LLVM OpenMP* Runtime Library
kmp.h
1 
2 /*
3  * kmp.h -- KPTS runtime header file.
4  */
5 
6 //===----------------------------------------------------------------------===//
7 //
8 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9 // See https://llvm.org/LICENSE.txt for license information.
10 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef KMP_H
15 #define KMP_H
16 
17 #include "kmp_config.h"
18 
19 /* #define BUILD_PARALLEL_ORDERED 1 */
20 
21 /* This fix replaces gettimeofday with clock_gettime for better scalability on
22  the Altix. Requires user code to be linked with -lrt. */
23 //#define FIX_SGI_CLOCK
24 
25 /* Defines for OpenMP 3.0 tasking and auto scheduling */
26 
27 #ifndef KMP_STATIC_STEAL_ENABLED
28 #define KMP_STATIC_STEAL_ENABLED 1
29 #endif
30 
31 #define TASK_CURRENT_NOT_QUEUED 0
32 #define TASK_CURRENT_QUEUED 1
33 
34 #ifdef BUILD_TIED_TASK_STACK
35 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37 // Number of entries in each task stack array
38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39 // Mask for determining index into stack block
40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41 #endif // BUILD_TIED_TASK_STACK
42 
43 #define TASK_NOT_PUSHED 1
44 #define TASK_SUCCESSFULLY_PUSHED 0
45 #define TASK_TIED 1
46 #define TASK_UNTIED 0
47 #define TASK_EXPLICIT 1
48 #define TASK_IMPLICIT 0
49 #define TASK_PROXY 1
50 #define TASK_FULL 0
51 #define TASK_DETACHABLE 1
52 #define TASK_UNDETACHABLE 0
53 
54 #define KMP_CANCEL_THREADS
55 #define KMP_THREAD_ATTR
56 
57 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
58 // built on Android
59 #if defined(__ANDROID__)
60 #undef KMP_CANCEL_THREADS
61 #endif
62 
63 #include <signal.h>
64 #include <stdarg.h>
65 #include <stddef.h>
66 #include <stdio.h>
67 #include <stdlib.h>
68 #include <string.h>
69 #include <limits>
70 #include <type_traits>
71 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72  Microsoft library. Some macros provided below to replace these functions */
73 #ifndef __ABSOFT_WIN
74 #include <sys/types.h>
75 #endif
76 #include <limits.h>
77 #include <time.h>
78 
79 #include <errno.h>
80 
81 #include "kmp_os.h"
82 
83 #include "kmp_safe_c_api.h"
84 
85 #if KMP_STATS_ENABLED
86 class kmp_stats_list;
87 #endif
88 
89 #if KMP_USE_HIER_SCHED
90 // Only include hierarchical scheduling if affinity is supported
91 #undef KMP_USE_HIER_SCHED
92 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93 #endif
94 
95 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96 #include "hwloc.h"
97 #ifndef HWLOC_OBJ_NUMANODE
98 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99 #endif
100 #ifndef HWLOC_OBJ_PACKAGE
101 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102 #endif
103 #endif
104 
105 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
106 #include <xmmintrin.h>
107 #endif
108 
109 // The below has to be defined before including "kmp_barrier.h".
110 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
111 #define KMP_INTERNAL_FREE(p) free(p)
112 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
113 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
114 
115 #include "kmp_debug.h"
116 #include "kmp_lock.h"
117 #include "kmp_version.h"
118 #include "kmp_barrier.h"
119 #if USE_DEBUGGER
120 #include "kmp_debugger.h"
121 #endif
122 #include "kmp_i18n.h"
123 
124 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
125 
126 #include "kmp_wrapper_malloc.h"
127 #if KMP_OS_UNIX
128 #include <unistd.h>
129 #if !defined NSIG && defined _NSIG
130 #define NSIG _NSIG
131 #endif
132 #endif
133 
134 #if KMP_OS_LINUX
135 #pragma weak clock_gettime
136 #endif
137 
138 #if OMPT_SUPPORT
139 #include "ompt-internal.h"
140 #endif
141 
142 #if OMPD_SUPPORT
143 #include "ompd-specific.h"
144 #endif
145 
146 #ifndef UNLIKELY
147 #define UNLIKELY(x) (x)
148 #endif
149 
150 // Affinity format function
151 #include "kmp_str.h"
152 
153 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
154 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
155 // free lists of limited size.
156 #ifndef USE_FAST_MEMORY
157 #define USE_FAST_MEMORY 3
158 #endif
159 
160 #ifndef KMP_NESTED_HOT_TEAMS
161 #define KMP_NESTED_HOT_TEAMS 0
162 #define USE_NESTED_HOT_ARG(x)
163 #else
164 #if KMP_NESTED_HOT_TEAMS
165 #define USE_NESTED_HOT_ARG(x) , x
166 #else
167 #define USE_NESTED_HOT_ARG(x)
168 #endif
169 #endif
170 
171 // Assume using BGET compare_exchange instruction instead of lock by default.
172 #ifndef USE_CMP_XCHG_FOR_BGET
173 #define USE_CMP_XCHG_FOR_BGET 1
174 #endif
175 
176 // Test to see if queuing lock is better than bootstrap lock for bget
177 // #ifndef USE_QUEUING_LOCK_FOR_BGET
178 // #define USE_QUEUING_LOCK_FOR_BGET
179 // #endif
180 
181 #define KMP_NSEC_PER_SEC 1000000000L
182 #define KMP_USEC_PER_SEC 1000000L
183 
192 enum {
197  /* 0x04 is no longer used */
206  KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
207  KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
208  KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
209 
210  KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
211  KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
212 
224  KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
225  KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
226  KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
227  KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
228  KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
229 };
230 
234 typedef struct ident {
235  kmp_int32 reserved_1;
236  kmp_int32 flags;
238  kmp_int32 reserved_2;
239 #if USE_ITT_BUILD
240 /* but currently used for storing region-specific ITT */
241 /* contextual information. */
242 #endif /* USE_ITT_BUILD */
243  kmp_int32 reserved_3;
244  char const *psource;
248  // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
249  kmp_int32 get_openmp_version() {
250  return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
251  }
252 } ident_t;
257 // Some forward declarations.
258 typedef union kmp_team kmp_team_t;
259 typedef struct kmp_taskdata kmp_taskdata_t;
260 typedef union kmp_task_team kmp_task_team_t;
261 typedef union kmp_team kmp_team_p;
262 typedef union kmp_info kmp_info_p;
263 typedef union kmp_root kmp_root_p;
264 
265 template <bool C = false, bool S = true> class kmp_flag_32;
266 template <bool C = false, bool S = true> class kmp_flag_64;
267 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
268 class kmp_flag_oncore;
269 
270 #ifdef __cplusplus
271 extern "C" {
272 #endif
273 
274 /* ------------------------------------------------------------------------ */
275 
276 /* Pack two 32-bit signed integers into a 64-bit signed integer */
277 /* ToDo: Fix word ordering for big-endian machines. */
278 #define KMP_PACK_64(HIGH_32, LOW_32) \
279  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
280 
281 // Generic string manipulation macros. Assume that _x is of type char *
282 #define SKIP_WS(_x) \
283  { \
284  while (*(_x) == ' ' || *(_x) == '\t') \
285  (_x)++; \
286  }
287 #define SKIP_DIGITS(_x) \
288  { \
289  while (*(_x) >= '0' && *(_x) <= '9') \
290  (_x)++; \
291  }
292 #define SKIP_TOKEN(_x) \
293  { \
294  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
295  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
296  (_x)++; \
297  }
298 #define SKIP_TO(_x, _c) \
299  { \
300  while (*(_x) != '\0' && *(_x) != (_c)) \
301  (_x)++; \
302  }
303 
304 /* ------------------------------------------------------------------------ */
305 
306 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
307 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
308 
309 /* ------------------------------------------------------------------------ */
310 /* Enumeration types */
311 
312 enum kmp_state_timer {
313  ts_stop,
314  ts_start,
315  ts_pause,
316 
317  ts_last_state
318 };
319 
320 enum dynamic_mode {
321  dynamic_default,
322 #ifdef USE_LOAD_BALANCE
323  dynamic_load_balance,
324 #endif /* USE_LOAD_BALANCE */
325  dynamic_random,
326  dynamic_thread_limit,
327  dynamic_max
328 };
329 
330 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
331  * not include it here */
332 #ifndef KMP_SCHED_TYPE_DEFINED
333 #define KMP_SCHED_TYPE_DEFINED
334 typedef enum kmp_sched {
335  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
336  // Note: need to adjust __kmp_sch_map global array in case enum is changed
337  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
338  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
339  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
340  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
341  kmp_sched_upper_std = 5, // upper bound for standard schedules
342  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
343  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
344 #if KMP_STATIC_STEAL_ENABLED
345  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
346 #endif
347  kmp_sched_upper,
348  kmp_sched_default = kmp_sched_static, // default scheduling
349  kmp_sched_monotonic = 0x80000000
350 } kmp_sched_t;
351 #endif
352 
357 enum sched_type : kmp_int32 {
359  kmp_sch_static_chunked = 33,
361  kmp_sch_dynamic_chunked = 35,
363  kmp_sch_runtime = 37,
365  kmp_sch_trapezoidal = 39,
366 
367  /* accessible only through KMP_SCHEDULE environment variable */
368  kmp_sch_static_greedy = 40,
369  kmp_sch_static_balanced = 41,
370  /* accessible only through KMP_SCHEDULE environment variable */
371  kmp_sch_guided_iterative_chunked = 42,
372  kmp_sch_guided_analytical_chunked = 43,
373  /* accessible only through KMP_SCHEDULE environment variable */
374  kmp_sch_static_steal = 44,
375 
376  /* static with chunk adjustment (e.g., simd) */
377  kmp_sch_static_balanced_chunked = 45,
381  /* accessible only through KMP_SCHEDULE environment variable */
385  kmp_ord_static_chunked = 65,
387  kmp_ord_dynamic_chunked = 67,
388  kmp_ord_guided_chunked = 68,
389  kmp_ord_runtime = 69,
391  kmp_ord_trapezoidal = 71,
394  /* Schedules for Distribute construct */
398  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
399  single iteration/chunk, even if the loop is serialized. For the schedule
400  types listed above, the entire iteration vector is returned if the loop is
401  serialized. This doesn't work for gcc/gcomp sections. */
402  kmp_nm_lower = 160,
404  kmp_nm_static_chunked =
405  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
407  kmp_nm_dynamic_chunked = 163,
409  kmp_nm_runtime = 165,
410  kmp_nm_auto = 166,
411  kmp_nm_trapezoidal = 167,
412 
413  /* accessible only through KMP_SCHEDULE environment variable */
414  kmp_nm_static_greedy = 168,
415  kmp_nm_static_balanced = 169,
416  /* accessible only through KMP_SCHEDULE environment variable */
417  kmp_nm_guided_iterative_chunked = 170,
418  kmp_nm_guided_analytical_chunked = 171,
419  kmp_nm_static_steal =
420  172, /* accessible only through OMP_SCHEDULE environment variable */
421 
422  kmp_nm_ord_static_chunked = 193,
424  kmp_nm_ord_dynamic_chunked = 195,
425  kmp_nm_ord_guided_chunked = 196,
426  kmp_nm_ord_runtime = 197,
428  kmp_nm_ord_trapezoidal = 199,
431  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
432  we need to distinguish the three possible cases (no modifier, monotonic
433  modifier, nonmonotonic modifier), we need separate bits for each modifier.
434  The absence of monotonic does not imply nonmonotonic, especially since 4.5
435  says that the behaviour of the "no modifier" case is implementation defined
436  in 4.5, but will become "nonmonotonic" in 5.0.
437 
438  Since we're passing a full 32 bit value, we can use a couple of high bits
439  for these flags; out of paranoia we avoid the sign bit.
440 
441  These modifiers can be or-ed into non-static schedules by the compiler to
442  pass the additional information. They will be stripped early in the
443  processing in __kmp_dispatch_init when setting up schedules, so most of the
444  code won't ever see schedules with these bits set. */
446  (1 << 29),
448  (1 << 30),
450 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
451  (enum sched_type)( \
453 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
454 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
455 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
456  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
457 #define SCHEDULE_GET_MODIFIERS(s) \
458  ((enum sched_type)( \
459  (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
460 #define SCHEDULE_SET_MODIFIERS(s, m) \
461  (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
462 #define SCHEDULE_NONMONOTONIC 0
463 #define SCHEDULE_MONOTONIC 1
464 
466 };
467 
468 // Apply modifiers on internal kind to standard kind
469 static inline void
470 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
471  enum sched_type internal_kind) {
472  if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
473  *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
474  }
475 }
476 
477 // Apply modifiers on standard kind to internal kind
478 static inline void
479 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
480  enum sched_type *internal_kind) {
481  if ((int)kind & (int)kmp_sched_monotonic) {
482  *internal_kind = (enum sched_type)((int)*internal_kind |
484  }
485 }
486 
487 // Get standard schedule without modifiers
488 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
489  return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
490 }
491 
492 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
493 typedef union kmp_r_sched {
494  struct {
495  enum sched_type r_sched_type;
496  int chunk;
497  };
498  kmp_int64 sched;
499 } kmp_r_sched_t;
500 
501 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
502 // internal schedule types
503 
504 enum library_type {
505  library_none,
506  library_serial,
507  library_turnaround,
508  library_throughput
509 };
510 
511 #if KMP_OS_LINUX
512 enum clock_function_type {
513  clock_function_gettimeofday,
514  clock_function_clock_gettime
515 };
516 #endif /* KMP_OS_LINUX */
517 
518 #if KMP_MIC_SUPPORTED
519 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
520 #endif
521 
522 /* -- fast reduction stuff ------------------------------------------------ */
523 
524 #undef KMP_FAST_REDUCTION_BARRIER
525 #define KMP_FAST_REDUCTION_BARRIER 1
526 
527 #undef KMP_FAST_REDUCTION_CORE_DUO
528 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
529 #define KMP_FAST_REDUCTION_CORE_DUO 1
530 #endif
531 
532 enum _reduction_method {
533  reduction_method_not_defined = 0,
534  critical_reduce_block = (1 << 8),
535  atomic_reduce_block = (2 << 8),
536  tree_reduce_block = (3 << 8),
537  empty_reduce_block = (4 << 8)
538 };
539 
540 // Description of the packed_reduction_method variable:
541 // The packed_reduction_method variable consists of two enum types variables
542 // that are packed together into 0-th byte and 1-st byte:
543 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
544 // barrier that will be used in fast reduction: bs_plain_barrier or
545 // bs_reduction_barrier
546 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
547 // be used in fast reduction;
548 // Reduction method is of 'enum _reduction_method' type and it's defined the way
549 // so that the bits of 0-th byte are empty, so no need to execute a shift
550 // instruction while packing/unpacking
551 
552 #if KMP_FAST_REDUCTION_BARRIER
553 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
554  ((reduction_method) | (barrier_type))
555 
556 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
557  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
558 
559 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
560  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
561 #else
562 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
563  (reduction_method)
564 
565 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
566  (packed_reduction_method)
567 
568 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
569 #endif
570 
571 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
572  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
573  (which_reduction_block))
574 
575 #if KMP_FAST_REDUCTION_BARRIER
576 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
577  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
578 
579 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
580  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
581 #endif
582 
583 typedef int PACKED_REDUCTION_METHOD_T;
584 
585 /* -- end of fast reduction stuff ----------------------------------------- */
586 
587 #if KMP_OS_WINDOWS
588 #define USE_CBLKDATA
589 #if KMP_MSVC_COMPAT
590 #pragma warning(push)
591 #pragma warning(disable : 271 310)
592 #endif
593 #include <windows.h>
594 #if KMP_MSVC_COMPAT
595 #pragma warning(pop)
596 #endif
597 #endif
598 
599 #if KMP_OS_UNIX
600 #include <dlfcn.h>
601 #include <pthread.h>
602 #endif
603 
604 enum kmp_hw_t : int {
605  KMP_HW_UNKNOWN = -1,
606  KMP_HW_SOCKET = 0,
607  KMP_HW_PROC_GROUP,
608  KMP_HW_NUMA,
609  KMP_HW_DIE,
610  KMP_HW_LLC,
611  KMP_HW_L3,
612  KMP_HW_TILE,
613  KMP_HW_MODULE,
614  KMP_HW_L2,
615  KMP_HW_L1,
616  KMP_HW_CORE,
617  KMP_HW_THREAD,
618  KMP_HW_LAST
619 };
620 
621 typedef enum kmp_hw_core_type_t {
622  KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
623 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
624  KMP_HW_CORE_TYPE_ATOM = 0x20,
625  KMP_HW_CORE_TYPE_CORE = 0x40,
626  KMP_HW_MAX_NUM_CORE_TYPES = 3,
627 #else
628  KMP_HW_MAX_NUM_CORE_TYPES = 1,
629 #endif
630 } kmp_hw_core_type_t;
631 
632 #define KMP_HW_MAX_NUM_CORE_EFFS 8
633 
634 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
635  KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
636 #define KMP_ASSERT_VALID_HW_TYPE(type) \
637  KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
638 
639 #define KMP_FOREACH_HW_TYPE(type) \
640  for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
641  type = (kmp_hw_t)((int)type + 1))
642 
643 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
644 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
645 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
646 
647 /* Only Linux* OS and Windows* OS support thread affinity. */
648 #if KMP_AFFINITY_SUPPORTED
649 
650 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
651 #if KMP_OS_WINDOWS
652 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
653 typedef struct GROUP_AFFINITY {
654  KAFFINITY Mask;
655  WORD Group;
656  WORD Reserved[3];
657 } GROUP_AFFINITY;
658 #endif /* _MSC_VER < 1600 */
659 #if KMP_GROUP_AFFINITY
660 extern int __kmp_num_proc_groups;
661 #else
662 static const int __kmp_num_proc_groups = 1;
663 #endif /* KMP_GROUP_AFFINITY */
664 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
665 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
666 
667 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
668 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
669 
670 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
671 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
672 
673 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
674  GROUP_AFFINITY *);
675 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
676 #endif /* KMP_OS_WINDOWS */
677 
678 #if KMP_USE_HWLOC
679 extern hwloc_topology_t __kmp_hwloc_topology;
680 extern int __kmp_hwloc_error;
681 #endif
682 
683 extern size_t __kmp_affin_mask_size;
684 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
685 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
686 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
687 #define KMP_CPU_SET_ITERATE(i, mask) \
688  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
689 #define KMP_CPU_SET(i, mask) (mask)->set(i)
690 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
691 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
692 #define KMP_CPU_ZERO(mask) (mask)->zero()
693 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
694 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
695 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
696 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
697 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
698 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
699 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
700 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
701 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
702 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
703 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
704 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
705  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
706 #define KMP_CPU_FREE_ARRAY(arr, n) \
707  __kmp_affinity_dispatch->deallocate_mask_array(arr)
708 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
709 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
710 #define __kmp_get_system_affinity(mask, abort_bool) \
711  (mask)->get_system_affinity(abort_bool)
712 #define __kmp_set_system_affinity(mask, abort_bool) \
713  (mask)->set_system_affinity(abort_bool)
714 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
715 
716 class KMPAffinity {
717 public:
718  class Mask {
719  public:
720  void *operator new(size_t n);
721  void operator delete(void *p);
722  void *operator new[](size_t n);
723  void operator delete[](void *p);
724  virtual ~Mask() {}
725  // Set bit i to 1
726  virtual void set(int i) {}
727  // Return bit i
728  virtual bool is_set(int i) const { return false; }
729  // Set bit i to 0
730  virtual void clear(int i) {}
731  // Zero out entire mask
732  virtual void zero() {}
733  // Copy src into this mask
734  virtual void copy(const Mask *src) {}
735  // this &= rhs
736  virtual void bitwise_and(const Mask *rhs) {}
737  // this |= rhs
738  virtual void bitwise_or(const Mask *rhs) {}
739  // this = ~this
740  virtual void bitwise_not() {}
741  // API for iterating over an affinity mask
742  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
743  virtual int begin() const { return 0; }
744  virtual int end() const { return 0; }
745  virtual int next(int previous) const { return 0; }
746 #if KMP_OS_WINDOWS
747  virtual int set_process_affinity(bool abort_on_error) const { return -1; }
748 #endif
749  // Set the system's affinity to this affinity mask's value
750  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
751  // Set this affinity mask to the current system affinity
752  virtual int get_system_affinity(bool abort_on_error) { return -1; }
753  // Only 1 DWORD in the mask should have any procs set.
