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 <stdlib.h>
18#include <atomic>
19
20#define KMP_FTN_PLAIN 1
21#define KMP_FTN_APPEND 2
22#define KMP_FTN_UPPER 3
23/*
24#define KMP_FTN_PREPEND 4
25#define KMP_FTN_UAPPEND 5
26*/
27
28#define KMP_PTR_SKIP (sizeof(void *))
29
30/* -------------------------- Compiler variations ------------------------ */
31
32#define KMP_OFF 0
33#define KMP_ON 1
34
35#define KMP_MEM_CONS_VOLATILE 0
36#define KMP_MEM_CONS_FENCE 1
37
38#ifndef KMP_MEM_CONS_MODEL
39#define KMP_MEM_CONS_MODEL KMP_MEM_CONS_VOLATILE
40#endif
41
42#ifndef __has_cpp_attribute
43#define __has_cpp_attribute(x) 0
44#endif
45
46#ifndef __has_attribute
47#define __has_attribute(x) 0
48#endif
49
50/* ------------------------- Compiler recognition ---------------------- */
51#define KMP_COMPILER_ICC 0
52#define KMP_COMPILER_GCC 0
53#define KMP_COMPILER_CLANG 0
54#define KMP_COMPILER_MSVC 0
55
56#if defined(__INTEL_COMPILER)
57#undef KMP_COMPILER_ICC
58#define KMP_COMPILER_ICC 1
59#elif defined(__clang__)
60#undef KMP_COMPILER_CLANG
61#define KMP_COMPILER_CLANG 1
62#elif defined(__GNUC__)
63#undef KMP_COMPILER_GCC
64#define KMP_COMPILER_GCC 1
65#elif defined(_MSC_VER)
66#undef KMP_COMPILER_MSVC
67#define KMP_COMPILER_MSVC 1
68#else
69#error Unknown compiler
70#endif
71
72#if (KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_FREEBSD) && !KMP_OS_CNK
73#define KMP_AFFINITY_SUPPORTED 1
74#if KMP_OS_WINDOWS && KMP_ARCH_X86_64
75#define KMP_GROUP_AFFINITY 1
76#else
77#define KMP_GROUP_AFFINITY 0
78#endif
79#else
80#define KMP_AFFINITY_SUPPORTED 0
81#define KMP_GROUP_AFFINITY 0
82#endif
83
84/* Check for quad-precision extension. */
85#define KMP_HAVE_QUAD 0
86#if KMP_ARCH_X86 || KMP_ARCH_X86_64
87#if KMP_COMPILER_ICC
88/* _Quad is already defined for icc */
89#undef KMP_HAVE_QUAD
90#define KMP_HAVE_QUAD 1
91#elif KMP_COMPILER_CLANG
92/* Clang doesn't support a software-implemented
93 128-bit extended precision type yet */
94typedef long double _Quad;
95#elif KMP_COMPILER_GCC
96/* GCC on NetBSD lacks __multc3/__divtc3 builtins needed for quad */
97#if !KMP_OS_NETBSD
98typedef __float128 _Quad;
99#undef KMP_HAVE_QUAD
100#define KMP_HAVE_QUAD 1
101#endif
102#elif KMP_COMPILER_MSVC
103typedef long double _Quad;
104#endif
105#else
106#if __LDBL_MAX_EXP__ >= 16384 && KMP_COMPILER_GCC
107typedef long double _Quad;
108#undef KMP_HAVE_QUAD
109#define KMP_HAVE_QUAD 1
110#endif
111#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
112
113#define KMP_USE_X87CONTROL 0
114#if KMP_OS_WINDOWS
115#define KMP_END_OF_LINE "\r\n"
116typedef char kmp_int8;
117typedef unsigned char kmp_uint8;
118typedef short kmp_int16;
119typedef unsigned short kmp_uint16;
120typedef int kmp_int32;
121typedef unsigned int kmp_uint32;
122#define KMP_INT32_SPEC "d"
123#define KMP_UINT32_SPEC "u"
124#ifndef KMP_STRUCT64
125typedef __int64 kmp_int64;
126typedef unsigned __int64 kmp_uint64;
127#define KMP_INT64_SPEC "I64d"
128#define KMP_UINT64_SPEC "I64u"
129#else
130struct kmp_struct64 {
131 kmp_int32 a, b;
132};
133typedef struct kmp_struct64 kmp_int64;
134typedef struct kmp_struct64 kmp_uint64;
135/* Not sure what to use for KMP_[U]INT64_SPEC here */
136#endif
137#if KMP_ARCH_X86 && KMP_MSVC_COMPAT
138#undef KMP_USE_X87CONTROL
139#define KMP_USE_X87CONTROL 1
140#endif
141#if KMP_ARCH_X86_64
142#define KMP_INTPTR 1
143typedef __int64 kmp_intptr_t;
144typedef unsigned __int64 kmp_uintptr_t;
145#define KMP_INTPTR_SPEC "I64d"
146#define KMP_UINTPTR_SPEC "I64u"
147#endif
148#endif /* KMP_OS_WINDOWS */
149
150#if KMP_OS_UNIX
151#define KMP_END_OF_LINE "\n"
152typedef char kmp_int8;
153typedef unsigned char kmp_uint8;
154typedef short kmp_int16;
155typedef unsigned short kmp_uint16;
156typedef int kmp_int32;
157typedef unsigned int kmp_uint32;
158typedef long long kmp_int64;
159typedef unsigned long long kmp_uint64;
160#define KMP_INT32_SPEC "d"
161#define KMP_UINT32_SPEC "u"
162#define KMP_INT64_SPEC "lld"
163#define KMP_UINT64_SPEC "llu"
164#endif /* KMP_OS_UNIX */
165
166#if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS
167#define KMP_SIZE_T_SPEC KMP_UINT32_SPEC
168#elif KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || \
169 KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64
170#define KMP_SIZE_T_SPEC KMP_UINT64_SPEC
171#else
172#error "Can't determine size_t printf format specifier."
