141#define TCD_8(a) (--(a))
1142#define TCR_SYNC_4(a) (a)
1143#define TCW_SYNC_4(a, b) (a) = (b)
1144#define TCX_SYNC_4(a, b, c) \
1145 KMP_COMPARE_AND_STORE_REL32((volatile kmp_int32 *)(volatile void *)&(a), \
1146 (kmp_int32)(b), (kmp_int32)(c))
1147#define TCR_SYNC_8(a) (a)
1148#define TCW_SYNC_8(a, b) (a) = (b)
1149#define TCX_SYNC_8(a, b, c) \
1150 KMP_COMPARE_AND_STORE_REL64((volatile kmp_int64 *)(volatile void *)&(a), \
1151 (kmp_int64)(b), (kmp_int64)(c))
1152
1153#if KMP_ARCH_X86 || KMP_ARCH_MIPS || KMP_ARCH_WASM || KMP_ARCH_PPC
1154// What about ARM?
1155#define TCR_PTR(a) ((void *)TCR_4(a))
1156#define TCW_PTR(a, b) TCW_4((a), (b))
1157#define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_4(a))
1158#define TCW_SYNC_PTR(a, b) TCW_SYNC_4((a), (b))
1159#define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_4((a), (b), (c)))
1160
1161#else /* 64 bit pointers */
1162
1163#define TCR_PTR(a) ((void *)TCR_8(a))
1164#define TCW_PTR(a, b) TCW_8((a), (b))
1165#define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_8(a))
1166#define TCW_SYNC_PTR(a, b) TCW_SYNC_8((a), (b))
1167#define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_8((a), (b), (c)))
1168
1169#endif /* KMP_ARCH_X86 */
1170
1171/* If these FTN_{TRUE,FALSE} values change, may need to change several places
1172 where they are used to check that language is Fortran, not C. */
1173
1174#ifndef FTN_TRUE
1175#define FTN_TRUE TRUE
1176#endif
1177
1178#ifndef FTN_FALSE
1179#define FTN_FALSE FALSE
1180#endif
1181
1182typedef void (*microtask_t)(int *gtid, int *npr, ...);
1183
1184#ifdef USE_VOLATILE_CAST
1185#define VOLATILE_CAST(x) (volatile x)
1186#else
1187#define VOLATILE_CAST(x) (x)
1188#endif
1189
1190#define KMP_WAIT __kmp_wait_4
1191#define KMP_WAIT_PTR __kmp_wait_4_ptr
1192#define KMP_EQ __kmp_eq_4
1193#define KMP_NEQ __kmp_neq_4
1194#define KMP_LT __kmp_lt_4
1195#define KMP_GE __kmp_ge_4
1196#define KMP_LE __kmp_le_4
1197
1198/* Workaround for Intel(R) 64 code gen bug when taking address of static array
1199 * (Intel(R) 64 Tracker #138) */
1200#if (KMP_ARCH_X86_64 || KMP_ARCH_PPC64) && KMP_OS_LINUX
1201#define STATIC_EFI2_WORKAROUND
1202#else
1203#define STATIC_EFI2_WORKAROUND static
1204#endif
1205
1206// Support of BGET usage
1207#ifndef KMP_USE_BGET
1208#define KMP_USE_BGET 1
1209#endif
1210
1211// Switches for OSS builds
1212#ifndef USE_CMPXCHG_FIX
1213#define USE_CMPXCHG_FIX 1
1214#endif
1215
1216// Enable dynamic user lock
1217#define KMP_USE_DYNAMIC_LOCK 1
1218
1219// Enable Intel(R) Transactional Synchronization Extensions (Intel(R) TSX) if
1220// dynamic user lock is turned on
1221#if KMP_USE_DYNAMIC_LOCK
1222// Visual studio can't handle the asm sections in this code
1223#define KMP_USE_TSX (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !KMP_COMPILER_MSVC
1224#ifdef KMP_USE_ADAPTIVE_LOCKS
1225#undef KMP_USE_ADAPTIVE_LOCKS
1226#endif
1227#define KMP_USE_ADAPTIVE_LOCKS KMP_USE_TSX
1228#endif
1229
1230// Enable tick time conversion of ticks to seconds
1231#if KMP_STATS_ENABLED
1232#define KMP_HAVE_TICK_TIME \
1233 (KMP_OS_LINUX && (KMP_MIC || KMP_ARCH_X86 || KMP_ARCH_X86_64))
1234#endif
1235
1236// Warning levels
1237enum kmp_warnings_level {
1238 kmp_warnings_off = 0, /* No warnings */