754  // Return the appropriate index, or -1 for an invalid mask.
755  virtual int get_proc_group() const { return -1; }
756  int get_max_cpu() const {
757  int cpu;
758  int max_cpu = -1;
759  KMP_CPU_SET_ITERATE(cpu, this) {
760  if (cpu > max_cpu)
761  max_cpu = cpu;
762  }
763  return max_cpu;
764  }
765  };
766  void *operator new(size_t n);
767  void operator delete(void *p);
768  // Need virtual destructor
769  virtual ~KMPAffinity() = default;
770  // Determine if affinity is capable
771  virtual void determine_capable(const char *env_var) {}
772  // Bind the current thread to os proc
773  virtual void bind_thread(int proc) {}
774  // Factory functions to allocate/deallocate a mask
775  virtual Mask *allocate_mask() { return nullptr; }
776  virtual void deallocate_mask(Mask *m) {}
777  virtual Mask *allocate_mask_array(int num) { return nullptr; }
778  virtual void deallocate_mask_array(Mask *m) {}
779  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
780  static void pick_api();
781  static void destroy_api();
782  enum api_type {
783  NATIVE_OS
784 #if KMP_USE_HWLOC
785  ,
786  HWLOC
787 #endif
788  };
789  virtual api_type get_api_type() const {
790  KMP_ASSERT(0);
791  return NATIVE_OS;
792  }
793 
794 private:
795  static bool picked_api;
796 };
797 
798 typedef KMPAffinity::Mask kmp_affin_mask_t;
799 extern KMPAffinity *__kmp_affinity_dispatch;
800 
801 // Declare local char buffers with this size for printing debug and info
802 // messages, using __kmp_affinity_print_mask().
803 #define KMP_AFFIN_MASK_PRINT_LEN 1024
804 
805 enum affinity_type {
806  affinity_none = 0,
807  affinity_physical,
808  affinity_logical,
809  affinity_compact,
810  affinity_scatter,
811  affinity_explicit,
812  affinity_balanced,
813  affinity_disabled, // not used outsize the env var parser
814  affinity_default
815 };
816 
817 enum affinity_top_method {
818  affinity_top_method_all = 0, // try all (supported) methods, in order
819 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
820  affinity_top_method_apicid,
821  affinity_top_method_x2apicid,
822  affinity_top_method_x2apicid_1f,
823 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
824  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
825 #if KMP_GROUP_AFFINITY
826  affinity_top_method_group,
827 #endif /* KMP_GROUP_AFFINITY */
828  affinity_top_method_flat,
829 #if KMP_USE_HWLOC
830  affinity_top_method_hwloc,
831 #endif
832  affinity_top_method_default
833 };
834 
835 #define affinity_respect_mask_default (2)
836 
837 typedef struct kmp_affinity_flags_t {
838  unsigned dups : 1;
839  unsigned verbose : 1;
840  unsigned warnings : 1;
841  unsigned respect : 2;
842  unsigned reset : 1;
843  unsigned initialized : 1;
844  unsigned reserved : 25;
845 } kmp_affinity_flags_t;
846 KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
847 
848 typedef struct kmp_affinity_ids_t {
849  int ids[KMP_HW_LAST];
850  int operator[](size_t idx) const { return ids[idx]; }
851  int &operator[](size_t idx) { return ids[idx]; }
852  kmp_affinity_ids_t &operator=(const kmp_affinity_ids_t &rhs) {
853  for (int i = 0; i < KMP_HW_LAST; ++i)
854  ids[i] = rhs[i];
855  return *this;
856  }
857 } kmp_affinity_ids_t;
858 
859 typedef struct kmp_affinity_attrs_t {
860  int core_type : 8;
861  int core_eff : 8;
862  unsigned valid : 1;
863  unsigned reserved : 15;
864 } kmp_affinity_attrs_t;
865 #define KMP_AFFINITY_ATTRS_UNKNOWN \
866  { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
867 
868 typedef struct kmp_affinity_t {
869  char *proclist;
870  enum affinity_type type;
871  kmp_hw_t gran;
872  int gran_levels;
873  int compact;
874  int offset;
875  kmp_affinity_flags_t flags;
876  unsigned num_masks;
877  kmp_affin_mask_t *masks;
878  kmp_affinity_ids_t *ids;
879  kmp_affinity_attrs_t *attrs;
880  unsigned num_os_id_masks;
881  kmp_affin_mask_t *os_id_masks;
882  const char *env_var;
883 } kmp_affinity_t;
884 
885 #define KMP_AFFINITY_INIT(env) \
886  { \
887  nullptr, affinity_default, KMP_HW_UNKNOWN, -1, 0, 0, \
888  {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE}, 0, \
889  nullptr, nullptr, nullptr, 0, nullptr, env \
890  }
891 
892 extern enum affinity_top_method __kmp_affinity_top_method;
893 extern kmp_affinity_t __kmp_affinity;
894 extern kmp_affinity_t __kmp_hh_affinity;
895 extern kmp_affinity_t *__kmp_affinities[2];
896 
897 extern void __kmp_affinity_bind_thread(int which);
898 
899 extern kmp_affin_mask_t *__kmp_affin_fullMask;
900 extern kmp_affin_mask_t *__kmp_affin_origMask;
901 extern char *__kmp_cpuinfo_file;
902 
903 #endif /* KMP_AFFINITY_SUPPORTED */
904 
905 // This needs to be kept in sync with the values in omp.h !!!
906 typedef enum kmp_proc_bind_t {
907  proc_bind_false = 0,
908  proc_bind_true,
909  proc_bind_primary,
910  proc_bind_close,
911  proc_bind_spread,
912  proc_bind_intel, // use KMP_AFFINITY interface
913  proc_bind_default
914 } kmp_proc_bind_t;
915 
916 typedef struct kmp_nested_proc_bind_t {
917  kmp_proc_bind_t *bind_types;
918  int size;
919  int used;
920 } kmp_nested_proc_bind_t;
921 
922 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
923 extern kmp_proc_bind_t __kmp_teams_proc_bind;
924 
925 extern int __kmp_display_affinity;
926 extern char *__kmp_affinity_format;
927 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
928 #if OMPT_SUPPORT
929 extern int __kmp_tool;
930 extern char *__kmp_tool_libraries;
931 #endif // OMPT_SUPPORT
932 
933 #if KMP_AFFINITY_SUPPORTED
934 #define KMP_PLACE_ALL (-1)
935 #define KMP_PLACE_UNDEFINED (-2)
936 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
937 #define KMP_AFFINITY_NON_PROC_BIND \
938  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
939  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
940  (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
941 #endif /* KMP_AFFINITY_SUPPORTED */
942 
943 extern int __kmp_affinity_num_places;
944 
945 typedef enum kmp_cancel_kind_t {
946  cancel_noreq = 0,
947  cancel_parallel = 1,
948  cancel_loop = 2,
949  cancel_sections = 3,
950  cancel_taskgroup = 4
951 } kmp_cancel_kind_t;
952 
953 // KMP_HW_SUBSET support:
954 typedef struct kmp_hws_item {
955  int num;
956  int offset;
957 } kmp_hws_item_t;
958 
959 extern kmp_hws_item_t __kmp_hws_socket;
960 extern kmp_hws_item_t __kmp_hws_die;
961 extern kmp_hws_item_t __kmp_hws_node;
962 extern kmp_hws_item_t __kmp_hws_tile;
963 extern kmp_hws_item_t __kmp_hws_core;
964 extern kmp_hws_item_t __kmp_hws_proc;
965 extern int __kmp_hws_requested;
966 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
967 
968 /* ------------------------------------------------------------------------ */
969 
970 #define KMP_PAD(type, sz) \
971  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
972 
973 // We need to avoid using -1 as a GTID as +1 is added to the gtid
974 // when storing it in a lock, and the value 0 is reserved.
975 #define KMP_GTID_DNE (-2) /* Does not exist */
976 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
977 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
978 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
979 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
980 
981 /* OpenMP 5.0 Memory Management support */
982 
983 #ifndef __OMP_H
984 // Duplicate type definitions from omp.h
985 typedef uintptr_t omp_uintptr_t;
986 
987 typedef enum {
988  omp_atk_sync_hint = 1,
989  omp_atk_alignment = 2,
990  omp_atk_access = 3,
991  omp_atk_pool_size = 4,
992  omp_atk_fallback = 5,
993  omp_atk_fb_data = 6,
994  omp_atk_pinned = 7,
995  omp_atk_partition = 8
996 } omp_alloctrait_key_t;
997 
998 typedef enum {
999  omp_atv_false = 0,
1000  omp_atv_true = 1,
1001  omp_atv_contended = 3,
1002  omp_atv_uncontended = 4,
1003  omp_atv_serialized = 5,
1004  omp_atv_sequential = omp_atv_serialized, // (deprecated)
1005  omp_atv_private = 6,
1006  omp_atv_all = 7,
1007  omp_atv_thread = 8,
1008  omp_atv_pteam = 9,
1009  omp_atv_cgroup = 10,
1010  omp_atv_default_mem_fb = 11,
1011  omp_atv_null_fb = 12,
1012  omp_atv_abort_fb = 13,
1013  omp_atv_allocator_fb = 14,
1014  omp_atv_environment = 15,
1015  omp_atv_nearest = 16,
1016  omp_atv_blocked = 17,
1017  omp_atv_interleaved = 18
1018 } omp_alloctrait_value_t;
1019 #define omp_atv_default ((omp_uintptr_t)-1)
1020 
1021 typedef void *omp_memspace_handle_t;
1022 extern omp_memspace_handle_t const omp_default_mem_space;
1023 extern omp_memspace_handle_t const omp_large_cap_mem_space;
1024 extern omp_memspace_handle_t const omp_const_mem_space;
1025 extern omp_memspace_handle_t const omp_high_bw_mem_space;
1026 extern omp_memspace_handle_t const omp_low_lat_mem_space;
1027 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1028 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1029 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1030 
1031 typedef struct {
1032  omp_alloctrait_key_t key;
1033  omp_uintptr_t value;
1034 } omp_alloctrait_t;
1035 
1036 typedef void *omp_allocator_handle_t;
1037 extern omp_allocator_handle_t const omp_null_allocator;
1038 extern omp_allocator_handle_t const omp_default_mem_alloc;
1039 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1040 extern omp_allocator_handle_t const omp_const_mem_alloc;
1041 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1042 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1043 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1044 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1045 extern omp_allocator_handle_t const omp_thread_mem_alloc;
1046 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1047 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1048 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1049 extern omp_allocator_handle_t const kmp_max_mem_alloc;
1050 extern omp_allocator_handle_t __kmp_def_allocator;
1051 
1052 // end of duplicate type definitions from omp.h
1053 #endif
1054 
1055 extern int __kmp_memkind_available;
1056 
1057 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1058 
1059 typedef struct kmp_allocator_t {
1060  omp_memspace_handle_t memspace;
1061  void **memkind; // pointer to memkind
1062  size_t alignment;
1063  omp_alloctrait_value_t fb;
1064  kmp_allocator_t *fb_data;
1065  kmp_uint64 pool_size;
1066  kmp_uint64 pool_used;
1067  bool pinned;
1068 } kmp_allocator_t;
1069 
1070 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1071  omp_memspace_handle_t,
1072  int ntraits,
1073  omp_alloctrait_t traits[]);
1074 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1075 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1076 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1077 // external interfaces, may be used by compiler
1078 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1079 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1080  omp_allocator_handle_t al);
1081 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1082  omp_allocator_handle_t al);
1083 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1084  omp_allocator_handle_t al,
1085  omp_allocator_handle_t free_al);
1086 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1087 // internal interfaces, contain real implementation
1088 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1089  omp_allocator_handle_t al);
1090 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1091  omp_allocator_handle_t al);
1092 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1093  omp_allocator_handle_t al,
1094  omp_allocator_handle_t free_al);
1095 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1096 
1097 extern void __kmp_init_memkind();
1098 extern void __kmp_fini_memkind();
1099 extern void __kmp_init_target_mem();
1100 
1101 /* ------------------------------------------------------------------------ */
1102 
1103 #if ENABLE_LIBOMPTARGET
1104 extern void __kmp_init_target_task();
1105 #endif
1106 
1107 /* ------------------------------------------------------------------------ */
1108 
1109 #define KMP_UINT64_MAX \
1110  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1111 
1112 #define KMP_MIN_NTH 1
1113 
1114 #ifndef KMP_MAX_NTH
1115 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1116 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1117 #else
1118 #define KMP_MAX_NTH INT_MAX
1119 #endif
1120 #endif /* KMP_MAX_NTH */
1121 
1122 #ifdef PTHREAD_STACK_MIN
1123 #define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1124 #else
1125 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1126 #endif
1127 
1128 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1129 
1130 #if KMP_ARCH_X86
1131 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1132 #elif KMP_ARCH_X86_64
1133 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1134 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1135 #else
1136 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1137 #endif
1138 
1139 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1140 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1141 #define KMP_MAX_MALLOC_POOL_INCR \
1142  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1143 
1144 #define KMP_MIN_STKOFFSET (0)
1145 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1146 #if KMP_OS_DARWIN
1147 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1148 #else
1149 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1150 #endif
1151 
1152 #define KMP_MIN_STKPADDING (0)
1153 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1154 
1155 #define KMP_BLOCKTIME_MULTIPLIER \
1156  (1000) /* number of blocktime units per second */
1157 #define KMP_MIN_BLOCKTIME (0)
1158 #define KMP_MAX_BLOCKTIME \
1159  (INT_MAX) /* Must be this for "infinite" setting the work */
1160 
1161 /* __kmp_blocktime is in milliseconds */
1162 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200))
1163 
1164 #if KMP_USE_MONITOR
1165 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1166 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1167 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1168 
1169 /* Calculate new number of monitor wakeups for a specific block time based on
1170  previous monitor_wakeups. Only allow increasing number of wakeups */
1171 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1172  (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1173  : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1174  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1175  ? (monitor_wakeups) \
1176  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1177 
1178 /* Calculate number of intervals for a specific block time based on
1179  monitor_wakeups */
1180 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1181  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1182  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1183 #else
1184 #define KMP_BLOCKTIME(team, tid) \
1185  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1186 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1187 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1188 extern kmp_uint64 __kmp_ticks_per_msec;
1189 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1190 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1191 #else
1192 #define KMP_NOW() __kmp_hardware_timestamp()
1193 #endif
1194 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1195 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1196  (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1197 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1198 #else
1199 // System time is retrieved sporadically while blocking.
1200 extern kmp_uint64 __kmp_now_nsec();
1201 #define KMP_NOW() __kmp_now_nsec()
1202 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1203 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1204  (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1205 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1206 #endif
1207 #endif // KMP_USE_MONITOR
1208 
1209 #define KMP_MIN_STATSCOLS 40
1210 #define KMP_MAX_STATSCOLS 4096
1211 #define KMP_DEFAULT_STATSCOLS 80
1212 
1213 #define KMP_MIN_INTERVAL 0
1214 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1215 #define KMP_DEFAULT_INTERVAL 0
1216 
1217 #define KMP_MIN_CHUNK 1
1218 #define KMP_MAX_CHUNK (INT_MAX - 1)
1219 #define KMP_DEFAULT_CHUNK 1
1220 
1221 #define KMP_MIN_DISP_NUM_BUFF 1
1222 #define KMP_DFLT_DISP_NUM_BUFF 7
1223 #define KMP_MAX_DISP_NUM_BUFF 4096
1224 
1225 #define KMP_MAX_ORDERED 8
1226 
1227 #define KMP_MAX_FIELDS 32
1228 
1229 #define KMP_MAX_BRANCH_BITS 31
1230 
1231 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1232 
1233 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1234 
1235 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1236 
1237 /* Minimum number of threads before switch to TLS gtid (experimentally
1238  determined) */
1239 /* josh TODO: what about OS X* tuning? */
1240 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1241 #define KMP_TLS_GTID_MIN 5
1242 #else
1243 #define KMP_TLS_GTID_MIN INT_MAX
1244 #endif
1245 
1246 #define KMP_MASTER_TID(tid) (0 == (tid))
1247 #define KMP_WORKER_TID(tid) (0 != (tid))
1248 
1249 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1250 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1251 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1252 
1253 #ifndef TRUE
1254 #define FALSE 0
1255 #define TRUE (!FALSE)
1256 #endif
1257 
1258 /* NOTE: all of the following constants must be even */
1259 
1260 #if KMP_OS_WINDOWS
1261 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1262 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1263 #elif KMP_OS_LINUX
1264 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1265 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1266 #elif KMP_OS_DARWIN
1267 /* TODO: tune for KMP_OS_DARWIN */
1268 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1269 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1270 #elif KMP_OS_DRAGONFLY
1271 /* TODO: tune for KMP_OS_DRAGONFLY */
1272 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1273 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1274 #elif KMP_OS_FREEBSD
1275 /* TODO: tune for KMP_OS_FREEBSD */
1276 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1277 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1278 #elif KMP_OS_NETBSD
1279 /* TODO: tune for KMP_OS_NETBSD */
1280 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1281 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1282 #elif KMP_OS_HURD
1283 /* TODO: tune for KMP_OS_HURD */
1284 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1285 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1286 #elif KMP_OS_OPENBSD
1287 /* TODO: tune for KMP_OS_OPENBSD */
1288 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1289 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1290 #endif
1291 
1292 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1293 typedef struct kmp_cpuid {
1294  kmp_uint32 eax;
1295  kmp_uint32 ebx;
1296  kmp_uint32 ecx;
1297  kmp_uint32 edx;
1298 } kmp_cpuid_t;
1299 
1300 typedef struct kmp_cpuinfo_flags_t {
1301  unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1302  unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1303  unsigned hybrid : 1;
1304  unsigned reserved : 29; // Ensure size of 32 bits
1305 } kmp_cpuinfo_flags_t;
1306 
1307 typedef struct kmp_cpuinfo {
1308  int initialized; // If 0, other fields are not initialized.