173#endif
174
175#if KMP_ARCH_X86
176#define KMP_SIZE_T_MAX (0xFFFFFFFF)
177#else
178#define KMP_SIZE_T_MAX (0xFFFFFFFFFFFFFFFF)
179#endif
180
181typedef size_t kmp_size_t;
182typedef float kmp_real32;
183typedef double kmp_real64;
184
185#ifndef KMP_INTPTR
186#define KMP_INTPTR 1
187typedef long kmp_intptr_t;
188typedef unsigned long kmp_uintptr_t;
189#define KMP_INTPTR_SPEC "ld"
190#define KMP_UINTPTR_SPEC "lu"
191#endif
192
193#ifdef BUILD_I8
194typedef kmp_int64 kmp_int;
195typedef kmp_uint64 kmp_uint;
196#else
197typedef kmp_int32 kmp_int;
198typedef kmp_uint32 kmp_uint;
199#endif /* BUILD_I8 */
200#define KMP_INT_MAX ((kmp_int32)0x7FFFFFFF)
201#define KMP_INT_MIN ((kmp_int32)0x80000000)
202
203#ifdef __cplusplus
204// macros to cast out qualifiers and to re-interpret types
205#define CCAST(type, var) const_cast<type>(var)
206#define RCAST(type, var) reinterpret_cast<type>(var)
207//-------------------------------------------------------------------------
208// template for debug prints specification ( d, u, lld, llu ), and to obtain
209// signed/unsigned flavors of a type
210template <typename T> struct traits_t {};
211// int
212template <> struct traits_t<signed int> {
213 typedef signed int signed_t;
214 typedef unsigned int unsigned_t;
215 typedef double floating_t;
216 static char const *spec;
217 static const signed_t max_value = 0x7fffffff;
218 static const signed_t min_value = 0x80000000;
219 static const int type_size = sizeof(signed_t);
220};
221// unsigned int
222template <> struct traits_t<unsigned int> {
223 typedef signed int signed_t;
224 typedef unsigned int unsigned_t;
225 typedef double floating_t;
226 static char const *spec;
227 static const unsigned_t max_value = 0xffffffff;
228 static const unsigned_t min_value = 0x00000000;
229 static const int type_size = sizeof(unsigned_t);
230};
231// long
232template <> struct traits_t<signed long> {
233 typedef signed long signed_t;
234 typedef unsigned long unsigned_t;
235 typedef long double floating_t;
236 static char const *spec;
237 static const int type_size = sizeof(signed_t);
238};
239// long long
240template <> struct traits_t<signed long long> {
241 typedef signed long long signed_t;
242 typedef unsigned long long unsigned_t;
243 typedef long double floating_t;
244 static char const *spec;
245 static const signed_t max_value = 0x7fffffffffffffffLL;
246 static const signed_t min_value = 0x8000000000000000LL;
247 static const int type_size = sizeof(signed_t);
248};
249// unsigned long long
250template <> struct traits_t<unsigned long long> {
251 typedef signed long long signed_t;
252 typedef unsigned long long unsigned_t;
253 typedef long double floating_t;
254 static char const *spec;
255 static const unsigned_t max_value = 0xffffffffffffffffLL;
256 static const unsigned_t min_value = 0x0000000000000000LL;
257 static const int type_size = sizeof(unsigned_t);
258};
259//-------------------------------------------------------------------------
260#else
261#define CCAST(type, var) (type)(var)
262#define RCAST(type, var) (type)(var)
263#endif // __cplusplus
264
265#define KMP_EXPORT extern /* export declaration in guide libraries */
266
267#if __GNUC__ >= 4 && !defined(__MINGW32__)
268#define __forceinline __inline
269#endif
270
271#if KMP_OS_WINDOWS
272#include <windows.h>
273
274static inline int KMP_GET_PAGE_SIZE(void) {
275 SYSTEM_INFO si;
276 GetSystemInfo(&si);
277 return si.dwPageSize;
278}
279#else
280#define KMP_GET_PAGE_SIZE() getpagesize()
281#endif
282
283#define PAGE_ALIGNED(_addr) \
284 (!((size_t)_addr & (size_t)(KMP_GET_PAGE_SIZE() - 1)))
285#define ALIGN_TO_PAGE(x) \
286 (void *)(((size_t)(x)) & ~((size_t)(KMP_GET_PAGE_SIZE() - 1)))
287
288/* ---------- Support for cache alignment, padding, etc. ----------------*/
289
290#ifdef __cplusplus
291extern "C" {
292#endif // __cplusplus
293
294#define INTERNODE_CACHE_LINE 4096 /* for multi-node systems */
295
296/* Define the default size of the cache line */
297#ifndef CACHE_LINE
298#define CACHE_LINE 128 /* cache line size in bytes */
299#else
300#if (CACHE_LINE < 64) && !defined(KMP_OS_DARWIN)
301// 2006-02-13: This produces too many warnings on OS X*. Disable for now
302#warning CACHE_LINE is too small.
