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