1309  int signature; // CPUID(1).EAX
1310  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1311  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1312  // Model << 4 ) + Model)
1313  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1314  kmp_cpuinfo_flags_t flags;
1315  int apic_id;
1316  int physical_id;
1317  int logical_id;
1318  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1319  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1320 } kmp_cpuinfo_t;
1321 
1322 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1323 
1324 #if KMP_OS_UNIX
1325 // subleaf is only needed for cache and topology discovery and can be set to
1326 // zero in most cases
1327 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1328  __asm__ __volatile__("cpuid"
1329  : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1330  : "a"(leaf), "c"(subleaf));
1331 }
1332 // Load p into FPU control word
1333 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1334  __asm__ __volatile__("fldcw %0" : : "m"(*p));
1335 }
1336 // Store FPU control word into p
1337 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1338  __asm__ __volatile__("fstcw %0" : "=m"(*p));
1339 }
1340 static inline void __kmp_clear_x87_fpu_status_word() {
1341 #if KMP_MIC
1342  // 32-bit protected mode x87 FPU state
1343  struct x87_fpu_state {
1344  unsigned cw;
1345  unsigned sw;
1346  unsigned tw;
1347  unsigned fip;
1348  unsigned fips;
1349  unsigned fdp;
1350  unsigned fds;
1351  };
1352  struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1353  __asm__ __volatile__("fstenv %0\n\t" // store FP env
1354  "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1355  "fldenv %0\n\t" // load FP env back
1356  : "+m"(fpu_state), "+m"(fpu_state.sw));
1357 #else
1358  __asm__ __volatile__("fnclex");
1359 #endif // KMP_MIC
1360 }
1361 #if __SSE__
1362 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1363 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1364 #else
1365 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1366 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1367 #endif
1368 #else
1369 // Windows still has these as external functions in assembly file
1370 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1371 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1372 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1373 extern void __kmp_clear_x87_fpu_status_word();
1374 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1375 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1376 #endif // KMP_OS_UNIX
1377 
1378 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1379 
1380 // User-level Monitor/Mwait
1381 #if KMP_HAVE_UMWAIT
1382 // We always try for UMWAIT first
1383 #if KMP_HAVE_WAITPKG_INTRINSICS
1384 #if KMP_HAVE_IMMINTRIN_H
1385 #include <immintrin.h>
1386 #elif KMP_HAVE_INTRIN_H
1387 #include <intrin.h>
1388 #endif
1389 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1390 
1391 KMP_ATTRIBUTE_TARGET_WAITPKG
1392 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1393 #if !KMP_HAVE_WAITPKG_INTRINSICS
1394  uint32_t timeHi = uint32_t(counter >> 32);
1395  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1396  char flag;
1397  __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1398  "setb %0"
1399  // The "=q" restraint means any register accessible as rl
1400  // in 32-bit mode: a, b, c, and d;
1401  // in 64-bit mode: any integer register
1402  : "=q"(flag)
1403  : "a"(timeLo), "d"(timeHi), "c"(hint)
1404  :);
1405  return flag;
1406 #else
1407  return _tpause(hint, counter);
1408 #endif
1409 }
1410 KMP_ATTRIBUTE_TARGET_WAITPKG
1411 static inline void __kmp_umonitor(void *cacheline) {
1412 #if !KMP_HAVE_WAITPKG_INTRINSICS
1413  __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1414  :
1415  : "a"(cacheline)
1416  :);
1417 #else
1418  _umonitor(cacheline);
1419 #endif
1420 }
1421 KMP_ATTRIBUTE_TARGET_WAITPKG
1422 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1423 #if !KMP_HAVE_WAITPKG_INTRINSICS
1424  uint32_t timeHi = uint32_t(counter >> 32);
1425  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1426  char flag;
1427  __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1428  "setb %0"
1429  // The "=q" restraint means any register accessible as rl
1430  // in 32-bit mode: a, b, c, and d;
1431  // in 64-bit mode: any integer register
1432  : "=q"(flag)
1433  : "a"(timeLo), "d"(timeHi), "c"(hint)
1434  :);
1435  return flag;
1436 #else
1437  return _umwait(hint, counter);
1438 #endif
1439 }
1440 #elif KMP_HAVE_MWAIT
1441 #if KMP_OS_UNIX
1442 #include <pmmintrin.h>
1443 #else
1444 #include <intrin.h>
1445 #endif
1446 #if KMP_OS_UNIX
1447 __attribute__((target("sse3")))
1448 #endif
1449 static inline void
1450 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1451  _mm_monitor(cacheline, extensions, hints);
1452 }
1453 #if KMP_OS_UNIX
1454 __attribute__((target("sse3")))
1455 #endif
1456 static inline void
1457 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1458  _mm_mwait(extensions, hints);
1459 }
1460 #endif // KMP_HAVE_UMWAIT
1461 
1462 #if KMP_ARCH_X86
1463 extern void __kmp_x86_pause(void);
1464 #elif KMP_MIC
1465 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1466 // regression after removal of extra PAUSE from spin loops. Changing
1467 // the delay from 100 to 300 showed even better performance than double PAUSE
1468 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1469 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1470 #else
1471 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1472 #endif
1473 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1474 #elif KMP_ARCH_PPC64
1475 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1476 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1477 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1478 #define KMP_CPU_PAUSE() \
1479  do { \
1480  KMP_PPC64_PRI_LOW(); \
1481  KMP_PPC64_PRI_MED(); \
1482  KMP_PPC64_PRI_LOC_MB(); \
1483  } while (0)
1484 #else
1485 #define KMP_CPU_PAUSE() /* nothing to do */
1486 #endif
1487 
1488 #define KMP_INIT_YIELD(count) \
1489  { (count) = __kmp_yield_init; }
1490 
1491 #define KMP_INIT_BACKOFF(time) \
1492  { (time) = __kmp_pause_init; }
1493 
1494 #define KMP_OVERSUBSCRIBED \
1495  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1496 
1497 #define KMP_TRY_YIELD \
1498  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1499 
1500 #define KMP_TRY_YIELD_OVERSUB \
1501  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1502 
1503 #define KMP_YIELD(cond) \
1504  { \
1505  KMP_CPU_PAUSE(); \
1506  if ((cond) && (KMP_TRY_YIELD)) \
1507  __kmp_yield(); \
1508  }
1509 
1510 #define KMP_YIELD_OVERSUB() \
1511  { \
1512  KMP_CPU_PAUSE(); \
1513  if ((KMP_TRY_YIELD_OVERSUB)) \
1514  __kmp_yield(); \
1515  }
1516 
1517 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1518 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1519 #define KMP_YIELD_SPIN(count) \
1520  { \
1521  KMP_CPU_PAUSE(); \
1522  if (KMP_TRY_YIELD) { \
1523  (count) -= 2; \
1524  if (!(count)) { \
1525  __kmp_yield(); \
1526  (count) = __kmp_yield_next; \
1527  } \
1528  } \
1529  }
1530 
1531 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1532 // (C0.2) state, which improves performance of other SMT threads on the same
1533 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1534 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1535 // available, fall back to the regular CPU pause and yield combination.
1536 #if KMP_HAVE_UMWAIT
1537 #define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1538 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1539  { \
1540  if (__kmp_tpause_enabled) { \
1541  if (KMP_OVERSUBSCRIBED) { \
1542  __kmp_tpause(0, (time)); \
1543  } else { \
1544  __kmp_tpause(__kmp_tpause_hint, (time)); \
1545  } \
1546  (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1547  } else { \
1548  KMP_CPU_PAUSE(); \
1549  if ((KMP_TRY_YIELD_OVERSUB)) { \
1550  __kmp_yield(); \
1551  } else if (__kmp_use_yield == 1) { \
1552  (count) -= 2; \
1553  if (!(count)) { \
1554  __kmp_yield(); \
1555  (count) = __kmp_yield_next; \
1556  } \
1557  } \
1558  } \
1559  }
1560 #else
1561 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1562  { \
1563  KMP_CPU_PAUSE(); \
1564  if ((KMP_TRY_YIELD_OVERSUB)) \
1565  __kmp_yield(); \
1566  else if (__kmp_use_yield == 1) { \
1567  (count) -= 2; \
1568  if (!(count)) { \
1569  __kmp_yield(); \
1570  (count) = __kmp_yield_next; \
1571  } \
1572  } \
1573  }
1574 #endif // KMP_HAVE_UMWAIT
1575 
1576 /* ------------------------------------------------------------------------ */
1577 /* Support datatypes for the orphaned construct nesting checks. */
1578 /* ------------------------------------------------------------------------ */
1579 
1580 /* When adding to this enum, add its corresponding string in cons_text_c[]
1581  * array in kmp_error.cpp */
1582 enum cons_type {
1583  ct_none,
1584  ct_parallel,
1585  ct_pdo,
1586  ct_pdo_ordered,
1587  ct_psections,
1588  ct_psingle,
1589  ct_critical,
1590  ct_ordered_in_parallel,
1591  ct_ordered_in_pdo,
1592  ct_master,
1593  ct_reduce,
1594  ct_barrier,
1595  ct_masked
1596 };
1597 
1598 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1599 
1600 struct cons_data {
1601  ident_t const *ident;
1602  enum cons_type type;
1603  int prev;
1604  kmp_user_lock_p
1605  name; /* address exclusively for critical section name comparison */
1606 };
1607 
1608 struct cons_header {
1609  int p_top, w_top, s_top;
1610  int stack_size, stack_top;
1611  struct cons_data *stack_data;
1612 };
1613 
1614 struct kmp_region_info {
1615  char *text;
1616  int offset[KMP_MAX_FIELDS];
1617  int length[KMP_MAX_FIELDS];
1618 };
1619 
1620 /* ---------------------------------------------------------------------- */
1621 /* ---------------------------------------------------------------------- */
1622 
1623 #if KMP_OS_WINDOWS
1624 typedef HANDLE kmp_thread_t;
1625 typedef DWORD kmp_key_t;
1626 #endif /* KMP_OS_WINDOWS */
1627 
1628 #if KMP_OS_UNIX
1629 typedef pthread_t kmp_thread_t;
1630 typedef pthread_key_t kmp_key_t;
1631 #endif
1632 
1633 extern kmp_key_t __kmp_gtid_threadprivate_key;
1634 
1635 typedef struct kmp_sys_info {
1636  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1637  long minflt; /* the number of page faults serviced without any I/O */
1638  long majflt; /* the number of page faults serviced that required I/O */
1639  long nswap; /* the number of times a process was "swapped" out of memory */
1640  long inblock; /* the number of times the file system had to perform input */
1641  long oublock; /* the number of times the file system had to perform output */
1642  long nvcsw; /* the number of times a context switch was voluntarily */
1643  long nivcsw; /* the number of times a context switch was forced */
1644 } kmp_sys_info_t;
1645 
1646 #if USE_ITT_BUILD
1647 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1648 // required type here. Later we will check the type meets requirements.
1649 typedef int kmp_itt_mark_t;
1650 #define KMP_ITT_DEBUG 0
1651 #endif /* USE_ITT_BUILD */
1652 
1653 typedef kmp_int32 kmp_critical_name[8];
1654 
1664 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1665 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1666  ...);
1667 
1672 /* ---------------------------------------------------------------------------
1673  */
1674 /* Threadprivate initialization/finalization function declarations */
1675 
1676 /* for non-array objects: __kmpc_threadprivate_register() */
1677 
1682 typedef void *(*kmpc_ctor)(void *);
1683 
1688 typedef void (*kmpc_dtor)(
1689  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1690  compiler */
1695 typedef void *(*kmpc_cctor)(void *, void *);
1696 
1697 /* for array objects: __kmpc_threadprivate_register_vec() */
1698 /* First arg: "this" pointer */
1699 /* Last arg: number of array elements */
1705 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1711 typedef void (*kmpc_dtor_vec)(void *, size_t);
1717 typedef void *(*kmpc_cctor_vec)(void *, void *,
1718  size_t); /* function unused by compiler */
1719 
1724 /* keeps tracked of threadprivate cache allocations for cleanup later */
1725 typedef struct kmp_cached_addr {
1726  void **addr; /* address of allocated cache */
1727  void ***compiler_cache; /* pointer to compiler's cache */
1728  void *data; /* pointer to global data */
1729  struct kmp_cached_addr *next; /* pointer to next cached address */
1730 } kmp_cached_addr_t;
1731 
1732 struct private_data {
1733  struct private_data *next; /* The next descriptor in the list */
1734  void *data; /* The data buffer for this descriptor */
1735  int more; /* The repeat count for this descriptor */
1736  size_t size; /* The data size for this descriptor */
1737 };
1738 
1739 struct private_common {
1740  struct private_common *next;
1741  struct private_common *link;
1742  void *gbl_addr;
1743  void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1744  size_t cmn_size;
1745 };
1746 
1747 struct shared_common {
1748  struct shared_common *next;
1749  struct private_data *pod_init;
1750  void *obj_init;
1751  void *gbl_addr;
1752  union {
1753  kmpc_ctor ctor;
1754  kmpc_ctor_vec ctorv;
1755  } ct;
1756  union {
1757  kmpc_cctor cctor;
1758  kmpc_cctor_vec cctorv;
1759  } cct;
1760  union {
1761  kmpc_dtor dtor;
1762  kmpc_dtor_vec dtorv;
1763  } dt;
1764  size_t vec_len;
1765  int is_vec;
1766  size_t cmn_size;
1767 };
1768 
1769 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1770 #define KMP_HASH_TABLE_SIZE \
1771  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1772 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1773 #define KMP_HASH(x) \
1774  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1775 
1776 struct common_table {
1777  struct private_common *data[KMP_HASH_TABLE_SIZE];
1778 };
1779 
1780 struct shared_table {
1781  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1782 };
1783 
1784 /* ------------------------------------------------------------------------ */
1785 
1786 #if KMP_USE_HIER_SCHED
1787 // Shared barrier data that exists inside a single unit of the scheduling
1788 // hierarchy
1789 typedef struct kmp_hier_private_bdata_t {
1790  kmp_int32 num_active;
1791  kmp_uint64 index;
1792  kmp_uint64 wait_val[2];
1793 } kmp_hier_private_bdata_t;
1794 #endif
1795 
1796 typedef struct kmp_sched_flags {
1797  unsigned ordered : 1;
1798  unsigned nomerge : 1;
1799  unsigned contains_last : 1;
1800 #if KMP_USE_HIER_SCHED
1801  unsigned use_hier : 1;
1802  unsigned unused : 28;
1803 #else
1804  unsigned unused : 29;
1805 #endif
1806 } kmp_sched_flags_t;
1807 
1808 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1809 
1810 #if KMP_STATIC_STEAL_ENABLED
1811 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1812  kmp_int32 count;
1813  kmp_int32 ub;
1814  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1815  kmp_int32 lb;
1816  kmp_int32 st;
1817  kmp_int32 tc;
1818  kmp_lock_t *steal_lock; // lock used for chunk stealing
1819  // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1820  // a) parm3 is properly aligned and
1821  // b) all parm1-4 are on the same cache line.
1822  // Because of parm1-4 are used together, performance seems to be better
1823  // if they are on the same cache line (not measured though).
1824 
1825  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1826  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1827  kmp_int32 parm2; // make no real change at least while padding is off.
1828  kmp_int32 parm3;
1829  kmp_int32 parm4;
1830  };
1831 
1832  kmp_uint32 ordered_lower;
1833  kmp_uint32 ordered_upper;
1834 #if KMP_OS_WINDOWS
1835  kmp_int32 last_upper;
1836 #endif /* KMP_OS_WINDOWS */
1837 } dispatch_private_info32_t;
1838 
1839 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1840  kmp_int64 count; // current chunk number for static & static-steal scheduling
1841  kmp_int64 ub; /* upper-bound */
1842  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1843  kmp_int64 lb; /* lower-bound */
1844  kmp_int64 st; /* stride */
1845  kmp_int64 tc; /* trip count (number of iterations) */
1846  kmp_lock_t *steal_lock; // lock used for chunk stealing
1847  /* parm[1-4] are used in different ways by different scheduling algorithms */
1848 
1849  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1850  // a) parm3 is properly aligned and
1851  // b) all parm1-4 are in the same cache line.
1852  // Because of parm1-4 are used together, performance seems to be better
1853  // if they are in the same line (not measured though).
1854 
1855  struct KMP_ALIGN(32) {
1856  kmp_int64 parm1;
1857  kmp_int64 parm2;
1858  kmp_int64 parm3;
1859  kmp_int64 parm4;
1860  };
1861 
1862  kmp_uint64 ordered_lower;
1863  kmp_uint64 ordered_upper;
1864 #if KMP_OS_WINDOWS
1865  kmp_int64 last_upper;
1866 #endif /* KMP_OS_WINDOWS */
1867 } dispatch_private_info64_t;
1868 #else /* KMP_STATIC_STEAL_ENABLED */
1869 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1870  kmp_int32 lb;
1871  kmp_int32 ub;
1872  kmp_int32 st;
1873  kmp_int32 tc;
1874 
1875  kmp_int32 parm1;
1876  kmp_int32 parm2;
1877  kmp_int32 parm3;
1878  kmp_int32 parm4;
1879 
1880  kmp_int32 count;
1881 
1882  kmp_uint32 ordered_lower;
1883  kmp_uint32 ordered_upper;
1884 #if KMP_OS_WINDOWS
1885  kmp_int32 last_upper;
1886 #endif /* KMP_OS_WINDOWS */
1887 } dispatch_private_info32_t;
1888 
1889 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1890  kmp_int64 lb; /* lower-bound */
1891  kmp_int64 ub; /* upper-bound */
1892  kmp_int64 st; /* stride */
1893  kmp_int64 tc; /* trip count (number of iterations) */
1894 
1895  /* parm[1-4] are used in different ways by different scheduling algorithms */
1896  kmp_int64 parm1;
1897  kmp_int64 parm2;
1898  kmp_int64 parm3;
1899  kmp_int64 parm4;
1900 
1901  kmp_int64 count; /* current chunk number for static scheduling */
1902 
1903  kmp_uint64 ordered_lower;
1904  kmp_uint64 ordered_upper;
1905 #if KMP_OS_WINDOWS
1906  kmp_int64 last_upper;
1907 #endif /* KMP_OS_WINDOWS */
1908 } dispatch_private_info64_t;
1909 #endif /* KMP_STATIC_STEAL_ENABLED */
1910 
1911 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1912  union private_info {
1913  dispatch_private_info32_t p32;
1914  dispatch_private_info64_t p64;
1915  } u;
1916  enum sched_type schedule; /* scheduling algorithm */
1917  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1918  std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1919  kmp_int32 ordered_bumped;
1920  // Stack of buffers for nest of serial regions
1921  struct dispatch_private_info *next;
1922  kmp_int32 type_size; /* the size of types in private_info */
1923 #if KMP_USE_HIER_SCHED
1924  kmp_int32 hier_id;
1925  void *parent; /* hierarchical scheduling parent pointer */
1926 #endif
1927  enum cons_type pushed_ws;
1928 } dispatch_private_info_t;
1929 
1930 typedef struct dispatch_shared_info32 {
1931  /* chunk index under dynamic, number of idle threads under static-steal;
1932  iteration index otherwise */
1933  volatile kmp_uint32 iteration;
1934  volatile kmp_int32 num_done;
1935  volatile kmp_uint32 ordered_iteration;
1936  // Dummy to retain the structure size after making ordered_iteration scalar
1937  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1938 } dispatch_shared_info32_t;
1939 
1940 typedef struct dispatch_shared_info64 {
1941  /* chunk index under dynamic, number of idle threads under static-steal;
1942  iteration index otherwise */
1943  volatile kmp_uint64 iteration;
1944  volatile kmp_int64 num_done;
1945  volatile kmp_uint64 ordered_iteration;
1946  // Dummy to retain the structure size after making ordered_iteration scalar
1947  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1948 } dispatch_shared_info64_t;
1949 
1950 typedef struct dispatch_shared_info {
1951  union shared_info {
1952  dispatch_shared_info32_t s32;
1953  dispatch_shared_info64_t s64;
1954  } u;
1955  volatile kmp_uint32 buffer_index;
1956  volatile kmp_int32 doacross_buf_idx; // teamwise index
1957  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1958  kmp_int32 doacross_num_done; // count finished threads
1959 #if KMP_USE_HIER_SCHED
1960  void *hier;
1961 #endif
1962 #if KMP_USE_HWLOC
1963  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1964  // machines (> 48 cores). Performance analysis showed that a cache thrash
1965  // was occurring and this padding helps alleviate the problem.