303#endif
304#endif /* CACHE_LINE */
305
306#define KMP_CACHE_PREFETCH(ADDR) /* nothing */
307
308// Define attribute that indicates that the fall through from the previous
309// case label is intentional and should not be diagnosed by a compiler
310// Code from libcxx/include/__config
311// Use a function like macro to imply that it must be followed by a semicolon
312#if __cplusplus > 201402L && __has_cpp_attribute(fallthrough)
313# define KMP_FALLTHROUGH() [[fallthrough]]
314#elif __has_cpp_attribute(clang::fallthrough)
315# define KMP_FALLTHROUGH() [[clang::fallthrough]]
316#elif __has_attribute(fallthrough) || __GNUC__ >= 7
317# define KMP_FALLTHROUGH() __attribute__((__fallthrough__))
318#else
319# define KMP_FALLTHROUGH() ((void)0)
320#endif
321
322// Define attribute that indicates a function does not return
323#if __cplusplus >= 201103L
324#define KMP_NORETURN [[noreturn]]
325#elif KMP_OS_WINDOWS
326#define KMP_NORETURN __declspec(noreturn)
327#else
328#define KMP_NORETURN __attribute__((noreturn))
329#endif
330
331#if KMP_OS_WINDOWS && KMP_MSVC_COMPAT
332#define KMP_ALIGN(bytes) __declspec(align(bytes))
333#define KMP_THREAD_LOCAL __declspec(thread)
334#define KMP_ALIAS /* Nothing */
335#else
336#define KMP_ALIGN(bytes) __attribute__((aligned(bytes)))
337#define KMP_THREAD_LOCAL __thread
338#define KMP_ALIAS(alias_of) __attribute__((alias(alias_of)))
339#endif
340
341#if KMP_HAVE_WEAK_ATTRIBUTE && !KMP_DYNAMIC_LIB
342#define KMP_WEAK_ATTRIBUTE_EXTERNAL __attribute__((weak))
343#else
344#define KMP_WEAK_ATTRIBUTE_EXTERNAL /* Nothing */
345#endif
346
347#if KMP_HAVE_WEAK_ATTRIBUTE
348#define KMP_WEAK_ATTRIBUTE_INTERNAL __attribute__((weak))
349#else
350#define KMP_WEAK_ATTRIBUTE_INTERNAL /* Nothing */
351#endif
352
353// Define KMP_VERSION_SYMBOL and KMP_EXPAND_NAME
354#ifndef KMP_STR
355#define KMP_STR(x) _KMP_STR(x)
356#define _KMP_STR(x) #x
357#endif
358
359#ifdef KMP_USE_VERSION_SYMBOLS
360// If using versioned symbols, KMP_EXPAND_NAME prepends
361// __kmp_api_ to the real API name
362#define KMP_EXPAND_NAME(api_name) _KMP_EXPAND_NAME(api_name)
363#define _KMP_EXPAND_NAME(api_name) __kmp_api_##api_name
364#define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str) \
365 _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, "VERSION")
366#define _KMP_VERSION_SYMBOL(api_name, ver_num, ver_str, default_ver) \
367 __typeof__(__kmp_api_##api_name) __kmp_api_##api_name##_##ver_num##_alias \
368 __attribute__((alias(KMP_STR(__kmp_api_##api_name)))); \
369 __asm__( \
370 ".symver " KMP_STR(__kmp_api_##api_name##_##ver_num##_alias) "," KMP_STR( \
371 api_name) "@" ver_str "\n\t"); \
372 __asm__(".symver " KMP_STR(__kmp_api_##api_name) "," KMP_STR( \
373 api_name) "@@" default_ver "\n\t")
374#else // KMP_USE_VERSION_SYMBOLS
375#define KMP_EXPAND_NAME(api_name) api_name
376#define KMP_VERSION_SYMBOL(api_name, ver_num, ver_str) /* Nothing */
377#endif // KMP_USE_VERSION_SYMBOLS
378
379/* Temporary note: if performance testing of this passes, we can remove
380 all references to KMP_DO_ALIGN and replace with KMP_ALIGN. */
381#define KMP_DO_ALIGN(bytes) KMP_ALIGN(bytes)
382#define KMP_ALIGN_CACHE KMP_ALIGN(CACHE_LINE)
383#define KMP_ALIGN_CACHE_INTERNODE KMP_ALIGN(INTERNODE_CACHE_LINE)
384
385/* General purpose fence types for memory operations */
386enum kmp_mem_fence_type {
387 kmp_no_fence, /* No memory fence */
388 kmp_acquire_fence, /* Acquire (read) memory fence */
389 kmp_release_fence, /* Release (write) memory fence */
390 kmp_full_fence /* Full (read+write) memory fence */
391};
392
393// Synchronization primitives
394
395#if KMP_ASM_INTRINS && KMP_OS_WINDOWS
396
397#if KMP_MSVC_COMPAT && !KMP_COMPILER_CLANG
398#pragma intrinsic(InterlockedExchangeAdd)
399#pragma intrinsic(InterlockedCompareExchange)
400#pragma intrinsic(InterlockedExchange)
401#pragma intrinsic(InterlockedExchange64)
402#endif
403
404// Using InterlockedIncrement / InterlockedDecrement causes a library loading
405// ordering problem, so we use InterlockedExchangeAdd instead.
406#define KMP_TEST_THEN_INC32(p) InterlockedExchangeAdd((volatile long *)(p), 1)
407#define KMP_TEST_THEN_INC_ACQ32(p) \
408 InterlockedExchangeAdd((volatile long *)(p), 1)
409#define KMP_TEST_THEN_ADD4_32(p) InterlockedExchangeAdd((volatile long *)(p), 4)
410#define KMP_TEST_THEN_ADD4_ACQ32(p) \
411 InterlockedExchangeAdd((volatile long *)(p), 4)
412#define KMP_TEST_THEN_DEC32(p) InterlockedExchangeAdd((volatile long *)(p), -1)
413#define KMP_TEST_THEN_DEC_ACQ32(p) \
414 InterlockedExchangeAdd((volatile long *)(p), -1)
415#define KMP_TEST_THEN_ADD32(p, v) \
416 InterlockedExchangeAdd((volatile long *)(p), (v))
417
418#define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
419 InterlockedCompareExchange((volatile long *)(p), (long)(sv), (long)(cv))
420
421#define KMP_XCHG_FIXED32(p, v) \
422 InterlockedExchange((volatile long *)(p), (long)(v))
423#define KMP_XCHG_FIXED64(p, v) \
424 InterlockedExchange64((volatile kmp_int64 *)(p), (kmp_int64)(v))
425
426inline kmp_real32 KMP_XCHG_REAL32(volatile kmp_real32 *p, kmp_real32 v) {
427 kmp_int32 tmp = InterlockedExchange((volatile long *)p, *(long *)&v);
428 return *(kmp_real32 *)&tmp;
429}
430
431// Routines that we still need to implement in assembly.