1239 kmp_warnings_low, /* Minimal warnings (default) */
1240 kmp_warnings_explicit = 6, /* Explicitly set to ON - more warnings */
1241 kmp_warnings_verbose /* reserved */
1242};
1243
1244#ifdef __cplusplus
1245} // extern "C"
1246#endif // __cplusplus
1247
1248// Safe C API
1249#include "kmp_safe_c_api.h"
1250
1251// Macros for C++11 atomic functions
1252#define KMP_ATOMIC_LD(p, order) (p)->load(std::memory_order_##order)
1253#define KMP_ATOMIC_OP(op, p, v, order) (p)->op(v, std::memory_order_##order)
1254
1255// For non-default load/store
1256#define KMP_ATOMIC_LD_ACQ(p) KMP_ATOMIC_LD(p, acquire)
1257#define KMP_ATOMIC_LD_RLX(p) KMP_ATOMIC_LD(p, relaxed)
1258#define KMP_ATOMIC_ST_REL(p, v) KMP_ATOMIC_OP(store, p, v, release)
1259#define KMP_ATOMIC_ST_RLX(p, v) KMP_ATOMIC_OP(store, p, v, relaxed)
1260
1261// For non-default fetch_<op>
1262#define KMP_ATOMIC_ADD(p, v) KMP_ATOMIC_OP(fetch_add, p, v, acq_rel)
1263#define KMP_ATOMIC_SUB(p, v) KMP_ATOMIC_OP(fetch_sub, p, v, acq_rel)
1264#define KMP_ATOMIC_AND(p, v) KMP_ATOMIC_OP(fetch_and, p, v, acq_rel)
1265#define KMP_ATOMIC_OR(p, v) KMP_ATOMIC_OP(fetch_or, p, v, acq_rel)
1266#define KMP_ATOMIC_INC(p) KMP_ATOMIC_OP(fetch_add, p, 1, acq_rel)
1267#define KMP_ATOMIC_DEC(p) KMP_ATOMIC_OP(fetch_sub, p, 1, acq_rel)
1268#define KMP_ATOMIC_ADD_RLX(p, v) KMP_ATOMIC_OP(fetch_add, p, v, relaxed)
1269#define KMP_ATOMIC_INC_RLX(p) KMP_ATOMIC_OP(fetch_add, p, 1, relaxed)
1270
1271// Callers of the following functions cannot see the side effect on "expected".
1272template <typename T>
1273bool __kmp_atomic_compare_store(std::atomic<T> *p, T expected, T desired) {
1274 return p->compare_exchange_strong(
1275 expected, desired, std::memory_order_acq_rel, std::memory_order_relaxed);
1276}
1277
1278template <typename T>
1279bool __kmp_atomic_compare_store_acq(std::atomic<T> *p, T expected, T desired) {
1280 return p->compare_exchange_strong(
1281 expected, desired, std::memory_order_acquire, std::memory_order_relaxed);
1282}
1283
1284template <typename T>
1285bool __kmp_atomic_compare_store_rel(std::atomic<T> *p, T expected, T desired) {
1286 return p->compare_exchange_strong(
1287 expected, desired, std::memory_order_release, std::memory_order_relaxed);
1288}
1289
1290// Symbol lookup on Linux/Windows
1291#if KMP_OS_WINDOWS
1292extern void *__kmp_lookup_symbol(const char *name, bool next = false);
1293#define KMP_DLSYM(name) __kmp_lookup_symbol(name)
1294#define KMP_DLSYM_NEXT(name) __kmp_lookup_symbol(name, true)
1295#elif KMP_OS_WASI
1296#define KMP_DLSYM(name) nullptr
1297#define KMP_DLSYM_NEXT(name) nullptr
1298#else
1299#define KMP_DLSYM(name) dlsym(RTLD_DEFAULT, name)
1300#define KMP_DLSYM_NEXT(name) dlsym(RTLD_NEXT, name)
1301#endif
1302
1303// MSVC doesn't have this, but clang/clang-cl does.
1304#ifndef __has_builtin
1305#define __has_builtin(x) 0
1306#endif
1307
1308// Same as LLVM_BUILTIN_UNREACHABLE. States that it is UB to reach this point.
1309#if __has_builtin(__builtin_unreachable) || defined(__GNUC__)
1310#define KMP_BUILTIN_UNREACHABLE __builtin_unreachable()
1311#elif defined(_MSC_VER)
1312#define KMP_BUILTIN_UNREACHABLE __assume(false)
1313#else
1314#define KMP_BUILTIN_UNREACHABLE
1315#endif
1316
1317#endif /* KMP_OS_H */