1966  char padding[64];
1967 #endif
1968 } dispatch_shared_info_t;
1969 
1970 typedef struct kmp_disp {
1971  /* Vector for ORDERED SECTION */
1972  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1973  /* Vector for END ORDERED SECTION */
1974  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1975 
1976  dispatch_shared_info_t *th_dispatch_sh_current;
1977  dispatch_private_info_t *th_dispatch_pr_current;
1978 
1979  dispatch_private_info_t *th_disp_buffer;
1980  kmp_uint32 th_disp_index;
1981  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1982  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1983  kmp_int64 *th_doacross_info; // info on loop bounds
1984 #if KMP_USE_INTERNODE_ALIGNMENT
1985  char more_padding[INTERNODE_CACHE_LINE];
1986 #endif
1987 } kmp_disp_t;
1988 
1989 /* ------------------------------------------------------------------------ */
1990 /* Barrier stuff */
1991 
1992 /* constants for barrier state update */
1993 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1994 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1995 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1996 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1997 
1998 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1999 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2000 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2001 
2002 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2003 #error "Barrier sleep bit must be smaller than barrier bump bit"
2004 #endif
2005 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2006 #error "Barrier unused bit must be smaller than barrier bump bit"
2007 #endif
2008 
2009 // Constants for release barrier wait state: currently, hierarchical only
2010 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2011 #define KMP_BARRIER_OWN_FLAG \
2012  1 // Normal state; worker waiting on own b_go flag in release
2013 #define KMP_BARRIER_PARENT_FLAG \
2014  2 // Special state; worker waiting on parent's b_go flag in release
2015 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2016  3 // Special state; tells worker to shift from parent to own b_go
2017 #define KMP_BARRIER_SWITCHING \
2018  4 // Special state; worker resets appropriate flag on wake-up
2019 
2020 #define KMP_NOT_SAFE_TO_REAP \
2021  0 // Thread th_reap_state: not safe to reap (tasking)
2022 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2023 
2024 // The flag_type describes the storage used for the flag.
2025 enum flag_type {
2026  flag32,
2027  flag64,
2028  atomic_flag64,
2029  flag_oncore,
2030  flag_unset
2031 };
2032 
2033 enum barrier_type {
2034  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2035  barriers if enabled) */
2036  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2037 #if KMP_FAST_REDUCTION_BARRIER
2038  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2039 #endif // KMP_FAST_REDUCTION_BARRIER
2040  bs_last_barrier /* Just a placeholder to mark the end */
2041 };
2042 
2043 // to work with reduction barriers just like with plain barriers
2044 #if !KMP_FAST_REDUCTION_BARRIER
2045 #define bs_reduction_barrier bs_plain_barrier
2046 #endif // KMP_FAST_REDUCTION_BARRIER
2047 
2048 typedef enum kmp_bar_pat { /* Barrier communication patterns */
2049  bp_linear_bar =
2050  0, /* Single level (degenerate) tree */
2051  bp_tree_bar =
2052  1, /* Balanced tree with branching factor 2^n */
2053  bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2054  branching factor 2^n */
2055  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2056  bp_dist_bar = 4, /* Distributed barrier */
2057  bp_last_bar /* Placeholder to mark the end */
2058 } kmp_bar_pat_e;
2059 
2060 #define KMP_BARRIER_ICV_PUSH 1
2061 
2062 /* Record for holding the values of the internal controls stack records */
2063 typedef struct kmp_internal_control {
2064  int serial_nesting_level; /* corresponds to the value of the
2065  th_team_serialized field */
2066  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2067  thread) */
2068  kmp_int8
2069  bt_set; /* internal control for whether blocktime is explicitly set */
2070  int blocktime; /* internal control for blocktime */
2071 #if KMP_USE_MONITOR
2072  int bt_intervals; /* internal control for blocktime intervals */
2073 #endif
2074  int nproc; /* internal control for #threads for next parallel region (per
2075  thread) */
2076  int thread_limit; /* internal control for thread-limit-var */
2077  int max_active_levels; /* internal control for max_active_levels */
2078  kmp_r_sched_t
2079  sched; /* internal control for runtime schedule {sched,chunk} pair */
2080  kmp_proc_bind_t proc_bind; /* internal control for affinity */
2081  kmp_int32 default_device; /* internal control for default device */
2082  struct kmp_internal_control *next;
2083 } kmp_internal_control_t;
2084 
2085 static inline void copy_icvs(kmp_internal_control_t *dst,
2086  kmp_internal_control_t *src) {
2087  *dst = *src;
2088 }
2089 
2090 /* Thread barrier needs volatile barrier fields */
2091 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2092  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2093  // uses of it). It is not explicitly aligned below, because we *don't* want
2094  // it to be padded -- instead, we fit b_go into the same cache line with
2095  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2096  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2097  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2098  // same NGO store
2099  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2100  KMP_ALIGN_CACHE volatile kmp_uint64
2101  b_arrived; // STATE => task reached synch point.
2102  kmp_uint32 *skip_per_level;
2103  kmp_uint32 my_level;
2104  kmp_int32 parent_tid;
2105  kmp_int32 old_tid;
2106  kmp_uint32 depth;
2107  struct kmp_bstate *parent_bar;
2108  kmp_team_t *team;
2109  kmp_uint64 leaf_state;
2110  kmp_uint32 nproc;
2111  kmp_uint8 base_leaf_kids;
2112  kmp_uint8 leaf_kids;
2113  kmp_uint8 offset;
2114  kmp_uint8 wait_flag;
2115  kmp_uint8 use_oncore_barrier;
2116 #if USE_DEBUGGER
2117  // The following field is intended for the debugger solely. Only the worker
2118  // thread itself accesses this field: the worker increases it by 1 when it
2119  // arrives to a barrier.
2120  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2121 #endif /* USE_DEBUGGER */
2122 } kmp_bstate_t;
2123 
2124 union KMP_ALIGN_CACHE kmp_barrier_union {
2125  double b_align; /* use worst case alignment */
2126  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2127  kmp_bstate_t bb;
2128 };
2129 
2130 typedef union kmp_barrier_union kmp_balign_t;
2131 
2132 /* Team barrier needs only non-volatile arrived counter */
2133 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2134  double b_align; /* use worst case alignment */
2135  char b_pad[CACHE_LINE];
2136  struct {
2137  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2138 #if USE_DEBUGGER
2139  // The following two fields are indended for the debugger solely. Only
2140  // primary thread of the team accesses these fields: the first one is
2141  // increased by 1 when the primary thread arrives to a barrier, the second
2142  // one is increased by one when all the threads arrived.
2143  kmp_uint b_master_arrived;
2144  kmp_uint b_team_arrived;
2145 #endif
2146  };
2147 };
2148 
2149 typedef union kmp_barrier_team_union kmp_balign_team_t;
2150 
2151 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2152  threads when a condition changes. This is to workaround an NPTL bug where
2153  padding was added to pthread_cond_t which caused the initialization routine
2154  to write outside of the structure if compiled on pre-NPTL threads. */
2155 #if KMP_OS_WINDOWS
2156 typedef struct kmp_win32_mutex {
2157  /* The Lock */
2158  CRITICAL_SECTION cs;
2159 } kmp_win32_mutex_t;
2160 
2161 typedef struct kmp_win32_cond {
2162  /* Count of the number of waiters. */
2163  int waiters_count_;
2164 
2165  /* Serialize access to <waiters_count_> */
2166  kmp_win32_mutex_t waiters_count_lock_;
2167 
2168  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2169  int release_count_;
2170 
2171  /* Keeps track of the current "generation" so that we don't allow */
2172  /* one thread to steal all the "releases" from the broadcast. */
2173  int wait_generation_count_;
2174 
2175  /* A manual-reset event that's used to block and release waiting threads. */
2176  HANDLE event_;
2177 } kmp_win32_cond_t;
2178 #endif
2179 
2180 #if KMP_OS_UNIX
2181 
2182 union KMP_ALIGN_CACHE kmp_cond_union {
2183  double c_align;
2184  char c_pad[CACHE_LINE];
2185  pthread_cond_t c_cond;
2186 };
2187 
2188 typedef union kmp_cond_union kmp_cond_align_t;
2189 
2190 union KMP_ALIGN_CACHE kmp_mutex_union {
2191  double m_align;
2192  char m_pad[CACHE_LINE];
2193  pthread_mutex_t m_mutex;
2194 };
2195 
2196 typedef union kmp_mutex_union kmp_mutex_align_t;
2197 
2198 #endif /* KMP_OS_UNIX */
2199 
2200 typedef struct kmp_desc_base {
2201  void *ds_stackbase;
2202  size_t ds_stacksize;
2203  int ds_stackgrow;
2204  kmp_thread_t ds_thread;
2205  volatile int ds_tid;
2206  int ds_gtid;
2207 #if KMP_OS_WINDOWS
2208  volatile int ds_alive;
2209  DWORD ds_thread_id;
2210 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2211  However, debugger support (libomp_db) cannot work with handles, because they
2212  uncomparable. For example, debugger requests info about thread with handle h.
2213  h is valid within debugger process, and meaningless within debugee process.
2214  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2215  within debugee process, but it is a *new* handle which does *not* equal to
2216  any other handle in debugee... The only way to compare handles is convert
2217  them to system-wide ids. GetThreadId() function is available only in
2218  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2219  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2220  thread id by call to GetCurrentThreadId() from within the thread and save it
2221  to let libomp_db identify threads. */
2222 #endif /* KMP_OS_WINDOWS */
2223 } kmp_desc_base_t;
2224 
2225 typedef union KMP_ALIGN_CACHE kmp_desc {
2226  double ds_align; /* use worst case alignment */
2227  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2228  kmp_desc_base_t ds;
2229 } kmp_desc_t;
2230 
2231 typedef struct kmp_local {
2232  volatile int this_construct; /* count of single's encountered by thread */
2233  void *reduce_data;
2234 #if KMP_USE_BGET
2235  void *bget_data;
2236  void *bget_list;
2237 #if !USE_CMP_XCHG_FOR_BGET
2238 #ifdef USE_QUEUING_LOCK_FOR_BGET
2239  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2240 #else
2241  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2242 // bootstrap lock so we can use it at library
2243 // shutdown.
2244 #endif /* USE_LOCK_FOR_BGET */
2245 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2246 #endif /* KMP_USE_BGET */
2247 
2248  PACKED_REDUCTION_METHOD_T
2249  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2250  __kmpc_end_reduce*() */
2251 
2252 } kmp_local_t;
2253 
2254 #define KMP_CHECK_UPDATE(a, b) \
2255  if ((a) != (b)) \
2256  (a) = (b)
2257 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2258  if ((a) != (b)) \
2259  TCW_SYNC_PTR((a), (b))
2260 
2261 #define get__blocktime(xteam, xtid) \
2262  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2263 #define get__bt_set(xteam, xtid) \
2264  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2265 #if KMP_USE_MONITOR
2266 #define get__bt_intervals(xteam, xtid) \
2267  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2268 #endif
2269 
2270 #define get__dynamic_2(xteam, xtid) \
2271  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2272 #define get__nproc_2(xteam, xtid) \
2273  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2274 #define get__sched_2(xteam, xtid) \
2275  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2276 
2277 #define set__blocktime_team(xteam, xtid, xval) \
2278  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2279  (xval))
2280 
2281 #if KMP_USE_MONITOR
2282 #define set__bt_intervals_team(xteam, xtid, xval) \
2283  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2284  (xval))
2285 #endif
2286 
2287 #define set__bt_set_team(xteam, xtid, xval) \
2288  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2289 
2290 #define set__dynamic(xthread, xval) \
2291  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2292 #define get__dynamic(xthread) \
2293  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2294 
2295 #define set__nproc(xthread, xval) \
2296  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2297 
2298 #define set__thread_limit(xthread, xval) \
2299  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2300 
2301 #define set__max_active_levels(xthread, xval) \
2302  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2303 
2304 #define get__max_active_levels(xthread) \
2305  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2306 
2307 #define set__sched(xthread, xval) \
2308  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2309 
2310 #define set__proc_bind(xthread, xval) \
2311  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2312 #define get__proc_bind(xthread) \
2313  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2314 
2315 // OpenMP tasking data structures
2316 
2317 typedef enum kmp_tasking_mode {
2318  tskm_immediate_exec = 0,
2319  tskm_extra_barrier = 1,
2320  tskm_task_teams = 2,
2321  tskm_max = 2
2322 } kmp_tasking_mode_t;
2323 
2324 extern kmp_tasking_mode_t
2325  __kmp_tasking_mode; /* determines how/when to execute tasks */
2326 extern int __kmp_task_stealing_constraint;
2327 extern int __kmp_enable_task_throttling;
2328 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2329 // specified, defaults to 0 otherwise
2330 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2331 extern kmp_int32 __kmp_max_task_priority;
2332 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2333 extern kmp_uint64 __kmp_taskloop_min_tasks;
2334 
2335 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2336  taskdata first */
2337 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2338 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2339 
2340 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2341 // were spawned and queued since the previous barrier release.
2342 #define KMP_TASKING_ENABLED(task_team) \
2343  (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2344 
2351 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2352 
2353 typedef union kmp_cmplrdata {
2354  kmp_int32 priority;
2355  kmp_routine_entry_t
2356  destructors; /* pointer to function to invoke deconstructors of
2357  firstprivate C++ objects */
2358  /* future data */
2359 } kmp_cmplrdata_t;
2360 
2361 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2364 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2365  void *shareds;
2366  kmp_routine_entry_t
2367  routine;
2368  kmp_int32 part_id;
2369  kmp_cmplrdata_t
2370  data1; /* Two known optional additions: destructors and priority */
2371  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2372  /* future data */
2373  /* private vars */
2374 } kmp_task_t;
2375 
2380 typedef struct kmp_taskgroup {
2381  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2382  std::atomic<kmp_int32>
2383  cancel_request; // request for cancellation of this taskgroup
2384  struct kmp_taskgroup *parent; // parent taskgroup
2385  // Block of data to perform task reduction
2386  void *reduce_data; // reduction related info
2387  kmp_int32 reduce_num_data; // number of data items to reduce
2388  uintptr_t *gomp_data; // gomp reduction data
2389 } kmp_taskgroup_t;
2390 
2391 // forward declarations
2392 typedef union kmp_depnode kmp_depnode_t;
2393 typedef struct kmp_depnode_list kmp_depnode_list_t;
2394 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2395 
2396 // macros for checking dep flag as an integer
2397 #define KMP_DEP_IN 0x1
2398 #define KMP_DEP_OUT 0x2
2399 #define KMP_DEP_INOUT 0x3
2400 #define KMP_DEP_MTX 0x4
2401 #define KMP_DEP_SET 0x8
2402 #define KMP_DEP_ALL 0x80
2403 // Compiler sends us this info:
2404 typedef struct kmp_depend_info {
2405  kmp_intptr_t base_addr;
2406  size_t len;
2407  union {
2408  kmp_uint8 flag; // flag as an unsigned char
2409  struct { // flag as a set of 8 bits
2410  unsigned in : 1;
2411  unsigned out : 1;
2412  unsigned mtx : 1;
2413  unsigned set : 1;
2414  unsigned unused : 3;
2415  unsigned all : 1;
2416  } flags;
2417  };
2418 } kmp_depend_info_t;
2419 
2420 // Internal structures to work with task dependencies:
2421 struct kmp_depnode_list {
2422  kmp_depnode_t *node;
2423  kmp_depnode_list_t *next;
2424 };
2425 
2426 // Max number of mutexinoutset dependencies per node
2427 #define MAX_MTX_DEPS 4
2428 
2429 typedef struct kmp_base_depnode {
2430  kmp_depnode_list_t *successors; /* used under lock */
2431  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2432  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2433  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2434  kmp_lock_t lock; /* guards shared fields: task, successors */
2435 #if KMP_SUPPORT_GRAPH_OUTPUT
2436  kmp_uint32 id;
2437 #endif
2438  std::atomic<kmp_int32> npredecessors;
2439  std::atomic<kmp_int32> nrefs;
2440 } kmp_base_depnode_t;
2441 
2442 union KMP_ALIGN_CACHE kmp_depnode {
2443  double dn_align; /* use worst case alignment */
2444  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2445  kmp_base_depnode_t dn;
2446 };
2447 
2448 struct kmp_dephash_entry {
2449  kmp_intptr_t addr;
2450  kmp_depnode_t *last_out;
2451  kmp_depnode_list_t *last_set;
2452  kmp_depnode_list_t *prev_set;
2453  kmp_uint8 last_flag;
2454  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2455  kmp_dephash_entry_t *next_in_bucket;
2456 };
2457 
2458 typedef struct kmp_dephash {
2459  kmp_dephash_entry_t **buckets;
2460  size_t size;
2461  kmp_depnode_t *last_all;
2462  size_t generation;
2463  kmp_uint32 nelements;
2464  kmp_uint32 nconflicts;
2465 } kmp_dephash_t;
2466 
2467 typedef struct kmp_task_affinity_info {
2468  kmp_intptr_t base_addr;
2469  size_t len;
2470  struct {
2471  bool flag1 : 1;
2472  bool flag2 : 1;
2473  kmp_int32 reserved : 30;
2474  } flags;
2475 } kmp_task_affinity_info_t;
2476 
2477 typedef enum kmp_event_type_t {
2478  KMP_EVENT_UNINITIALIZED = 0,
2479  KMP_EVENT_ALLOW_COMPLETION = 1
2480 } kmp_event_type_t;
2481 
2482 typedef struct {
2483  kmp_event_type_t type;
2484  kmp_tas_lock_t lock;
2485  union {
2486  kmp_task_t *task;
2487  } ed;
2488 } kmp_event_t;
2489 
2490 #if OMPX_TASKGRAPH
2491 // Initial number of allocated nodes while recording
2492 #define INIT_MAPSIZE 50
2493 
2494 typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2495  unsigned nowait : 1;
2496  unsigned re_record : 1;
2497  unsigned reserved : 30;
2498 } kmp_taskgraph_flags_t;
2499 
2501 typedef struct kmp_node_info {
2502  kmp_task_t *task; // Pointer to the actual task
2503  kmp_int32 *successors; // Array of the succesors ids
2504  kmp_int32 nsuccessors; // Number of succesors of the node
2505  std::atomic<kmp_int32>
2506  npredecessors_counter; // Number of predessors on the fly
2507  kmp_int32 npredecessors; // Total number of predecessors
2508  kmp_int32 successors_size; // Number of allocated succesors ids
2509  kmp_taskdata_t *parent_task; // Parent implicit task
2510 } kmp_node_info_t;
2511 
2513 typedef enum kmp_tdg_status {
2514  KMP_TDG_NONE = 0,
2515  KMP_TDG_RECORDING = 1,
2516  KMP_TDG_READY = 2
2517 } kmp_tdg_status_t;
2518 
2520 typedef struct kmp_tdg_info {
2521  kmp_int32 tdg_id; // Unique idenfifier of the TDG
2522  kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2523  kmp_int32 map_size; // Number of allocated TDG nodes
2524  kmp_int32 num_roots; // Number of roots tasks int the TDG
2525  kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2526  kmp_node_info_t *record_map; // Array of TDG nodes
2527  kmp_tdg_status_t tdg_status =
2528  KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2529  std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2530  kmp_bootstrap_lock_t
2531  graph_lock; // Protect graph attributes when updated via taskloop_recur
2532  // Taskloop reduction related
2533  void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2534  // __kmpc_taskred_init
2535  kmp_int32 rec_num_taskred;
2536 } kmp_tdg_info_t;
2537 
2538 extern int __kmp_tdg_dot;
2539 extern kmp_int32 __kmp_max_tdgs;
2540 extern kmp_tdg_info_t **__kmp_global_tdgs;
2541 extern kmp_int32 __kmp_curr_tdg_idx;
2542 extern kmp_int32 __kmp_successors_size;
2543 extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2544 extern kmp_int32 __kmp_num_tdg;
2545 #endif
2546 
2547 #ifdef BUILD_TIED_TASK_STACK
2548 
2549 /* Tied Task stack definitions */
2550 typedef struct kmp_stack_block {
2551  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2552  struct kmp_stack_block *sb_next;
2553  struct kmp_stack_block *sb_prev;
2554 } kmp_stack_block_t;
2555 
2556 typedef struct kmp_task_stack {
2557  kmp_stack_block_t ts_first_block; // first block of stack entries
2558  kmp_taskdata_t **ts_top; // pointer to the top of stack
2559  kmp_int32 ts_entries; // number of entries on the stack
2560 } kmp_task_stack_t;
2561 
2562 #endif // BUILD_TIED_TASK_STACK
2563 
2564 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2565  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2566  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2567  unsigned final : 1; /* task is final(1) so execute immediately */
2568  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2569  code path */
2570  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2571  invoke destructors from the runtime */
2572  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2573  context of the RTL) */
2574  unsigned priority_specified : 1; /* set if the compiler provides priority
2575  setting for the task */
2576  unsigned detachable : 1; /* 1 == can detach */
2577  unsigned hidden_helper : 1; /* 1 == hidden helper task */
2578  unsigned reserved : 8; /* reserved for compiler use */
2579 
2580  /* Library flags */ /* Total library flags must be 16 bits */
2581  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2582  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2583  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2584  // (1) or may be deferred (0)
2585  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2586  // (0) [>= 2 threads]
2587  /* If either team_serial or tasking_ser is set, task team may be NULL */
2588  /* Task State Flags: */
2589  unsigned started : 1; /* 1==started, 0==not started */
2590  unsigned executing : 1; /* 1==executing, 0==not executing */
2591  unsigned complete : 1; /* 1==complete, 0==not complete */
2592  unsigned freed : 1; /* 1==freed, 0==allocated */
2593  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2594 #if OMPX_TASKGRAPH
2595  unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2596  unsigned reserved31 : 6; /* reserved for library use */
2597 #else
2598  unsigned reserved31 : 7; /* reserved for library use */
2599 #endif
2600 
2601 } kmp_tasking_flags_t;
2602 
2603 typedef struct kmp_target_data {
2604  void *async_handle; // libomptarget async handle for task completion query
2605 } kmp_target_data_t;
2606 
2607 struct kmp_taskdata { /* aligned during dynamic allocation */
2608  kmp_int32 td_task_id; /* id, assigned by debugger */
2609  kmp_tasking_flags_t td_flags; /* task flags */
2610  kmp_team_t *td_team; /* team for this task */
2611  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2612  /* Currently not used except for perhaps IDB */
2613  kmp_taskdata_t *td_parent; /* parent task */
2614  kmp_int32 td_level; /* task nesting level */
2615  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2616  ident_t *td_ident; /* task identifier */
2617  // Taskwait data.