432extern kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 v);
433extern kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 v);
434extern kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 v);
435extern kmp_int32 __kmp_test_then_add32(volatile kmp_int32 *p, kmp_int32 v);
436extern kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 v);
437extern kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 v);
438extern kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 v);
439extern kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 v);
440extern kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 v);
441
442extern kmp_int8 __kmp_compare_and_store8(volatile kmp_int8 *p, kmp_int8 cv,
443 kmp_int8 sv);
444extern kmp_int16 __kmp_compare_and_store16(volatile kmp_int16 *p, kmp_int16 cv,
445 kmp_int16 sv);
446extern kmp_int32 __kmp_compare_and_store32(volatile kmp_int32 *p, kmp_int32 cv,
447 kmp_int32 sv);
448extern kmp_int32 __kmp_compare_and_store64(volatile kmp_int64 *p, kmp_int64 cv,
449 kmp_int64 sv);
450extern kmp_int8 __kmp_compare_and_store_ret8(volatile kmp_int8 *p, kmp_int8 cv,
451 kmp_int8 sv);
452extern kmp_int16 __kmp_compare_and_store_ret16(volatile kmp_int16 *p,
453 kmp_int16 cv, kmp_int16 sv);
454extern kmp_int32 __kmp_compare_and_store_ret32(volatile kmp_int32 *p,
455 kmp_int32 cv, kmp_int32 sv);
456extern kmp_int64 __kmp_compare_and_store_ret64(volatile kmp_int64 *p,
457 kmp_int64 cv, kmp_int64 sv);
458
459extern kmp_int8 __kmp_xchg_fixed8(volatile kmp_int8 *p, kmp_int8 v);
460extern kmp_int16 __kmp_xchg_fixed16(volatile kmp_int16 *p, kmp_int16 v);
461extern kmp_int32 __kmp_xchg_fixed32(volatile kmp_int32 *p, kmp_int32 v);
462extern kmp_int64 __kmp_xchg_fixed64(volatile kmp_int64 *p, kmp_int64 v);
463extern kmp_real32 __kmp_xchg_real32(volatile kmp_real32 *p, kmp_real32 v);
464extern kmp_real64 __kmp_xchg_real64(volatile kmp_real64 *p, kmp_real64 v);
465
466//#define KMP_TEST_THEN_INC32(p) __kmp_test_then_add32((p), 1)
467//#define KMP_TEST_THEN_INC_ACQ32(p) __kmp_test_then_add32((p), 1)
468#define KMP_TEST_THEN_INC64(p) __kmp_test_then_add64((p), 1LL)
469#define KMP_TEST_THEN_INC_ACQ64(p) __kmp_test_then_add64((p), 1LL)
470//#define KMP_TEST_THEN_ADD4_32(p) __kmp_test_then_add32((p), 4)
471//#define KMP_TEST_THEN_ADD4_ACQ32(p) __kmp_test_then_add32((p), 4)
472#define KMP_TEST_THEN_ADD4_64(p) __kmp_test_then_add64((p), 4LL)
473#define KMP_TEST_THEN_ADD4_ACQ64(p) __kmp_test_then_add64((p), 4LL)
474//#define KMP_TEST_THEN_DEC32(p) __kmp_test_then_add32((p), -1)
475//#define KMP_TEST_THEN_DEC_ACQ32(p) __kmp_test_then_add32((p), -1)
476#define KMP_TEST_THEN_DEC64(p) __kmp_test_then_add64((p), -1LL)
477#define KMP_TEST_THEN_DEC_ACQ64(p) __kmp_test_then_add64((p), -1LL)
478//#define KMP_TEST_THEN_ADD32(p, v) __kmp_test_then_add32((p), (v))
479#define KMP_TEST_THEN_ADD8(p, v) __kmp_test_then_add8((p), (v))
480#define KMP_TEST_THEN_ADD64(p, v) __kmp_test_then_add64((p), (v))
481
482#define KMP_TEST_THEN_OR8(p, v) __kmp_test_then_or8((p), (v))
483#define KMP_TEST_THEN_AND8(p, v) __kmp_test_then_and8((p), (v))
484#define KMP_TEST_THEN_OR32(p, v) __kmp_test_then_or32((p), (v))
485#define KMP_TEST_THEN_AND32(p, v) __kmp_test_then_and32((p), (v))
486#define KMP_TEST_THEN_OR64(p, v) __kmp_test_then_or64((p), (v))
487#define KMP_TEST_THEN_AND64(p, v) __kmp_test_then_and64((p), (v))
488
489#define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
490 __kmp_compare_and_store8((p), (cv), (sv))
491#define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
492 __kmp_compare_and_store8((p), (cv), (sv))
493#define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
494 __kmp_compare_and_store16((p), (cv), (sv))
495#define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
496 __kmp_compare_and_store16((p), (cv), (sv))
497#define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
498 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
499 (kmp_int32)(sv))
500#define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
501 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
502 (kmp_int32)(sv))
503#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
504 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
505 (kmp_int64)(sv))
506#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
507 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
508 (kmp_int64)(sv))
509
510#if KMP_ARCH_X86
511#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
512 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
513 (kmp_int32)(sv))
514#else /* 64 bit pointers */
515#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
516 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
517 (kmp_int64)(sv))
518#endif /* KMP_ARCH_X86 */
519
520#define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
521 __kmp_compare_and_store_ret8((p), (cv), (sv))
522#define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
523 __kmp_compare_and_store_ret16((p), (cv), (sv))
524#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
525 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
526 (kmp_int64)(sv))
527
528#define KMP_XCHG_FIXED8(p, v) \
529 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
530#define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
531//#define KMP_XCHG_FIXED32(p, v) __kmp_xchg_fixed32((p), (v));
532//#define KMP_XCHG_FIXED64(p, v) __kmp_xchg_fixed64((p), (v));
533//#define KMP_XCHG_REAL32(p, v) __kmp_xchg_real32((p), (v));