2618  ident_t *td_taskwait_ident;
2619  kmp_uint32 td_taskwait_counter;
2620  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2621  KMP_ALIGN_CACHE kmp_internal_control_t
2622  td_icvs; /* Internal control variables for the task */
2623  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2624  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2625  deallocated */
2626  std::atomic<kmp_int32>
2627  td_incomplete_child_tasks; /* Child tasks not yet complete */
2628  kmp_taskgroup_t
2629  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2630  kmp_dephash_t
2631  *td_dephash; // Dependencies for children tasks are tracked from here
2632  kmp_depnode_t
2633  *td_depnode; // Pointer to graph node if this task has dependencies
2634  kmp_task_team_t *td_task_team;
2635  size_t td_size_alloc; // Size of task structure, including shareds etc.
2636 #if defined(KMP_GOMP_COMPAT)
2637  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2638  kmp_int32 td_size_loop_bounds;
2639 #endif
2640  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2641 #if defined(KMP_GOMP_COMPAT)
2642  // GOMP sends in a copy function for copy constructors
2643  void (*td_copy_func)(void *, void *);
2644 #endif
2645  kmp_event_t td_allow_completion_event;
2646 #if OMPT_SUPPORT
2647  ompt_task_info_t ompt_task_info;
2648 #endif
2649 #if OMPX_TASKGRAPH
2650  bool is_taskgraph = 0; // whether the task is within a TDG
2651  kmp_tdg_info_t *tdg; // used to associate task with a TDG
2652 #endif
2653  kmp_target_data_t td_target_data;
2654 }; // struct kmp_taskdata
2655 
2656 // Make sure padding above worked
2657 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2658 
2659 // Data for task team but per thread
2660 typedef struct kmp_base_thread_data {
2661  kmp_info_p *td_thr; // Pointer back to thread info
2662  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2663  // queued?
2664  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2665  kmp_taskdata_t *
2666  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2667  kmp_int32 td_deque_size; // Size of deck
2668  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2669  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2670  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2671  // GEH: shouldn't this be volatile since used in while-spin?
2672  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2673 #ifdef BUILD_TIED_TASK_STACK
2674  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2675 // scheduling constraint
2676 #endif // BUILD_TIED_TASK_STACK
2677 } kmp_base_thread_data_t;
2678 
2679 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2680 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2681 
2682 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2683 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2684 
2685 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2686  kmp_base_thread_data_t td;
2687  double td_align; /* use worst case alignment */
2688  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2689 } kmp_thread_data_t;
2690 
2691 typedef struct kmp_task_pri {
2692  kmp_thread_data_t td;
2693  kmp_int32 priority;
2694  kmp_task_pri *next;
2695 } kmp_task_pri_t;
2696 
2697 // Data for task teams which are used when tasking is enabled for the team
2698 typedef struct kmp_base_task_team {
2699  kmp_bootstrap_lock_t
2700  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2701  /* must be bootstrap lock since used at library shutdown*/
2702 
2703  // TODO: check performance vs kmp_tas_lock_t
2704  kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2705  kmp_task_pri_t *tt_task_pri_list;
2706 
2707  kmp_task_team_t *tt_next; /* For linking the task team free list */
2708  kmp_thread_data_t
2709  *tt_threads_data; /* Array of per-thread structures for task team */
2710  /* Data survives task team deallocation */
2711  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2712  executing this team? */
2713  /* TRUE means tt_threads_data is set up and initialized */
2714  kmp_int32 tt_nproc; /* #threads in team */
2715  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2716  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2717  kmp_int32 tt_untied_task_encountered;
2718  std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2719  // There is hidden helper thread encountered in this task team so that we must
2720  // wait when waiting on task team
2721  kmp_int32 tt_hidden_helper_task_encountered;
2722 
2723  KMP_ALIGN_CACHE
2724  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2725 
2726  KMP_ALIGN_CACHE
2727  volatile kmp_uint32
2728  tt_active; /* is the team still actively executing tasks */
2729 } kmp_base_task_team_t;
2730 
2731 union KMP_ALIGN_CACHE kmp_task_team {
2732  kmp_base_task_team_t tt;
2733  double tt_align; /* use worst case alignment */
2734  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2735 };
2736 
2737 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2738 // Free lists keep same-size free memory slots for fast memory allocation
2739 // routines
2740 typedef struct kmp_free_list {
2741  void *th_free_list_self; // Self-allocated tasks free list
2742  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2743  // threads
2744  void *th_free_list_other; // Non-self free list (to be returned to owner's
2745  // sync list)
2746 } kmp_free_list_t;
2747 #endif
2748 #if KMP_NESTED_HOT_TEAMS
2749 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2750 // are not put in teams pool, and they don't put threads in threads pool.
2751 typedef struct kmp_hot_team_ptr {
2752  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2753  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2754 } kmp_hot_team_ptr_t;
2755 #endif
2756 typedef struct kmp_teams_size {
2757  kmp_int32 nteams; // number of teams in a league
2758  kmp_int32 nth; // number of threads in each team of the league
2759 } kmp_teams_size_t;
2760 
2761 // This struct stores a thread that acts as a "root" for a contention
2762 // group. Contention groups are rooted at kmp_root threads, but also at
2763 // each primary thread of each team created in the teams construct.
2764 // This struct therefore also stores a thread_limit associated with
2765 // that contention group, and a counter to track the number of threads
2766 // active in that contention group. Each thread has a list of these: CG
2767 // root threads have an entry in their list in which cg_root refers to
2768 // the thread itself, whereas other workers in the CG will have a
2769 // single entry where cg_root is same as the entry containing their CG
2770 // root. When a thread encounters a teams construct, it will add a new
2771 // entry to the front of its list, because it now roots a new CG.
2772 typedef struct kmp_cg_root {
2773  kmp_info_p *cg_root; // "root" thread for a contention group
2774  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2775  // thread_limit clause for teams primary threads
2776  kmp_int32 cg_thread_limit;
2777  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2778  struct kmp_cg_root *up; // pointer to higher level CG root in list
2779 } kmp_cg_root_t;
2780 
2781 // OpenMP thread data structures
2782 
2783 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2784  /* Start with the readonly data which is cache aligned and padded. This is
2785  written before the thread starts working by the primary thread. Uber
2786  masters may update themselves later. Usage does not consider serialized
2787  regions. */
2788  kmp_desc_t th_info;
2789  kmp_team_p *th_team; /* team we belong to */
2790  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2791  kmp_info_p *th_next_pool; /* next available thread in the pool */
2792  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2793  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2794 
2795  /* The following are cached from the team info structure */
2796  /* TODO use these in more places as determined to be needed via profiling */
2797  int th_team_nproc; /* number of threads in a team */
2798  kmp_info_p *th_team_master; /* the team's primary thread */
2799  int th_team_serialized; /* team is serialized */
2800  microtask_t th_teams_microtask; /* save entry address for teams construct */
2801  int th_teams_level; /* save initial level of teams construct */
2802 /* it is 0 on device but may be any on host */
2803 
2804 /* The blocktime info is copied from the team struct to the thread struct */
2805 /* at the start of a barrier, and the values stored in the team are used */
2806 /* at points in the code where the team struct is no longer guaranteed */
2807 /* to exist (from the POV of worker threads). */
2808 #if KMP_USE_MONITOR
2809  int th_team_bt_intervals;
2810  int th_team_bt_set;
2811 #else
2812  kmp_uint64 th_team_bt_intervals;
2813 #endif
2814 
2815 #if KMP_AFFINITY_SUPPORTED
2816  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2817  kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2818  kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2819 #endif
2820  omp_allocator_handle_t th_def_allocator; /* default allocator */
2821  /* The data set by the primary thread at reinit, then R/W by the worker */
2822  KMP_ALIGN_CACHE int
2823  th_set_nproc; /* if > 0, then only use this request for the next fork */
2824 #if KMP_NESTED_HOT_TEAMS
2825  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2826 #endif
2827  kmp_proc_bind_t
2828  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2829  kmp_teams_size_t
2830  th_teams_size; /* number of teams/threads in teams construct */
2831 #if KMP_AFFINITY_SUPPORTED
2832  int th_current_place; /* place currently bound to */
2833  int th_new_place; /* place to bind to in par reg */
2834  int th_first_place; /* first place in partition */
2835  int th_last_place; /* last place in partition */
2836 #endif
2837  int th_prev_level; /* previous level for affinity format */
2838  int th_prev_num_threads; /* previous num_threads for affinity format */
2839 #if USE_ITT_BUILD
2840  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2841  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2842  kmp_uint64 th_frame_time; /* frame timestamp */
2843 #endif /* USE_ITT_BUILD */
2844  kmp_local_t th_local;
2845  struct private_common *th_pri_head;
2846 
2847  /* Now the data only used by the worker (after initial allocation) */
2848  /* TODO the first serial team should actually be stored in the info_t
2849  structure. this will help reduce initial allocation overhead */
2850  KMP_ALIGN_CACHE kmp_team_p
2851  *th_serial_team; /*serialized team held in reserve*/
2852 
2853 #if OMPT_SUPPORT
2854  ompt_thread_info_t ompt_thread_info;
2855 #endif
2856 
2857  /* The following are also read by the primary thread during reinit */
2858  struct common_table *th_pri_common;
2859 
2860  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2861  /* while awaiting queuing lock acquire */
2862 
2863  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2864  flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2865 
2866  ident_t *th_ident;
2867  unsigned th_x; // Random number generator data
2868  unsigned th_a; // Random number generator data
2869 
2870  /* Tasking-related data for the thread */
2871  kmp_task_team_t *th_task_team; // Task team struct
2872  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2873  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2874  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2875  // at nested levels
2876  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2877  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2878  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2879  // tasking, thus safe to reap
2880 
2881  /* More stuff for keeping track of active/sleeping threads (this part is
2882  written by the worker thread) */
2883  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2884  int th_active; // ! sleeping; 32 bits for TCR/TCW
2885  std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2886  // 0 = not used in team; 1 = used in team;
2887  // 2 = transitioning to not used in team; 3 = transitioning to used in team
2888  struct cons_header *th_cons; // used for consistency check
2889 #if KMP_USE_HIER_SCHED
2890  // used for hierarchical scheduling
2891  kmp_hier_private_bdata_t *th_hier_bar_data;
2892 #endif
2893 
2894  /* Add the syncronizing data which is cache aligned and padded. */
2895  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2896 
2897  KMP_ALIGN_CACHE volatile kmp_int32
2898  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2899 
2900 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2901 #define NUM_LISTS 4
2902  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2903 // allocation routines
2904 #endif
2905 
2906 #if KMP_OS_WINDOWS
2907  kmp_win32_cond_t th_suspend_cv;
2908  kmp_win32_mutex_t th_suspend_mx;
2909  std::atomic<int> th_suspend_init;
2910 #endif
2911 #if KMP_OS_UNIX
2912  kmp_cond_align_t th_suspend_cv;
2913  kmp_mutex_align_t th_suspend_mx;
2914  std::atomic<int> th_suspend_init_count;
2915 #endif
2916 
2917 #if USE_ITT_BUILD
2918  kmp_itt_mark_t th_itt_mark_single;
2919 // alignment ???
2920 #endif /* USE_ITT_BUILD */
2921 #if KMP_STATS_ENABLED
2922  kmp_stats_list *th_stats;
2923 #endif
2924 #if KMP_OS_UNIX
2925  std::atomic<bool> th_blocking;
2926 #endif
2927  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2928 } kmp_base_info_t;
2929 
2930 typedef union KMP_ALIGN_CACHE kmp_info {
2931  double th_align; /* use worst case alignment */
2932  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2933  kmp_base_info_t th;
2934 } kmp_info_t;
2935 
2936 // OpenMP thread team data structures
2937 
2938 typedef struct kmp_base_data {
2939  volatile kmp_uint32 t_value;
2940 } kmp_base_data_t;
2941 
2942 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2943  double dt_align; /* use worst case alignment */
2944  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2945  kmp_base_data_t dt;
2946 } kmp_sleep_team_t;
2947 
2948 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2949  double dt_align; /* use worst case alignment */
2950  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2951  kmp_base_data_t dt;
2952 } kmp_ordered_team_t;
2953 
2954 typedef int (*launch_t)(int gtid);
2955 
2956 /* Minimum number of ARGV entries to malloc if necessary */
2957 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2958 
2959 // Set up how many argv pointers will fit in cache lines containing
2960 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2961 // larger value for more space between the primary write/worker read section and
2962 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2963 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2964 #define KMP_INLINE_ARGV_BYTES \
2965  (4 * CACHE_LINE - \
2966  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2967  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2968  CACHE_LINE))
2969 #else
2970 #define KMP_INLINE_ARGV_BYTES \
2971  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2972 #endif
2973 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2974 
2975 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2976  // Synchronization Data
2977  // ---------------------------------------------------------------------------
2978  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2979  kmp_balign_team_t t_bar[bs_last_barrier];
2980  std::atomic<int> t_construct; // count of single directive encountered by team
2981  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2982 
2983  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2984  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2985  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2986 
2987  // Primary thread only
2988  // ---------------------------------------------------------------------------
2989  KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2990  int t_master_this_cons; // "this_construct" single counter of primary thread
2991  // in parent team
2992  ident_t *t_ident; // if volatile, have to change too much other crud to
2993  // volatile too
2994  kmp_team_p *t_parent; // parent team
2995  kmp_team_p *t_next_pool; // next free team in the team pool
2996  kmp_disp_t *t_dispatch; // thread's dispatch data
2997  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2998  kmp_proc_bind_t t_proc_bind; // bind type for par region
2999 #if USE_ITT_BUILD
3000  kmp_uint64 t_region_time; // region begin timestamp
3001 #endif /* USE_ITT_BUILD */
3002 
3003  // Primary thread write, workers read
3004  // --------------------------------------------------------------------------
3005  KMP_ALIGN_CACHE void **t_argv;
3006  int t_argc;
3007  int t_nproc; // number of threads in team
3008  microtask_t t_pkfn;
3009  launch_t t_invoke; // procedure to launch the microtask
3010 
3011 #if OMPT_SUPPORT
3012  ompt_team_info_t ompt_team_info;
3013  ompt_lw_taskteam_t *ompt_serialized_team_info;
3014 #endif
3015 
3016 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3017  kmp_int8 t_fp_control_saved;
3018  kmp_int8 t_pad2b;
3019  kmp_int16 t_x87_fpu_control_word; // FP control regs
3020  kmp_uint32 t_mxcsr;
3021 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3022 
3023  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3024 
3025  KMP_ALIGN_CACHE kmp_info_t **t_threads;
3026  kmp_taskdata_t
3027  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3028  int t_level; // nested parallel level
3029 
3030  KMP_ALIGN_CACHE int t_max_argc;
3031  int t_max_nproc; // max threads this team can handle (dynamically expandable)
3032  int t_serialized; // levels deep of serialized teams
3033  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3034  int t_id; // team's id, assigned by debugger.
3035  int t_active_level; // nested active parallel level
3036  kmp_r_sched_t t_sched; // run-time schedule for the team
3037 #if KMP_AFFINITY_SUPPORTED
3038  int t_first_place; // first & last place in parent thread's partition.
3039  int t_last_place; // Restore these values to primary thread after par region.
3040 #endif // KMP_AFFINITY_SUPPORTED
3041  int t_display_affinity;
3042  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3043  // omp_set_num_threads() call
3044  omp_allocator_handle_t t_def_allocator; /* default allocator */
3045 
3046 // Read/write by workers as well
3047 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3048  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3049  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3050  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3051  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3052  char dummy_padding[1024];
3053 #endif
3054  // Internal control stack for additional nested teams.
3055  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3056  // for SERIALIZED teams nested 2 or more levels deep
3057  // typed flag to store request state of cancellation
3058  std::atomic<kmp_int32> t_cancel_request;
3059  int t_master_active; // save on fork, restore on join
3060  void *t_copypriv_data; // team specific pointer to copyprivate data array
3061 #if KMP_OS_WINDOWS
3062  std::atomic<kmp_uint32> t_copyin_counter;
3063 #endif
3064 #if USE_ITT_BUILD
3065  void *t_stack_id; // team specific stack stitching id (for ittnotify)
3066 #endif /* USE_ITT_BUILD */
3067  distributedBarrier *b; // Distributed barrier data associated with team
3068 } kmp_base_team_t;
3069 
3070 union KMP_ALIGN_CACHE kmp_team {
3071  kmp_base_team_t t;
3072  double t_align; /* use worst case alignment */
3073  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3074 };
3075 
3076 typedef union KMP_ALIGN_CACHE kmp_time_global {
3077  double dt_align; /* use worst case alignment */
3078  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3079  kmp_base_data_t dt;
3080 } kmp_time_global_t;
3081 
3082 typedef struct kmp_base_global {
3083  /* cache-aligned */
3084  kmp_time_global_t g_time;
3085 
3086  /* non cache-aligned */
3087  volatile int g_abort;
3088  volatile int g_done;
3089 
3090  int g_dynamic;
3091  enum dynamic_mode g_dynamic_mode;
3092 } kmp_base_global_t;
3093 
3094 typedef union KMP_ALIGN_CACHE kmp_global {
3095  kmp_base_global_t g;
3096  double g_align; /* use worst case alignment */
3097  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3098 } kmp_global_t;
3099 
3100 typedef struct kmp_base_root {
3101  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3102  // (r_in_parallel>= 0)
3103  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3104  // the synch overhead or keeping r_active
3105  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3106  // keeps a count of active parallel regions per root
3107  std::atomic<int> r_in_parallel;
3108  // GEH: This is misnamed, should be r_active_levels
3109  kmp_team_t *r_root_team;
3110  kmp_team_t *r_hot_team;
3111  kmp_info_t *r_uber_thread;
3112  kmp_lock_t r_begin_lock;
3113  volatile int r_begin;
3114  int r_blocktime; /* blocktime for this root and descendants */
3115 #if KMP_AFFINITY_SUPPORTED
3116  int r_affinity_assigned;
3117 #endif // KMP_AFFINITY_SUPPORTED
3118 } kmp_base_root_t;
3119 
3120 typedef union KMP_ALIGN_CACHE kmp_root {
3121  kmp_base_root_t r;
3122  double r_align; /* use worst case alignment */
3123  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3124 } kmp_root_t;
3125 
3126 struct fortran_inx_info {
3127  kmp_int32 data;
3128 };
3129 
3130 // This list type exists to hold old __kmp_threads arrays so that
3131 // old references to them may complete while reallocation takes place when
3132 // expanding the array. The items in this list are kept alive until library
3133 // shutdown.
3134 typedef struct kmp_old_threads_list_t {
3135  kmp_info_t **threads;
3136  struct kmp_old_threads_list_t *next;
3137 } kmp_old_threads_list_t;
3138 
3139 /* ------------------------------------------------------------------------ */
3140 
3141 extern int __kmp_settings;
3142 extern int __kmp_duplicate_library_ok;
3143 #if USE_ITT_BUILD
3144 extern int __kmp_forkjoin_frames;
3145 extern int __kmp_forkjoin_frames_mode;
3146 #endif
3147 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3148 extern int __kmp_determ_red;
3149 
3150 #ifdef KMP_DEBUG
3151 extern int kmp_a_debug;
3152 extern int kmp_b_debug;
3153 extern int kmp_c_debug;
3154 extern int kmp_d_debug;
3155 extern int kmp_e_debug;
3156 extern int kmp_f_debug;
3157 #endif /* KMP_DEBUG */
3158 
3159 /* For debug information logging using rotating buffer */
3160 #define KMP_DEBUG_BUF_LINES_INIT 512
3161 #define KMP_DEBUG_BUF_LINES_MIN 1
3162 
3163 #define KMP_DEBUG_BUF_CHARS_INIT 128
3164 #define KMP_DEBUG_BUF_CHARS_MIN 2
3165 
3166 extern int
3167  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3168 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3169 extern int
3170  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3171 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3172  entry pointer */
3173 
3174 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3175 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3176  printed in buffer so far */
3177 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3178  recommended in warnings */
3179 /* end rotating debug buffer */
3180 
3181 #ifdef KMP_DEBUG
3182 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3183 
3184 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3185 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3186 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3187 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3188 extern int __kmp_par_range_lb;
3189 extern int __kmp_par_range_ub;
3190 #endif
3191 
3192 /* For printing out dynamic storage map for threads and teams */
3193 extern int
3194  __kmp_storage_map; /* True means print storage map for threads and teams */
3195 extern int __kmp_storage_map_verbose; /* True means storage map includes
3196  placement info */
3197 extern int __kmp_storage_map_verbose_specified;
3198 
3199 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3200 extern kmp_cpuinfo_t __kmp_cpuinfo;
3201 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3202 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3203 static inline bool __kmp_is_hybrid_cpu() { return true; }
3204 #else
3205 static inline bool __kmp_is_hybrid_cpu() { return false; }
3206 #endif
3207 
3208 extern volatile int __kmp_init_serial;
3209 extern volatile int __kmp_init_gtid;
3210 extern volatile int __kmp_init_common;
3211 extern volatile int __kmp_need_register_serial;
3212 extern volatile int __kmp_init_middle;
3213 extern volatile int __kmp_init_parallel;
3214 #if KMP_USE_MONITOR
3215 extern volatile int __kmp_init_monitor;
3216 #endif
3217 extern volatile int __kmp_init_user_locks;
3218 extern volatile int __kmp_init_hidden_helper_threads;
3219 extern int __kmp_init_counter;
3220 extern int __kmp_root_counter;
3221 extern int __kmp_version;
3222 
3223 /* list of address of allocated caches for commons */
3224 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3225 
3226 /* Barrier algorithm types and options */
3227 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3228 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3229 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3230 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3231 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3232 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3233 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3234 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3235 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3236 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3237 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3238 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3239 
3240 /* Global Locks */
3241 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3242 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3243 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3244 extern kmp_bootstrap_lock_t
3245  __kmp_exit_lock; /* exit() is not always thread-safe */
3246 #if KMP_USE_MONITOR
3247 extern kmp_bootstrap_lock_t
3248  __kmp_monitor_lock; /* control monitor thread creation */
3249 #endif
3250 extern kmp_bootstrap_lock_t
3251  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3252  __kmp_threads expansion to co-exist */
3253 
3254 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3255 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3256 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3257 
3258 extern enum library_type __kmp_library;
3259 
3260 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3261 extern enum sched_type __kmp_static; /* default static scheduling method */
3262 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3263 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3264 extern int __kmp_chunk; /* default runtime chunk size */
3265 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3266 
3267 extern size_t __kmp_stksize; /* stack size per thread */
3268 #if KMP_USE_MONITOR
3269 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3270 #endif
3271 extern size_t __kmp_stkoffset; /* stack offset per thread */
3272 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3273 
3274 extern size_t
3275  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3276 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3277 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3278 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3279 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3280 extern int __kmp_generate_warnings; /* should we issue warnings? */
3281 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3282 
3283 #ifdef DEBUG_SUSPEND
3284 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3285 #endif
3286 
3287 extern kmp_int32 __kmp_use_yield;
3288 extern kmp_int32 __kmp_use_yield_exp_set;
3289 extern kmp_uint32 __kmp_yield_init;
3290 extern kmp_uint32 __kmp_yield_next;
3291 extern kmp_uint64 __kmp_pause_init;
3292 
3293 /* ------------------------------------------------------------------------- */
3294 extern int __kmp_allThreadsSpecified;
3295 
3296 extern size_t __kmp_align_alloc;
3297 /* following data protected by initialization routines */
3298 extern int __kmp_xproc; /* number of processors in the system */
3299 extern int __kmp_avail_proc; /* number of processors available to the process */
3300 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3301 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3302 // maximum total number of concurrently-existing threads on device
3303 extern int __kmp_max_nth;
3304 // maximum total number of concurrently-existing threads in a contention group
3305 extern int __kmp_cg_max_nth;
3306 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3307 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3308  __kmp_root */
3309 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3310  region a la OMP_NUM_THREADS */
3311 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3312  initialization */
3313 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3314  used (fixed) */
3315 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3316  (__kmpc_threadprivate_cached()) */
3317 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3318  blocking (env setting) */
3319 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3320 #if KMP_USE_MONITOR
3321 extern int
3322  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3323 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3324  blocking */
3325 #endif
3326 #ifdef KMP_ADJUST_BLOCKTIME
3327 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3328 #endif /* KMP_ADJUST_BLOCKTIME */
3329 #ifdef KMP_DFLT_NTH_CORES
3330 extern int __kmp_ncores; /* Total number of cores for threads placement */
3331 #endif
3332 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3333 extern int __kmp_abort_delay;
3334 
3335 extern int __kmp_need_register_atfork_specified;
3336 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3337  to install fork handler */
3338 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3339  0 - not set, will be set at runtime
3340  1 - using stack search
3341  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3342  X*) or TlsGetValue(Windows* OS))
3343  3 - static TLS (__declspec(thread) __kmp_gtid),
3344  Linux* OS .so only. */
3345 extern int
3346  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3347 #ifdef KMP_TDATA_GTID
3348 extern KMP_THREAD_LOCAL int __kmp_gtid;
3349 #endif
3350 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3351 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3352 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3353 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3354 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3355 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3356 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3357 
3358 // max_active_levels for nested parallelism enabled by default via
3359 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3360 extern int __kmp_dflt_max_active_levels;
3361 // Indicates whether value of __kmp_dflt_max_active_levels was already
3362 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3363 extern bool __kmp_dflt_max_active_levels_set;
3364 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3365  concurrent execution per team */
3366 #if KMP_NESTED_HOT_TEAMS
3367 extern int __kmp_hot_teams_mode;
3368 extern int __kmp_hot_teams_max_level;
3369 #endif
3370 
3371 #if KMP_OS_LINUX
3372 extern enum clock_function_type __kmp_clock_function;
3373 extern int __kmp_clock_function_param;
3374 #endif /* KMP_OS_LINUX */
3375 
3376 #if KMP_MIC_SUPPORTED
3377 extern enum mic_type __kmp_mic_type;
3378 #endif
3379 
3380 #ifdef USE_LOAD_BALANCE
3381 extern double __kmp_load_balance_interval; // load balance algorithm interval
3382 #endif /* USE_LOAD_BALANCE */
3383 
3384 // OpenMP 3.1 - Nested num threads array
3385 typedef struct kmp_nested_nthreads_t {
3386  int *nth;
3387  int size;
3388  int used;
3389 } kmp_nested_nthreads_t;
3390 
3391 extern kmp_nested_nthreads_t __kmp_nested_nth;
3392 
3393 #if KMP_USE_ADAPTIVE_LOCKS
3394 
3395 // Parameters for the speculative lock backoff system.
3396 struct kmp_adaptive_backoff_params_t {
3397  // Number of soft retries before it counts as a hard retry.
3398  kmp_uint32 max_soft_retries;
3399  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3400  // the right
3401  kmp_uint32 max_badness;
3402 };
3403 
3404 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3405 
3406 #if KMP_DEBUG_ADAPTIVE_LOCKS
3407 extern const char *__kmp_speculative_statsfile;
3408 #endif
3409 
3410 #endif // KMP_USE_ADAPTIVE_LOCKS
3411 
3412 extern int __kmp_display_env; /* TRUE or FALSE */
3413 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3414 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3415 extern int __kmp_nteams;
3416 extern int __kmp_teams_thread_limit;
3417 
3418 /* ------------------------------------------------------------------------- */
3419 
3420 /* the following are protected by the fork/join lock */
3421 /* write: lock read: anytime */
3422 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3423 /* Holds old arrays of __kmp_threads until library shutdown */
3424 extern kmp_old_threads_list_t *__kmp_old_threads_list;
3425 /* read/write: lock */
3426 extern volatile kmp_team_t *__kmp_team_pool;
3427 extern volatile kmp_info_t *__kmp_thread_pool;
3428 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3429 
3430 // total num threads reachable from some root thread including all root threads
3431 extern volatile int __kmp_nth;
3432 /* total number of threads reachable from some root thread including all root
3433  threads, and those in the thread pool */
3434 extern volatile int __kmp_all_nth;
3435 extern std::atomic<int> __kmp_thread_pool_active_nth;
3436 
3437 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3438 /* end data protected by fork/join lock */
3439 /* ------------------------------------------------------------------------- */
3440 
3441 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3442 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3443 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3444 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3445 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3446 
3447 // AT: Which way is correct?
3448 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3449 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3450 #define __kmp_get_team_num_threads(gtid) \
3451  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3452 
3453 static inline bool KMP_UBER_GTID(int gtid) {
3454  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3455  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3456  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3457  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3458 }
3459 
3460 static inline int __kmp_tid_from_gtid(int gtid) {
3461  KMP_DEBUG_ASSERT(gtid >= 0);
3462  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3463 }
3464 
3465 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3466  KMP_DEBUG_ASSERT(tid >= 0 && team);
3467  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3468 }
3469 
3470 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3471  KMP_DEBUG_ASSERT(thr);
3472  return thr->th.th_info.ds.ds_gtid;
3473 }
3474 
3475 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3476  KMP_DEBUG_ASSERT(gtid >= 0);
3477  return __kmp_threads[gtid];
3478 }
3479 
3480 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3481  KMP_DEBUG_ASSERT(gtid >= 0);
3482  return __kmp_threads[gtid]->th.th_team;
3483 }
3484 
3485 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3486  if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3487  KMP_FATAL(ThreadIdentInvalid);
3488 }
3489 
3490 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3491 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3492 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3493 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3494 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3495 #endif
3496 
3497 #if KMP_HAVE_UMWAIT
3498 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3499 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3500 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3501 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3502 #endif
3503 
3504 /* ------------------------------------------------------------------------- */
3505 
3506 extern kmp_global_t __kmp_global; /* global status */
3507 
3508 extern kmp_info_t __kmp_monitor;
3509 // For Debugging Support Library
3510 extern std::atomic<kmp_int32> __kmp_team_counter;
3511 // For Debugging Support Library
3512 extern std::atomic<kmp_int32> __kmp_task_counter;
3513 
3514 #if USE_DEBUGGER
3515 #define _KMP_GEN_ID(counter) \
3516  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3517 #else
3518 #define _KMP_GEN_ID(counter) (~0)
3519 #endif /* USE_DEBUGGER */
3520 
3521 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3522 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3523 
3524 /* ------------------------------------------------------------------------ */
3525 
3526 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3527  size_t size, char const *format, ...);
3528 
3529 extern void __kmp_serial_initialize(void);
3530 extern void __kmp_middle_initialize(void);
3531 extern void __kmp_parallel_initialize(void);
3532 
3533 extern void __kmp_internal_begin(void);
3534 extern void __kmp_internal_end_library(int gtid);
3535 extern void __kmp_internal_end_thread(int gtid);
3536 extern void __kmp_internal_end_atexit(void);
3537 extern void __kmp_internal_end_dtor(void);
3538 extern void __kmp_internal_end_dest(void *);
3539 
3540 extern int __kmp_register_root(int initial_thread);
3541 extern void __kmp_unregister_root(int gtid);
3542 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3543 
3544 extern int __kmp_ignore_mppbeg(void);
3545 extern int __kmp_ignore_mppend(void);
3546 
3547 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3548 extern void __kmp_exit_single(int gtid);
3549 
3550 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3551 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3552 
3553 #ifdef USE_LOAD_BALANCE
3554 extern int __kmp_get_load_balance(int);
3555 #endif
3556 
3557 extern int __kmp_get_global_thread_id(void);
3558 extern int __kmp_get_global_thread_id_reg(void);
3559 extern void __kmp_exit_thread(int exit_status);
3560 extern void __kmp_abort(char const *format, ...);
3561 extern void __kmp_abort_thread(void);
3562 KMP_NORETURN extern void __kmp_abort_process(void);
3563 extern void __kmp_warn(char const *format, ...);
3564 
3565 extern void __kmp_set_num_threads(int new_nth, int gtid);
3566 
3567 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3568 // registered.