534#define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
535
536#elif (KMP_ASM_INTRINS && KMP_OS_UNIX) || !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
537
538/* cast p to correct type so that proper intrinsic will be used */
539#define KMP_TEST_THEN_INC32(p) \
540 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
541#define KMP_TEST_THEN_INC_ACQ32(p) \
542 __sync_fetch_and_add((volatile kmp_int32 *)(p), 1)
543#if KMP_ARCH_MIPS
544#define KMP_TEST_THEN_INC64(p) \
545 __atomic_fetch_add((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
546#define KMP_TEST_THEN_INC_ACQ64(p) \
547 __atomic_fetch_add((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
548#else
549#define KMP_TEST_THEN_INC64(p) \
550 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
551#define KMP_TEST_THEN_INC_ACQ64(p) \
552 __sync_fetch_and_add((volatile kmp_int64 *)(p), 1LL)
553#endif
554#define KMP_TEST_THEN_ADD4_32(p) \
555 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
556#define KMP_TEST_THEN_ADD4_ACQ32(p) \
557 __sync_fetch_and_add((volatile kmp_int32 *)(p), 4)
558#if KMP_ARCH_MIPS
559#define KMP_TEST_THEN_ADD4_64(p) \
560 __atomic_fetch_add((volatile kmp_int64 *)(p), 4LL, __ATOMIC_SEQ_CST)
561#define KMP_TEST_THEN_ADD4_ACQ64(p) \
562 __atomic_fetch_add((volatile kmp_int64 *)(p), 4LL, __ATOMIC_SEQ_CST)
563#define KMP_TEST_THEN_DEC64(p) \
564 __atomic_fetch_sub((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
565#define KMP_TEST_THEN_DEC_ACQ64(p) \
566 __atomic_fetch_sub((volatile kmp_int64 *)(p), 1LL, __ATOMIC_SEQ_CST)
567#else
568#define KMP_TEST_THEN_ADD4_64(p) \
569 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
570#define KMP_TEST_THEN_ADD4_ACQ64(p) \
571 __sync_fetch_and_add((volatile kmp_int64 *)(p), 4LL)
572#define KMP_TEST_THEN_DEC64(p) \
573 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
574#define KMP_TEST_THEN_DEC_ACQ64(p) \
575 __sync_fetch_and_sub((volatile kmp_int64 *)(p), 1LL)
576#endif
577#define KMP_TEST_THEN_DEC32(p) \
578 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
579#define KMP_TEST_THEN_DEC_ACQ32(p) \
580 __sync_fetch_and_sub((volatile kmp_int32 *)(p), 1)
581#define KMP_TEST_THEN_ADD8(p, v) \
582 __sync_fetch_and_add((volatile kmp_int8 *)(p), (kmp_int8)(v))
583#define KMP_TEST_THEN_ADD32(p, v) \
584 __sync_fetch_and_add((volatile kmp_int32 *)(p), (kmp_int32)(v))
585#if KMP_ARCH_MIPS
586#define KMP_TEST_THEN_ADD64(p, v) \
587 __atomic_fetch_add((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
588 __ATOMIC_SEQ_CST)
589#else
590#define KMP_TEST_THEN_ADD64(p, v) \
591 __sync_fetch_and_add((volatile kmp_int64 *)(p), (kmp_int64)(v))
592#endif
593
594#define KMP_TEST_THEN_OR8(p, v) \
595 __sync_fetch_and_or((volatile kmp_int8 *)(p), (kmp_int8)(v))
596#define KMP_TEST_THEN_AND8(p, v) \
597 __sync_fetch_and_and((volatile kmp_int8 *)(p), (kmp_int8)(v))
598#define KMP_TEST_THEN_OR32(p, v) \
599 __sync_fetch_and_or((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
600#define KMP_TEST_THEN_AND32(p, v) \
601 __sync_fetch_and_and((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
602#if KMP_ARCH_MIPS
603#define KMP_TEST_THEN_OR64(p, v) \
604 __atomic_fetch_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
605 __ATOMIC_SEQ_CST)
606#define KMP_TEST_THEN_AND64(p, v) \
607 __atomic_fetch_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v), \
608 __ATOMIC_SEQ_CST)
609#else
610#define KMP_TEST_THEN_OR64(p, v) \
611 __sync_fetch_and_or((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
612#define KMP_TEST_THEN_AND64(p, v) \
613 __sync_fetch_and_and((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
614#endif
615
616#define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
617 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
618 (kmp_uint8)(sv))
619#define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
620 __sync_bool_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
621 (kmp_uint8)(sv))
622#define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
623 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
624 (kmp_uint16)(sv))
625#define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
626 __sync_bool_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
627 (kmp_uint16)(sv))
628#define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
629 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
630 (kmp_uint32)(sv))
631#define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
632 __sync_bool_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
633 (kmp_uint32)(sv))
634#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
635 __sync_bool_compare_and_swap((void *volatile *)(p), (void *)(cv), \
636 (void *)(sv))
637
638#define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
639 __sync_val_compare_and_swap((volatile kmp_uint8 *)(p), (kmp_uint8)(cv), \
640 (kmp_uint8)(sv))
641#define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
642 __sync_val_compare_and_swap((volatile kmp_uint16 *)(p), (kmp_uint16)(cv), \
643 (kmp_uint16)(sv))
644#define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
645 __sync_val_compare_and_swap((volatile kmp_uint32 *)(p), (kmp_uint32)(cv), \
646 (kmp_uint32)(sv))
647#if KMP_ARCH_MIPS
648static inline bool mips_sync_bool_compare_and_swap(
649 volatile kmp_uint64 *p, kmp_uint64 cv, kmp_uint64 sv) {
650 return __atomic_compare_exchange(p, &cv, &sv, false, __ATOMIC_SEQ_CST,
651 __ATOMIC_SEQ_CST);
652}