3569 static inline kmp_info_t *__kmp_entry_thread() {
3570  int gtid = __kmp_entry_gtid();
3571 
3572  return __kmp_threads[gtid];
3573 }
3574 
3575 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3576 extern int __kmp_get_max_active_levels(int gtid);
3577 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3578 extern int __kmp_get_team_size(int gtid, int level);
3579 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3580 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3581 
3582 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3583 extern void __kmp_init_random(kmp_info_t *thread);
3584 
3585 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3586 extern void __kmp_adjust_num_threads(int new_nproc);
3587 extern void __kmp_check_stksize(size_t *val);
3588 
3589 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3590 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3591 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3592 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3593 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3594 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3595 
3596 #if USE_FAST_MEMORY
3597 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3598  size_t size KMP_SRC_LOC_DECL);
3599 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3600 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3601 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3602 #define __kmp_fast_allocate(this_thr, size) \
3603  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3604 #define __kmp_fast_free(this_thr, ptr) \
3605  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3606 #endif
3607 
3608 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3609 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3610  size_t elsize KMP_SRC_LOC_DECL);
3611 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3612  size_t size KMP_SRC_LOC_DECL);
3613 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3614 #define __kmp_thread_malloc(th, size) \
3615  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3616 #define __kmp_thread_calloc(th, nelem, elsize) \
3617  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3618 #define __kmp_thread_realloc(th, ptr, size) \
3619  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3620 #define __kmp_thread_free(th, ptr) \
3621  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3622 
3623 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3624 
3625 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3626  kmp_proc_bind_t proc_bind);
3627 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3628  int num_threads);
3629 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3630  int num_teams_ub, int num_threads);
3631 
3632 extern void __kmp_yield();
3633 
3634 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3635  enum sched_type schedule, kmp_int32 lb,
3636  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3637 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3638  enum sched_type schedule, kmp_uint32 lb,
3639  kmp_uint32 ub, kmp_int32 st,
3640  kmp_int32 chunk);
3641 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3642  enum sched_type schedule, kmp_int64 lb,
3643  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3644 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3645  enum sched_type schedule, kmp_uint64 lb,
3646  kmp_uint64 ub, kmp_int64 st,
3647  kmp_int64 chunk);
3648 
3649 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3650  kmp_int32 *p_last, kmp_int32 *p_lb,
3651  kmp_int32 *p_ub, kmp_int32 *p_st);
3652 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3653  kmp_int32 *p_last, kmp_uint32 *p_lb,
3654  kmp_uint32 *p_ub, kmp_int32 *p_st);
3655 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3656  kmp_int32 *p_last, kmp_int64 *p_lb,
3657  kmp_int64 *p_ub, kmp_int64 *p_st);
3658 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3659  kmp_int32 *p_last, kmp_uint64 *p_lb,
3660  kmp_uint64 *p_ub, kmp_int64 *p_st);
3661 
3662 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3663 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3664 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3665 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3666 
3667 #ifdef KMP_GOMP_COMPAT
3668 
3669 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3670  enum sched_type schedule, kmp_int32 lb,
3671  kmp_int32 ub, kmp_int32 st,
3672  kmp_int32 chunk, int push_ws);
3673 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3674  enum sched_type schedule, kmp_uint32 lb,
3675  kmp_uint32 ub, kmp_int32 st,
3676  kmp_int32 chunk, int push_ws);
3677 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3678  enum sched_type schedule, kmp_int64 lb,
3679  kmp_int64 ub, kmp_int64 st,
3680  kmp_int64 chunk, int push_ws);
3681 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3682  enum sched_type schedule, kmp_uint64 lb,
3683  kmp_uint64 ub, kmp_int64 st,
3684  kmp_int64 chunk, int push_ws);
3685 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3686 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3687 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3688 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3689 
3690 #endif /* KMP_GOMP_COMPAT */
3691 
3692 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3693 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3694 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3695 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3696 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3697 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3698  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3699  void *obj);
3700 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3701  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3702 
3703 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3704  int final_spin
3705 #if USE_ITT_BUILD
3706  ,
3707  void *itt_sync_obj
3708 #endif
3709 );
3710 extern void __kmp_release_64(kmp_flag_64<> *flag);
3711 
3712 extern void __kmp_infinite_loop(void);
3713 
3714 extern void __kmp_cleanup(void);
3715 
3716 #if KMP_HANDLE_SIGNALS
3717 extern int __kmp_handle_signals;
3718 extern void __kmp_install_signals(int parallel_init);
3719 extern void __kmp_remove_signals(void);
3720 #endif
3721 
3722 extern void __kmp_clear_system_time(void);
3723 extern void __kmp_read_system_time(double *delta);
3724 
3725 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3726 
3727 extern void __kmp_expand_host_name(char *buffer, size_t size);
3728 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3729 
3730 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3731 extern void
3732 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3733 #endif
3734 
3735 extern void
3736 __kmp_runtime_initialize(void); /* machine specific initialization */
3737 extern void __kmp_runtime_destroy(void);
3738 
3739 #if KMP_AFFINITY_SUPPORTED
3740 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3741  kmp_affin_mask_t *mask);
3742 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3743  kmp_affin_mask_t *mask);
3744 extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3745 extern void __kmp_affinity_uninitialize(void);
3746 extern void __kmp_affinity_set_init_mask(
3747  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3748 extern void __kmp_affinity_set_place(int gtid);
3749 extern void __kmp_affinity_determine_capable(const char *env_var);
3750 extern int __kmp_aux_set_affinity(void **mask);
3751 extern int __kmp_aux_get_affinity(void **mask);
3752 extern int __kmp_aux_get_affinity_max_proc();
3753 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3754 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3755 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3756 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3757 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3758 extern int kmp_set_thread_affinity_mask_initial(void);
3759 #endif
3760 static inline void __kmp_assign_root_init_mask() {
3761  int gtid = __kmp_entry_gtid();
3762  kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3763  if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3764  __kmp_affinity_set_init_mask(gtid, TRUE);
3765  r->r.r_affinity_assigned = TRUE;
3766  }
3767 }
3768 static inline void __kmp_reset_root_init_mask(int gtid) {
3769  if (!KMP_AFFINITY_CAPABLE())
3770  return;
3771  kmp_info_t *th = __kmp_threads[gtid];
3772  kmp_root_t *r = th->th.th_root;
3773  if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3774  __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3775  KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3776  r->r.r_affinity_assigned = FALSE;
3777  }
3778 }
3779 #else /* KMP_AFFINITY_SUPPORTED */
3780 #define __kmp_assign_root_init_mask() /* Nothing */
3781 static inline void __kmp_reset_root_init_mask(int gtid) {}
3782 #endif /* KMP_AFFINITY_SUPPORTED */
3783 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3784 // format string is for affinity, so platforms that do not support
3785 // affinity can still use the other fields, e.g., %n for num_threads
3786 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3787  kmp_str_buf_t *buffer);
3788 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3789 
3790 extern void __kmp_cleanup_hierarchy();
3791 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3792 
3793 #if KMP_USE_FUTEX
3794 
3795 extern int __kmp_futex_determine_capable(void);
3796 
3797 #endif // KMP_USE_FUTEX
3798 
3799 extern void __kmp_gtid_set_specific(int gtid);
3800 extern int __kmp_gtid_get_specific(void);
3801 
3802 extern double __kmp_read_cpu_time(void);
3803 
3804 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3805 
3806 #if KMP_USE_MONITOR
3807 extern void __kmp_create_monitor(kmp_info_t *th);
3808 #endif
3809 
3810 extern void *__kmp_launch_thread(kmp_info_t *thr);
3811 
3812 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3813 
3814 #if KMP_OS_WINDOWS
3815 extern int __kmp_still_running(kmp_info_t *th);
3816 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3817 extern void __kmp_free_handle(kmp_thread_t tHandle);
3818 #endif
3819 
3820 #if KMP_USE_MONITOR
3821 extern void __kmp_reap_monitor(kmp_info_t *th);
3822 #endif
3823 extern void __kmp_reap_worker(kmp_info_t *th);
3824 extern void __kmp_terminate_thread(int gtid);
3825 
3826 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3827 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3828 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3829 
3830 extern void __kmp_elapsed(double *);
3831 extern void __kmp_elapsed_tick(double *);
3832 
3833 extern void __kmp_enable(int old_state);
3834 extern void __kmp_disable(int *old_state);
3835 
3836 extern void __kmp_thread_sleep(int millis);
3837 
3838 extern void __kmp_common_initialize(void);
3839 extern void __kmp_common_destroy(void);
3840 extern void __kmp_common_destroy_gtid(int gtid);
3841 
3842 #if KMP_OS_UNIX
3843 extern void __kmp_register_atfork(void);
3844 #endif
3845 extern void __kmp_suspend_initialize(void);
3846 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3847 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3848 
3849 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3850  int tid);
3851 extern kmp_team_t *
3852 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3853 #if OMPT_SUPPORT
3854  ompt_data_t ompt_parallel_data,
3855 #endif
3856  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3857  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3858 extern void __kmp_free_thread(kmp_info_t *);
3859 extern void __kmp_free_team(kmp_root_t *,
3860  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3861 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3862 
3863 /* ------------------------------------------------------------------------ */
3864 
3865 extern void __kmp_initialize_bget(kmp_info_t *th);
3866 extern void __kmp_finalize_bget(kmp_info_t *th);
3867 
3868 KMP_EXPORT void *kmpc_malloc(size_t size);
3869 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3870 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3871 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3872 KMP_EXPORT void kmpc_free(void *ptr);
3873 
3874 /* declarations for internal use */
3875 
3876 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3877  size_t reduce_size, void *reduce_data,
3878  void (*reduce)(void *, void *));
3879 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3880 extern int __kmp_barrier_gomp_cancel(int gtid);
3881 
3886 enum fork_context_e {
3887  fork_context_gnu,
3889  fork_context_intel,
3890  fork_context_last
3891 };
3892 extern int __kmp_fork_call(ident_t *loc, int gtid,
3893  enum fork_context_e fork_context, kmp_int32 argc,
3894  microtask_t microtask, launch_t invoker,
3895  kmp_va_list ap);
3896 
3897 extern void __kmp_join_call(ident_t *loc, int gtid
3898 #if OMPT_SUPPORT
3899  ,
3900  enum fork_context_e fork_context
3901 #endif
3902  ,
3903  int exit_teams = 0);
3904 
3905 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3906 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3907 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3908 extern int __kmp_invoke_task_func(int gtid);
3909 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3910  kmp_info_t *this_thr,
3911  kmp_team_t *team);
3912 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3913  kmp_info_t *this_thr,
3914  kmp_team_t *team);
3915 
3916 // should never have been exported
3917 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3918 extern int __kmp_invoke_teams_master(int gtid);
3919 extern void __kmp_teams_master(int gtid);
3920 extern int __kmp_aux_get_team_num();
3921 extern int __kmp_aux_get_num_teams();
3922 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3923 extern void __kmp_user_set_library(enum library_type arg);
3924 extern void __kmp_aux_set_library(enum library_type arg);
3925 extern void __kmp_aux_set_stacksize(size_t arg);
3926 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3927 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3928 
3929 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3930 void kmpc_set_blocktime(int arg);
3931 void ompc_set_nested(int flag);
3932 void ompc_set_dynamic(int flag);
3933 void ompc_set_num_threads(int arg);
3934 
3935 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3936  kmp_team_t *team, int tid);
3937 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3938 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3939  kmp_tasking_flags_t *flags,
3940  size_t sizeof_kmp_task_t,
3941  size_t sizeof_shareds,
3942  kmp_routine_entry_t task_entry);
3943 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3944  kmp_team_t *team, int tid,
3945  int set_curr_task);
3946 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3947 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3948 
3949 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3950  int gtid,
3951  kmp_task_t *task);
3952 extern void __kmp_fulfill_event(kmp_event_t *event);
3953 
3954 extern void __kmp_free_task_team(kmp_info_t *thread,
3955  kmp_task_team_t *task_team);
3956 extern void __kmp_reap_task_teams(void);
3957 extern void __kmp_wait_to_unref_task_teams(void);
3958 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3959  int always);
3960 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3961 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3962 #if USE_ITT_BUILD
3963  ,
3964  void *itt_sync_obj
3965 #endif /* USE_ITT_BUILD */
3966  ,
3967  int wait = 1);
3968 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3969  int gtid);
3970 
3971 extern int __kmp_is_address_mapped(void *addr);
3972 extern kmp_uint64 __kmp_hardware_timestamp(void);
3973 
3974 #if KMP_OS_UNIX
3975 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3976 #endif
3977 
3978 /* ------------------------------------------------------------------------ */
3979 //
3980 // Assembly routines that have no compiler intrinsic replacement
3981 //
3982 
3983 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3984  void *argv[]
3985 #if OMPT_SUPPORT
3986  ,
3987  void **exit_frame_ptr
3988 #endif
3989 );
3990 
3991 /* ------------------------------------------------------------------------ */
3992 
3993 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3994 KMP_EXPORT void __kmpc_end(ident_t *);
3995 
3996 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3997  kmpc_ctor_vec ctor,
3998  kmpc_cctor_vec cctor,
3999  kmpc_dtor_vec dtor,
4000  size_t vector_length);
4001 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4002  kmpc_ctor ctor, kmpc_cctor cctor,
4003  kmpc_dtor dtor);
4004 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4005  void *data, size_t size);
4006 
4007 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4008 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4009 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4010 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4011 
4012 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4013 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4014  kmpc_micro microtask, ...);
4015 KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4016  kmpc_micro microtask, kmp_int32 cond,
4017  void *args);
4018 
4019 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4020 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4021 
4022 KMP_EXPORT void __kmpc_flush(ident_t *);
4023 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4024 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4025 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4026 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4027  kmp_int32 filter);
4028 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4029 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4030 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4031 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4032  kmp_critical_name *);
4033 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4034  kmp_critical_name *);
4035 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4036  kmp_critical_name *, uint32_t hint);
4037 
4038 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4039 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4040 
4041 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4042  kmp_int32 global_tid);
4043 
4044 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4045 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4046 
4047 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4048 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4049  kmp_int32 numberOfSections);
4050 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4051 
4052 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4053  kmp_int32 schedtype, kmp_int32 *plastiter,
4054  kmp_int *plower, kmp_int *pupper,
4055  kmp_int *pstride, kmp_int incr,
4056  kmp_int chunk);
4057 
4058 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4059 
4060 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4061  size_t cpy_size, void *cpy_data,
4062  void (*cpy_func)(void *, void *),
4063  kmp_int32 didit);
4064 
4065 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4066  void *cpy_data);
4067 
4068 extern void KMPC_SET_NUM_THREADS(int arg);
4069 extern void KMPC_SET_DYNAMIC(int flag);
4070 extern void KMPC_SET_NESTED(int flag);
4071 
4072 /* OMP 3.0 tasking interface routines */
4073 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4074  kmp_task_t *new_task);
4075 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4076  kmp_int32 flags,
4077  size_t sizeof_kmp_task_t,
4078  size_t sizeof_shareds,
4079  kmp_routine_entry_t task_entry);
4080 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4081  ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4082  size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4083 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4084  kmp_task_t *task);
4085 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4086  kmp_task_t *task);
4087 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4088  kmp_task_t *new_task);
4089 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4090 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4091  int end_part);
4092 
4093 #if TASK_UNUSED
4094 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4095 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4096  kmp_task_t *task);
4097 #endif // TASK_UNUSED
4098 
4099 /* ------------------------------------------------------------------------ */
4100 
4101 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4102 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4103 
4104 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4105  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4106  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4107  kmp_depend_info_t *noalias_dep_list);
4108 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4109  kmp_int32 ndeps,
4110  kmp_depend_info_t *dep_list,
4111  kmp_int32 ndeps_noalias,
4112  kmp_depend_info_t *noalias_dep_list);
4113 /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4114  * Placeholder for taskwait with nowait clause.*/
4115 KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4116  kmp_int32 ndeps,
4117  kmp_depend_info_t *dep_list,
4118  kmp_int32 ndeps_noalias,
4119  kmp_depend_info_t *noalias_dep_list,
4120  kmp_int32 has_no_wait);
4121 
4122 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4123  bool serialize_immediate);
4124 
4125 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4126  kmp_int32 cncl_kind);
4127 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4128  kmp_int32 cncl_kind);
4129 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4130 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4131 
4132 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4133 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4134 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4135  kmp_int32 if_val, kmp_uint64 *lb,
4136  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4137  kmp_int32 sched, kmp_uint64 grainsize,
4138  void *task_dup);
4139 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4140  kmp_task_t *task, kmp_int32 if_val,
4141  kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4142  kmp_int32 nogroup, kmp_int32 sched,
4143  kmp_uint64 grainsize, kmp_int32 modifier,
4144  void *task_dup);
4145 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4146 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4147 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4148 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4149  int is_ws, int num,
4150  void *data);
4151 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4152  int num, void *data);
4153 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4154  int is_ws);
4155 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4156  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4157  kmp_task_affinity_info_t *affin_list);
4158 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4159 KMP_EXPORT int __kmp_get_max_teams(void);
4160 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4161 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4162 
4163 /* Interface target task integration */
4164 KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4165 KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4166 
4167 /* Lock interface routines (fast versions with gtid passed in) */
4168 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4169  void **user_lock);
4170 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4171  void **user_lock);
4172 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4173  void **user_lock);
4174 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4175  void **user_lock);
4176 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4177 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4178  void **user_lock);
4179 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4180  void **user_lock);
4181 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4182  void **user_lock);
4183 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4184 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4185  void **user_lock);
4186 
4187 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4188  void **user_lock, uintptr_t hint);
4189 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4190  void **user_lock,
4191  uintptr_t hint);
4192 
4193 #if OMPX_TASKGRAPH
4194 // Taskgraph's Record & Replay mechanism
4195 // __kmp_tdg_is_recording: check whether a given TDG is recording
4196 // status: the tdg's current status
4197 static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4198  return status == KMP_TDG_RECORDING;
4199 }
4200 
4201 KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4202  kmp_int32 input_flags,
4203  kmp_int32 tdg_id);
4204 KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4205  kmp_int32 input_flags, kmp_int32 tdg_id);
4206 #endif
4207 /* Interface to fast scalable reduce methods routines */
4208 
4209 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4210  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4211  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4212  kmp_critical_name *lck);
4213 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4214  kmp_critical_name *lck);
4215 KMP_EXPORT kmp_int32 __kmpc_reduce(
4216  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4217  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4218  kmp_critical_name *lck);
4219 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4220  kmp_critical_name *lck);
4221 
4222 /* Internal fast reduction routines */
4223 
4224 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4225  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4226  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4227  kmp_critical_name *lck);
4228 
4229 // this function is for testing set/get/determine reduce method
4230 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4231 
4232 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4233 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4234 
4235 // C++ port
4236 // missing 'extern "C"' declarations
4237 
4238 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4239 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4240 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4241  kmp_int32 num_threads);
4242 
4243 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4244  int proc_bind);
4245 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4246  kmp_int32 num_teams,
4247  kmp_int32 num_threads);
4248 /* Function for OpenMP 5.1 num_teams clause */
4249 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4250  kmp_int32 num_teams_lb,
4251  kmp_int32 num_teams_ub,
4252  kmp_int32 num_threads);
4253 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4254  kmpc_micro microtask, ...);
4255 struct kmp_dim { // loop bounds info casted to kmp_int64
4256  kmp_int64 lo; // lower
4257  kmp_int64 up; // upper
4258  kmp_int64 st; // stride
4259 };
4260 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4261  kmp_int32 num_dims,
4262  const struct kmp_dim *dims);
4263 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4264  const kmp_int64 *vec);
4265 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4266  const kmp_int64 *vec);
4267 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4268 
4269 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4270  void *data, size_t size,
4271  void ***cache);
4272 
4273 // The routines below are not exported.
4274 // Consider making them 'static' in corresponding source files.
4275 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4276  void *data_addr, size_t pc_size);
4277 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4278  void *data_addr,
4279  size_t pc_size);
4280 void __kmp_threadprivate_resize_cache(int newCapacity);
4281 void __kmp_cleanup_threadprivate_caches();
4282 
4283 // ompc_, kmpc_ entries moved from omp.h.
4284 #if KMP_OS_WINDOWS
4285 #define KMPC_CONVENTION __cdecl
4286 #else
4287 #define KMPC_CONVENTION
4288 #endif
4289 
4290 #ifndef __OMP_H
4291 typedef enum omp_sched_t {
4292  omp_sched_static = 1,
4293  omp_sched_dynamic = 2,
4294  omp_sched_guided = 3,
4295  omp_sched_auto = 4
4296 } omp_sched_t;
4297 typedef void *kmp_affinity_mask_t;
4298 #endif
4299 
4300 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4301 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4302 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4303 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4304 KMP_EXPORT int KMPC_CONVENTION
4305 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4306 KMP_EXPORT int KMPC_CONVENTION
4307 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4308 KMP_EXPORT int KMPC_CONVENTION
4309 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4310 
4311 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4312 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4313 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4314 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4315 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4316 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4317 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4318 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4319 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4320  char const *format);
4321 
4322 enum kmp_target_offload_kind {
4323  tgt_disabled = 0,
4324  tgt_default = 1,
4325  tgt_mandatory = 2
4326 };
4327 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4328 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4329 extern kmp_target_offload_kind_t __kmp_target_offload;
4330 extern int __kmpc_get_target_offload();
4331 
4332 // Constants used in libomptarget
4333 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4334 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4335 
4336 // OMP Pause Resource
4337 
4338 // The following enum is used both to set the status in __kmp_pause_status, and
4339 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4340 typedef enum kmp_pause_status_t {
4341  kmp_not_paused = 0, // status is not paused, or, requesting resume
4342  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4343  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4344 } kmp_pause_status_t;
4345 
4346 // This stores the pause state of the runtime
4347 extern kmp_pause_status_t __kmp_pause_status;
4348 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4349 extern int __kmp_pause_resource(kmp_pause_status_t level);
4350 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4351 extern void __kmp_resume_if_soft_paused();
4352 // Hard resume simply resets the status to not paused. Library will appear to
4353 // be uninitialized after hard pause. Let OMP constructs trigger required
4354 // initializations.