653static inline bool mips_sync_val_compare_and_swap(
654 volatile kmp_uint64 *p, kmp_uint64 cv, kmp_uint64 sv) {
655 __atomic_compare_exchange(p, &cv, &sv, false, __ATOMIC_SEQ_CST,
656 __ATOMIC_SEQ_CST);
657 return cv;
658}
659#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
660 mips_sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv),\
661 (kmp_uint64)(sv))
662#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
663 mips_sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv),\
664 (kmp_uint64)(sv))
665#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
666 mips_sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
667 (kmp_uint64)(sv))
668#else
669#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
670 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
671 (kmp_uint64)(sv))
672#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
673 __sync_bool_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
674 (kmp_uint64)(sv))
675#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
676 __sync_val_compare_and_swap((volatile kmp_uint64 *)(p), (kmp_uint64)(cv), \
677 (kmp_uint64)(sv))
678#endif
679
680#define KMP_XCHG_FIXED8(p, v) \
681 __sync_lock_test_and_set((volatile kmp_uint8 *)(p), (kmp_uint8)(v))
682#define KMP_XCHG_FIXED16(p, v) \
683 __sync_lock_test_and_set((volatile kmp_uint16 *)(p), (kmp_uint16)(v))
684#define KMP_XCHG_FIXED32(p, v) \
685 __sync_lock_test_and_set((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
686#define KMP_XCHG_FIXED64(p, v) \
687 __sync_lock_test_and_set((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
688
689inline kmp_real32 KMP_XCHG_REAL32(volatile kmp_real32 *p, kmp_real32 v) {
690 kmp_int32 tmp =
691 __sync_lock_test_and_set((volatile kmp_uint32 *)(p), *(kmp_uint32 *)&v);
692 return *(kmp_real32 *)&tmp;
693}
694
695inline kmp_real64 KMP_XCHG_REAL64(volatile kmp_real64 *p, kmp_real64 v) {
696 kmp_int64 tmp =
697 __sync_lock_test_and_set((volatile kmp_uint64 *)(p), *(kmp_uint64 *)&v);
698 return *(kmp_real64 *)&tmp;
699}
700
701#else
702
703extern kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 v);
704extern kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 v);
705extern kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 v);
706extern kmp_int32 __kmp_test_then_add32(volatile kmp_int32 *p, kmp_int32 v);
707extern kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 v);
708extern kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 v);
709extern kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 v);
710extern kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 v);
711extern kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 v);
712
713extern kmp_int8 __kmp_compare_and_store8(volatile kmp_int8 *p, kmp_int8 cv,
714 kmp_int8 sv);
715extern kmp_int16 __kmp_compare_and_store16(volatile kmp_int16 *p, kmp_int16 cv,
716 kmp_int16 sv);
717extern kmp_int32 __kmp_compare_and_store32(volatile kmp_int32 *p, kmp_int32 cv,
718 kmp_int32 sv);
719extern kmp_int32 __kmp_compare_and_store64(volatile kmp_int64 *p, kmp_int64 cv,
720 kmp_int64 sv);
721extern kmp_int8 __kmp_compare_and_store_ret8(volatile kmp_int8 *p, kmp_int8 cv,
722 kmp_int8 sv);
723extern kmp_int16 __kmp_compare_and_store_ret16(volatile kmp_int16 *p,
724 kmp_int16 cv, kmp_int16 sv);
725extern kmp_int32 __kmp_compare_and_store_ret32(volatile kmp_int32 *p,
726 kmp_int32 cv, kmp_int32 sv);
727extern kmp_int64 __kmp_compare_and_store_ret64(volatile kmp_int64 *p,
728 kmp_int64 cv, kmp_int64 sv);
729
730extern kmp_int8 __kmp_xchg_fixed8(volatile kmp_int8 *p, kmp_int8 v);
731extern kmp_int16 __kmp_xchg_fixed16(volatile kmp_int16 *p, kmp_int16 v);
732extern kmp_int32 __kmp_xchg_fixed32(volatile kmp_int32 *p, kmp_int32 v);
733extern kmp_int64 __kmp_xchg_fixed64(volatile kmp_int64 *p, kmp_int64 v);
734extern kmp_real32 __kmp_xchg_real32(volatile kmp_real32 *p, kmp_real32 v);
735extern kmp_real64 __kmp_xchg_real64(volatile kmp_real64 *p, kmp_real64 v);
736
737#define KMP_TEST_THEN_INC32(p) \
738 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
739#define KMP_TEST_THEN_INC_ACQ32(p) \
740 __kmp_test_then_add32((volatile kmp_int32 *)(p), 1)
741#define KMP_TEST_THEN_INC64(p) \
742 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
743#define KMP_TEST_THEN_INC_ACQ64(p) \
744 __kmp_test_then_add64((volatile kmp_int64 *)(p), 1LL)
745#define KMP_TEST_THEN_ADD4_32(p) \
746 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
747#define KMP_TEST_THEN_ADD4_ACQ32(p) \
748 __kmp_test_then_add32((volatile kmp_int32 *)(p), 4)
749#define KMP_TEST_THEN_ADD4_64(p) \
750 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
751#define KMP_TEST_THEN_ADD4_ACQ64(p) \
752 __kmp_test_then_add64((volatile kmp_int64 *)(p), 4LL)
753#define KMP_TEST_THEN_DEC32(p) \
754 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
755#define KMP_TEST_THEN_DEC_ACQ32(p) \
756 __kmp_test_then_add32((volatile kmp_int32 *)(p), -1)
757#define KMP_TEST_THEN_DEC64(p) \
758 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
759#define KMP_TEST_THEN_DEC_ACQ64(p) \
760 __kmp_test_then_add64((volatile kmp_int64 *)(p), -1LL)
761#define KMP_TEST_THEN_ADD8(p, v) \
762 __kmp_test_then_add8((volatile kmp_int8 *)(p), (kmp_int8)(v))
763#define KMP_TEST_THEN_ADD32(p, v) \
764 __kmp_test_then_add32((volatile kmp_int32 *)(p), (kmp_int32)(v))
765#define KMP_TEST_THEN_ADD64(p, v) \
766 __kmp_test_then_add64((volatile kmp_int64 *)(p), (kmp_int64)(v))
767
768#define KMP_TEST_THEN_OR8(p, v) \
769 __kmp_test_then_or8((volatile kmp_int8 *)(p), (kmp_int8)(v))
770#define KMP_TEST_THEN_AND8(p, v) \
771 __kmp_test_then_and8((volatile kmp_int8 *)(p), (kmp_int8)(v))
772#define KMP_TEST_THEN_OR32(p, v) \
773 __kmp_test_then_or32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
774#define KMP_TEST_THEN_AND32(p, v) \
775 __kmp_test_then_and32((volatile kmp_uint32 *)(p), (kmp_uint32)(v))
776#define KMP_TEST_THEN_OR64(p, v) \
777 __kmp_test_then_or64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
778#define KMP_TEST_THEN_AND64(p, v) \
779 __kmp_test_then_and64((volatile kmp_uint64 *)(p), (kmp_uint64)(v))
780
781#define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv) \
782 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
783 (kmp_int8)(sv))
784#define KMP_COMPARE_AND_STORE_REL8(p, cv, sv) \
785 __kmp_compare_and_store8((volatile kmp_int8 *)(p), (kmp_int8)(cv), \
786 (kmp_int8)(sv))
787#define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) \
788 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
789 (kmp_int16)(sv))
790#define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) \
791 __kmp_compare_and_store16((volatile kmp_int16 *)(p), (kmp_int16)(cv), \
792 (kmp_int16)(sv))
793#define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) \
794 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
795 (kmp_int32)(sv))
796#define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) \
797 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
798 (kmp_int32)(sv))
799#define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) \
800 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
801 (kmp_int64)(sv))
802#define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) \
803 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
804 (kmp_int64)(sv))
805
806#if KMP_ARCH_X86
807#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
808 __kmp_compare_and_store32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
809 (kmp_int32)(sv))
810#else /* 64 bit pointers */
811#define KMP_COMPARE_AND_STORE_PTR(p, cv, sv) \
812 __kmp_compare_and_store64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
813 (kmp_int64)(sv))
814#endif /* KMP_ARCH_X86 */
815
816#define KMP_COMPARE_AND_STORE_RET8(p, cv, sv) \
817 __kmp_compare_and_store_ret8((p), (cv), (sv))
818#define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) \
819 __kmp_compare_and_store_ret16((p), (cv), (sv))
820#define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) \
821 __kmp_compare_and_store_ret32((volatile kmp_int32 *)(p), (kmp_int32)(cv), \
822 (kmp_int32)(sv))
823#define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) \
824 __kmp_compare_and_store_ret64((volatile kmp_int64 *)(p), (kmp_int64)(cv), \
825 (kmp_int64)(sv))
826
827#define KMP_XCHG_FIXED8(p, v) \
828 __kmp_xchg_fixed8((volatile kmp_int8 *)(p), (kmp_int8)(v));
829#define KMP_XCHG_FIXED16(p, v) __kmp_xchg_fixed16((p), (v));
830#define KMP_XCHG_FIXED32(p, v) __kmp_xchg_fixed32((p), (v));
831#define KMP_XCHG_FIXED64(p, v) __kmp_xchg_fixed64((p), (v));
832#define KMP_XCHG_REAL32(p, v) __kmp_xchg_real32((p), (v));
833#define KMP_XCHG_REAL64(p, v) __kmp_xchg_real64((p), (v));
834
835#endif /* KMP_ASM_INTRINS */
836
837/* ------------- relaxed consistency memory model stuff ------------------ */
838
839#if KMP_OS_WINDOWS
840#ifdef __ABSOFT_WIN
841#define KMP_MB() asm("nop")
842#define KMP_IMB() asm("nop")
843#else
844#define KMP_MB() /* _asm{ nop } */
845#define KMP_IMB() /* _asm{ nop } */
846#endif
847#endif /* KMP_OS_WINDOWS */
848
849#if KMP_ARCH_PPC64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS || \
850 KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64
851#define KMP_MB() __sync_synchronize()
852#endif
853
854#ifndef KMP_MB
855#define KMP_MB() /* nothing to do */
856#endif
857
858#ifndef KMP_IMB
859#define KMP_IMB() /* nothing to do */
860#endif
861
862#ifndef KMP_ST_REL32
863#define KMP_ST_REL32(A, D) (*(A) = (D))
864#endif
865
866#ifndef KMP_ST_REL64
867#define KMP_ST_REL64(A, D) (*(A) = (D))
868#endif
869
870#ifndef KMP_LD_ACQ32
871#define KMP_LD_ACQ32(A) (*(A))
872#endif
873
874#ifndef KMP_LD_ACQ64
875#define KMP_LD_ACQ64(A) (*(A))
876#endif
877
878/* ------------------------------------------------------------------------ */
879// FIXME - maybe this should this be
880//
881// #define TCR_4(a) (*(volatile kmp_int32 *)(&a))
882// #define TCW_4(a,b) (a) = (*(volatile kmp_int32 *)&(b))
883//
884// #define TCR_8(a) (*(volatile kmp_int64 *)(a))
885// #define TCW_8(a,b) (a) = (*(volatile kmp_int64 *)(&b))
886//
887// I'm fairly certain this is the correct thing to do, but I'm afraid
888// of performance regressions.
889
890#define TCR_1(a) (a)
891#define TCW_1(a, b) (a) = (b)
892#define TCR_4(a) (a)
893#define TCW_4(a, b) (a) = (b)
894#define TCI_4(a) (++(a))
895#define TCD_4(a) (--(a))
896#define TCR_8(a) (a)
897#define TCW_8(a, b) (a) = (b)
898#define TCI_8(a) (++(a))
899#define TCD_8(a) (--(a))
900#define TCR_SYNC_4(a) (a)
901#define TCW_SYNC_4(a, b) (a) = (b)
902#define TCX_SYNC_4(a, b, c) \
903 KMP_COMPARE_AND_STORE_REL32((volatile kmp_int32 *)(volatile void *)&(a), \
904 (kmp_int32)(b), (kmp_int32)(c))
905#define TCR_SYNC_8(a) (a)
906#define TCW_SYNC_8(a, b) (a) = (b)
907#define TCX_SYNC_8(a, b, c) \
908 KMP_COMPARE_AND_STORE_REL64((volatile kmp_int64 *)(volatile void *)&(a), \
909 (kmp_int64)(b), (kmp_int64)(c))
910
911#if KMP_ARCH_X86 || KMP_ARCH_MIPS
912// What about ARM?
913#define TCR_PTR(a) ((void *)TCR_4(a))
914#define TCW_PTR(a, b) TCW_4((a), (b))
915#define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_4(a))
916#define TCW_SYNC_PTR(a, b) TCW_SYNC_4((a), (b))
917#define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_4((a), (b), (c)))
918
919#else /* 64 bit pointers */
920
921#define TCR_PTR(a) ((void *)TCR_8(a))
922#define TCW_PTR(a, b) TCW_8((a), (b))
923#define TCR_SYNC_PTR(a) ((void *)TCR_SYNC_8(a))
924#define TCW_SYNC_PTR(a, b) TCW_SYNC_8((a), (b))
925#define TCX_SYNC_PTR(a, b, c) ((void *)TCX_SYNC_8((a), (b), (c)))
926
927#endif /* KMP_ARCH_X86 */
928
929/* If these FTN_{TRUE,FALSE} values change, may need to change several places
930 where they are used to check that language is Fortran, not C. */
931
932#ifndef FTN_TRUE
933#define FTN_TRUE TRUE
934#endif
935
936#ifndef FTN_FALSE
937#define FTN_FALSE FALSE
938#endif
939
940typedef void (*microtask_t)(int *gtid, int *npr, ...);
941
942#ifdef USE_VOLATILE_CAST
943#define VOLATILE_CAST(x) (volatile x)
944#else
945#define VOLATILE_CAST(x) (x)
946#endif
947
948#define KMP_WAIT __kmp_wait_4
949#define KMP_WAIT_PTR __kmp_wait_4_ptr
950#define KMP_EQ __kmp_eq_4
951#define KMP_NEQ __kmp_neq_4
952#define KMP_LT __kmp_lt_4
953#define KMP_GE __kmp_ge_4
954#define KMP_LE __kmp_le_4
955
956/* Workaround for Intel(R) 64 code gen bug when taking address of static array
957 * (Intel(R) 64 Tracker #138) */
958#if (KMP_ARCH_X86_64 || KMP_ARCH_PPC64) && KMP_OS_LINUX
959#define STATIC_EFI2_WORKAROUND
960#else
961#define STATIC_EFI2_WORKAROUND static
962#endif
963
964// Support of BGET usage
965#ifndef KMP_USE_BGET
966#define KMP_USE_BGET 1
967#endif
968
969// Switches for OSS builds
970#ifndef USE_CMPXCHG_FIX
971#define USE_CMPXCHG_FIX 1
972#endif
973
974// Enable dynamic user lock
975#define KMP_USE_DYNAMIC_LOCK 1
976
977// Enable Intel(R) Transactional Synchronization Extensions (Intel(R) TSX) if
978// dynamic user lock is turned on
979#if KMP_USE_DYNAMIC_LOCK
980// Visual studio can't handle the asm sections in this code
981#define KMP_USE_TSX (KMP_ARCH_X86 || KMP_ARCH_X86_64) && !KMP_COMPILER_MSVC
982#ifdef KMP_USE_ADAPTIVE_LOCKS
983#undef KMP_USE_ADAPTIVE_LOCKS
984#endif
985#define KMP_USE_ADAPTIVE_LOCKS KMP_USE_TSX
986#endif
987
988// Enable tick time conversion of ticks to seconds
989#if KMP_STATS_ENABLED
990#define KMP_HAVE_TICK_TIME \
991 (KMP_OS_LINUX && (KMP_MIC || KMP_ARCH_X86 || KMP_ARCH_X86_64))
992#endif
993
994// Warning levels
995enum kmp_warnings_level {
996 kmp_warnings_off = 0, /* No warnings */
997 kmp_warnings_low, /* Minimal warnings (default) */
998 kmp_warnings_explicit = 6, /* Explicitly set to ON - more warnings */
999 kmp_warnings_verbose /* reserved */
1000};
1001
1002#ifdef __cplusplus
1003} // extern "C"
1004#endif // __cplusplus
1005
1006// Macros for C++11 atomic functions
1007#define KMP_ATOMIC_LD(p, order) (p)->load(std::memory_order_##order)
1008#define KMP_ATOMIC_OP(op, p, v, order) (p)->op(v, std::memory_order_##order)
1009
1010// For non-default load/store
1011#define KMP_ATOMIC_LD_ACQ(p) KMP_ATOMIC_LD(p, acquire)
1012#define KMP_ATOMIC_LD_RLX(p) KMP_ATOMIC_LD(p, relaxed)
1013#define KMP_ATOMIC_ST_REL(p, v) KMP_ATOMIC_OP(store, p, v, release)
1014#define KMP_ATOMIC_ST_RLX(p, v) KMP_ATOMIC_OP(store, p, v, relaxed)
1015
1016// For non-default fetch_<op>
1017#define KMP_ATOMIC_ADD(p, v) KMP_ATOMIC_OP(fetch_add, p, v, acq_rel)
1018#define KMP_ATOMIC_SUB(p, v) KMP_ATOMIC_OP(fetch_sub, p, v, acq_rel)
1019#define KMP_ATOMIC_AND(p, v) KMP_ATOMIC_OP(fetch_and, p, v, acq_rel)
1020#define KMP_ATOMIC_OR(p, v) KMP_ATOMIC_OP(fetch_or, p, v, acq_rel)
1021#define KMP_ATOMIC_INC(p) KMP_ATOMIC_OP(fetch_add, p, 1, acq_rel)
1022#define KMP_ATOMIC_DEC(p) KMP_ATOMIC_OP(fetch_sub, p, 1, acq_rel)
1023#define KMP_ATOMIC_ADD_RLX(p, v) KMP_ATOMIC_OP(fetch_add, p, v, relaxed)
1024#define KMP_ATOMIC_INC_RLX(p) KMP_ATOMIC_OP(fetch_add, p, 1, relaxed)
1025
1026// Callers of the following functions cannot see the side effect on "expected".
1027template <typename T>
1028bool __kmp_atomic_compare_store(std::atomic<T> *p, T expected, T desired) {
1029 return p->compare_exchange_strong(
1030 expected, desired, std::memory_order_acq_rel, std::memory_order_relaxed);
1031}
1032
1033template <typename T>
1034bool __kmp_atomic_compare_store_acq(std::atomic<T> *p, T expected, T desired) {
1035 return p->compare_exchange_strong(
1036 expected, desired, std::memory_order_acquire, std::memory_order_relaxed);
1037}
1038
1039template <typename T>
1040bool __kmp_atomic_compare_store_rel(std::atomic<T> *p, T expected, T desired) {
1041 return p->compare_exchange_strong(
1042 expected, desired, std::memory_order_release, std::memory_order_relaxed);
1043}
1044
1045#endif /* KMP_OS_H */
1046// Safe C API
1047#include "kmp_safe_c_api.h"