4355 static inline void __kmp_resume_if_hard_paused() {
4356  if (__kmp_pause_status == kmp_hard_paused) {
4357  __kmp_pause_status = kmp_not_paused;
4358  }
4359 }
4360 
4361 extern void __kmp_omp_display_env(int verbose);
4362 
4363 // 1: it is initializing hidden helper team
4364 extern volatile int __kmp_init_hidden_helper;
4365 // 1: the hidden helper team is done
4366 extern volatile int __kmp_hidden_helper_team_done;
4367 // 1: enable hidden helper task
4368 extern kmp_int32 __kmp_enable_hidden_helper;
4369 // Main thread of hidden helper team
4370 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4371 // Descriptors for the hidden helper threads
4372 extern kmp_info_t **__kmp_hidden_helper_threads;
4373 // Number of hidden helper threads
4374 extern kmp_int32 __kmp_hidden_helper_threads_num;
4375 // Number of hidden helper tasks that have not been executed yet
4376 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4377 
4378 extern void __kmp_hidden_helper_initialize();
4379 extern void __kmp_hidden_helper_threads_initz_routine();
4380 extern void __kmp_do_initialize_hidden_helper_threads();
4381 extern void __kmp_hidden_helper_threads_initz_wait();
4382 extern void __kmp_hidden_helper_initz_release();
4383 extern void __kmp_hidden_helper_threads_deinitz_wait();
4384 extern void __kmp_hidden_helper_threads_deinitz_release();
4385 extern void __kmp_hidden_helper_main_thread_wait();
4386 extern void __kmp_hidden_helper_worker_thread_wait();
4387 extern void __kmp_hidden_helper_worker_thread_signal();
4388 extern void __kmp_hidden_helper_main_thread_release();
4389 
4390 // Check whether a given thread is a hidden helper thread
4391 #define KMP_HIDDEN_HELPER_THREAD(gtid) \
4392  ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4393 
4394 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4395  ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4396 
4397 #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4398  ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4399 
4400 #define KMP_HIDDEN_HELPER_TEAM(team) \
4401  (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4402 
4403 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4404 // main thread, is skipped.
4405 #define KMP_GTID_TO_SHADOW_GTID(gtid) \
4406  ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4407 
4408 // Return the adjusted gtid value by subtracting from gtid the number
4409 // of hidden helper threads. This adjusted value is the gtid the thread would
4410 // have received if there were no hidden helper threads.
4411 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4412  int adjusted_gtid = gtid;
4413  if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4414  gtid - __kmp_hidden_helper_threads_num >= 0) {
4415  adjusted_gtid -= __kmp_hidden_helper_threads_num;
4416  }
4417  return adjusted_gtid;
4418 }
4419 
4420 // Support for error directive
4421 typedef enum kmp_severity_t {
4422  severity_warning = 1,
4423  severity_fatal = 2
4424 } kmp_severity_t;
4425 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4426 
4427 // Support for scope directive
4428 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4429 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4430 
4431 #ifdef __cplusplus
4432 }
4433 #endif
4434 
4435 template <bool C, bool S>
4436 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4437 template <bool C, bool S>
4438 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4439 template <bool C, bool S>
4440 extern void __kmp_atomic_suspend_64(int th_gtid,
4441  kmp_atomic_flag_64<C, S> *flag);
4442 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4443 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4444 template <bool C, bool S>
4445 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4446 template <bool C, bool S>
4447 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4448 template <bool C, bool S>
4449 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4450 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4451 #endif
4452 template <bool C, bool S>
4453 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4454 template <bool C, bool S>
4455 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4456 template <bool C, bool S>
4457 extern void __kmp_atomic_resume_64(int target_gtid,
4458  kmp_atomic_flag_64<C, S> *flag);
4459 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4460 
4461 template <bool C, bool S>
4462 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4463  kmp_flag_32<C, S> *flag, int final_spin,
4464  int *thread_finished,
4465 #if USE_ITT_BUILD
4466  void *itt_sync_obj,
4467 #endif /* USE_ITT_BUILD */
4468  kmp_int32 is_constrained);
4469 template <bool C, bool S>
4470 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4471  kmp_flag_64<C, S> *flag, int final_spin,
4472  int *thread_finished,
4473 #if USE_ITT_BUILD
4474  void *itt_sync_obj,
4475 #endif /* USE_ITT_BUILD */
4476  kmp_int32 is_constrained);
4477 template <bool C, bool S>
4478 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4479  kmp_atomic_flag_64<C, S> *flag,
4480  int final_spin, int *thread_finished,
4481 #if USE_ITT_BUILD
4482  void *itt_sync_obj,
4483 #endif /* USE_ITT_BUILD */
4484  kmp_int32 is_constrained);
4485 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4486  kmp_flag_oncore *flag, int final_spin,
4487  int *thread_finished,
4488 #if USE_ITT_BUILD
4489  void *itt_sync_obj,
4490 #endif /* USE_ITT_BUILD */
4491  kmp_int32 is_constrained);
4492 
4493 extern int __kmp_nesting_mode;
4494 extern int __kmp_nesting_mode_nlevels;
4495 extern int *__kmp_nesting_nth_level;
4496 extern void __kmp_init_nesting_mode();
4497 extern void __kmp_set_nesting_mode_threads();
4498 
4506  FILE *f;
4507 
4508  void close() {
4509  if (f && f != stdout && f != stderr) {
4510  fclose(f);
4511  f = nullptr;
4512  }
4513  }
4514 
4515 public:
4516  kmp_safe_raii_file_t() : f(nullptr) {}
4517  kmp_safe_raii_file_t(const char *filename, const char *mode,
4518  const char *env_var = nullptr)
4519  : f(nullptr) {
4520  open(filename, mode, env_var);
4521  }
4522  ~kmp_safe_raii_file_t() { close(); }
4523 
4527  void open(const char *filename, const char *mode,
4528  const char *env_var = nullptr) {
4529  KMP_ASSERT(!f);
4530  f = fopen(filename, mode);
4531  if (!f) {
4532  int code = errno;
4533  if (env_var) {
4534  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4535  KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4536  } else {
4537  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4538  __kmp_msg_null);
4539  }
4540  }
4541  }
4544  int try_open(const char *filename, const char *mode) {
4545  KMP_ASSERT(!f);
4546  f = fopen(filename, mode);
4547  if (!f)
4548  return errno;
4549  return 0;
4550  }
4553  void set_stdout() {
4554  KMP_ASSERT(!f);
4555  f = stdout;
4556  }
4559  void set_stderr() {
4560  KMP_ASSERT(!f);
4561  f = stderr;
4562  }
4563  operator bool() { return bool(f); }
4564  operator FILE *() { return f; }
4565 };
4566 
4567 template <typename SourceType, typename TargetType,
4568  bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4569  bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4570  bool isSourceSigned = std::is_signed<SourceType>::value,
4571  bool isTargetSigned = std::is_signed<TargetType>::value>
4572 struct kmp_convert {};
4573 
4574 // Both types are signed; Source smaller
4575 template <typename SourceType, typename TargetType>
4576 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4577  static TargetType to(SourceType src) { return (TargetType)src; }
4578 };
4579 // Source equal
4580 template <typename SourceType, typename TargetType>
4581 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4582  static TargetType to(SourceType src) { return src; }
4583 };
4584 // Source bigger
4585 template <typename SourceType, typename TargetType>
4586 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4587  static TargetType to(SourceType src) {
4588  KMP_ASSERT(src <= static_cast<SourceType>(
4589  (std::numeric_limits<TargetType>::max)()));
4590  KMP_ASSERT(src >= static_cast<SourceType>(
4591  (std::numeric_limits<TargetType>::min)()));
4592  return (TargetType)src;
4593  }
4594 };
4595 
4596 // Source signed, Target unsigned
4597 // Source smaller
4598 template <typename SourceType, typename TargetType>
4599 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4600  static TargetType to(SourceType src) {
4601  KMP_ASSERT(src >= 0);
4602  return (TargetType)src;
4603  }
4604 };
4605 // Source equal
4606 template <typename SourceType, typename TargetType>
4607 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4608  static TargetType to(SourceType src) {
4609  KMP_ASSERT(src >= 0);
4610  return (TargetType)src;
4611  }
4612 };
4613 // Source bigger
4614 template <typename SourceType, typename TargetType>
4615 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4616  static TargetType to(SourceType src) {
4617  KMP_ASSERT(src >= 0);
4618  KMP_ASSERT(src <= static_cast<SourceType>(
4619  (std::numeric_limits<TargetType>::max)()));
4620  return (TargetType)src;
4621  }
4622 };
4623 
4624 // Source unsigned, Target signed
4625 // Source smaller
4626 template <typename SourceType, typename TargetType>
4627 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4628  static TargetType to(SourceType src) { return (TargetType)src; }
4629 };
4630 // Source equal
4631 template <typename SourceType, typename TargetType>
4632 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4633  static TargetType to(SourceType src) {
4634  KMP_ASSERT(src <= static_cast<SourceType>(
4635  (std::numeric_limits<TargetType>::max)()));
4636  return (TargetType)src;
4637  }
4638 };
4639 // Source bigger
4640 template <typename SourceType, typename TargetType>
4641 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4642  static TargetType to(SourceType src) {
4643  KMP_ASSERT(src <= static_cast<SourceType>(
4644  (std::numeric_limits<TargetType>::max)()));
4645  return (TargetType)src;
4646  }
4647 };
4648 
4649 // Source unsigned, Target unsigned
4650 // Source smaller
4651 template <typename SourceType, typename TargetType>
4652 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4653  static TargetType to(SourceType src) { return (TargetType)src; }
4654 };
4655 // Source equal
4656 template <typename SourceType, typename TargetType>
4657 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4658  static TargetType to(SourceType src) { return src; }
4659 };
4660 // Source bigger
4661 template <typename SourceType, typename TargetType>
4662 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4663  static TargetType to(SourceType src) {
4664  KMP_ASSERT(src <= static_cast<SourceType>(
4665  (std::numeric_limits<TargetType>::max)()));
4666  return (TargetType)src;
4667  }
4668 };
4669 
4670 template <typename T1, typename T2>
4671 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4672  *dest = kmp_convert<T1, T2>::to(src);
4673 }
4674 
4675 #endif /* KMP_H */
__kmpc_dispatch_fini_8
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
Definition: kmp_dispatch.cpp:2832
KMP_IDENT_IMB
@ KMP_IDENT_IMB
Definition: kmp.h:194
__kmpc_end_single
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:1883
__kmpc_dispatch_next_4u
int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint32 *p_lb, kmp_uint32 *p_ub, kmp_int32 *p_st)
Definition: kmp_dispatch.cpp:2772
__kmpc_fork_call
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
Definition: kmp_csupport.cpp:262
KMP_IDENT_WORK_LOOP
@ KMP_IDENT_WORK_LOOP
Definition: kmp.h:214
KMP_IDENT_WORK_SECTIONS
@ KMP_IDENT_WORK_SECTIONS
Definition: kmp.h:216
__kmpc_proxy_task_completed
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
Definition: kmp_tasking.cpp:4492
__kmpc_reduce
KMP_EXPORT kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
Definition: kmp_csupport.cpp:3765
__kmpc_dispatch_init_8u
void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint64 lb, kmp_uint64 ub, kmp_int64 st, kmp_int64 chunk)
Definition: kmp_dispatch.cpp:2676
__kmpc_omp_task_with_deps
KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
Definition: kmp_taskdeps.cpp:662
__kmpc_push_num_teams_51
KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams_lb, kmp_int32 num_teams_ub, kmp_int32 num_threads)
Definition: kmp_csupport.cpp:400
KMP_IDENT_WORK_DISTRIBUTE
@ KMP_IDENT_WORK_DISTRIBUTE
Definition: kmp.h:218
__kmpc_omp_get_target_async_handle_ptr
KMP_EXPORT void ** __kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid)
Definition: kmp_tasking.cpp:5415
__kmpc_masked
KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, kmp_int32 filter)
Definition: kmp_csupport.cpp:858
ident::reserved_2
kmp_int32 reserved_2
Definition: kmp.h:238
__kmpc_global_thread_num
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
Definition: kmp_csupport.cpp:99
kmp_sch_default
@ kmp_sch_default
Definition: kmp.h:465
KMP_IDENT_BARRIER_IMPL
@ KMP_IDENT_BARRIER_IMPL
Definition: kmp.h:205
__kmpc_end
KMP_EXPORT void __kmpc_end(ident_t *)
Definition: kmp_csupport.cpp:59
kmpc_ctor
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1682
kmp_nm_static
@ kmp_nm_static
Definition: kmp.h:406
__kmpc_master
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:769
__kmpc_dispatch_next_4
int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int32 *p_lb, kmp_int32 *p_ub, kmp_int32 *p_st)
Definition: kmp_dispatch.cpp:2756
__kmpc_single
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:1830
__kmpc_taskred_init
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
Definition: kmp_tasking.cpp:2599
__kmpc_end_serialized_parallel
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:540
__kmpc_task_reduction_init
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
Definition: kmp_tasking.cpp:2572
kmp_sch_guided_chunked
@ kmp_sch_guided_chunked
Definition: kmp.h:362
__kmpc_dispatch_init_4
void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st, kmp_int32 chunk)
Definition: kmp_dispatch.cpp:2638
__kmpc_flush
KMP_EXPORT void __kmpc_flush(ident_t *)
Definition: kmp_csupport.cpp:698
__kmpc_next_section
KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, kmp_int32 numberOfSections)
Definition: kmp_dispatch.cpp:2387
kmp_safe_raii_file_t::open
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4527
kmp_safe_raii_file_t
Definition: kmp.h:4505
ident::psource
const char * psource
Definition: kmp.h:244
kmp_sch_guided_simd
@ kmp_sch_guided_simd
Definition: kmp.h:378
__kmpc_task_reduction_get_th_data
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
Definition: kmp_tasking.cpp:2642
__kmpc_end_reduce_nowait
KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
Definition: kmp_csupport.cpp:3696
__kmpc_bound_thread_num
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
Definition: kmp_csupport.cpp:134
__kmpc_end_reduce
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
Definition: kmp_csupport.cpp:3887
__kmpc_fork_call_if
KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs, kmpc_micro microtask, kmp_int32 cond, void *args)
Definition: kmp_csupport.cpp:343
__kmpc_taskloop
void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, void *task_dup)
Definition: kmp_tasking.cpp:5369
kmp_sch_upper
@ kmp_sch_upper
Definition: kmp.h:382
kmp_sch_modifier_monotonic
@ kmp_sch_modifier_monotonic
Definition: kmp.h:445
ident::reserved_3
kmp_int32 reserved_3
Definition: kmp.h:243
kmp_safe_raii_file_t::set_stdout
void set_stdout()
Definition: kmp.h:4553
__kmpc_omp_has_task_team
KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid)
Definition: kmp_tasking.cpp:5436
__kmpc_taskloop_5
void __kmpc_taskloop_5(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, int modifier, void *task_dup)
Definition: kmp_tasking.cpp:5396
__kmpc_ordered
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:943
kmp_safe_raii_file_t::set_stderr
void set_stderr()
Definition: kmp.h:4559
KMP_IDENT_BARRIER_EXPL
@ KMP_IDENT_BARRIER_EXPL
Definition: kmp.h:203
__kmpc_end_ordered
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:1016
sched_type
sched_type
Definition: kmp.h:357
__kmpc_begin
KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags)
Definition: kmp_csupport.cpp:36
__kmpc_doacross_init
void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims, const struct kmp_dim *dims)
Definition: kmp_csupport.cpp:4051
__kmpc_threadprivate_register_vec
KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length)
Definition: kmp_threadprivate.cpp:746
kmp_nm_lower
@ kmp_nm_lower
Definition: kmp.h:402
__kmpc_dispatch_next_8
int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int64 *p_lb, kmp_int64 *p_ub, kmp_int64 *p_st)
Definition: kmp_dispatch.cpp:2789
kmp_ord_static
@ kmp_ord_static
Definition: kmp.h:386
__kmpc_fork_teams
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
Definition: kmp_csupport.cpp:421
kmp_nm_upper
@ kmp_nm_upper
Definition: kmp.h:429
__kmpc_dispatch_fini_8u
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
Definition: kmp_dispatch.cpp:2846
ident::reserved_1
kmp_int32 reserved_1
Definition: kmp.h:235
kmp_nm_ord_auto
@ kmp_nm_ord_auto
Definition: kmp.h:427
KMP_IDENT_AUTOPAR
@ KMP_IDENT_AUTOPAR
Definition: kmp.h:199
__kmpc_end_barrier_master
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:1751
kmp_sch_modifier_nonmonotonic
@ kmp_sch_modifier_nonmonotonic
Definition: kmp.h:447
__kmpc_proxy_task_completed_ooo
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
Definition: kmp_tasking.cpp:4556
ident
Definition: kmp.h:234
__kmpc_critical
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
Definition: kmp_csupport.cpp:1251
ident_t
struct ident ident_t
__kmpc_global_num_threads
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
Definition: kmp_csupport.cpp:121
kmpc_dtor_vec
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1711
__kmpc_task_reduction_modifier_fini
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
Definition: kmp_tasking.cpp:2851
__kmpc_bound_num_threads
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
Definition: kmp_csupport.cpp:144
__kmpc_end_masked
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:914
__kmpc_ok_to_fork
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
Definition: kmp_csupport.cpp:156
kmpc_ctor_vec
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1705
kmp_distribute_static
@ kmp_distribute_static
Definition: kmp.h:396
kmp_sch_auto
@ kmp_sch_auto
Definition: kmp.h:364
__kmpc_copyprivate_light
KMP_EXPORT void * __kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid, void *cpy_data)
Definition: kmp_csupport.cpp:2254
kmpc_dtor
void(* kmpc_dtor)(void *)
Definition: kmp.h:1688
__kmpc_barrier
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:720
__kmpc_end_master
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:826
kmp_sch_runtime_simd
@ kmp_sch_runtime_simd
Definition: kmp.h:379
kmpc_cctor
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1695
__kmpc_in_parallel
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
Definition: kmp_csupport.cpp:218
KMP_IDENT_KMPC
@ KMP_IDENT_KMPC
Definition: kmp.h:196
__kmpc_task_reduction_modifier_init
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
Definition: kmp_tasking.cpp:2817
__kmpc_reduce_nowait
KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
Definition: kmp_csupport.cpp:3536
__kmpc_threadprivate_register
KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor)
Definition: kmp_threadprivate.cpp:504
__kmpc_push_num_threads
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
Definition: kmp_csupport.cpp:231
__kmpc_dispatch_fini_4u
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
Definition: kmp_dispatch.cpp:2839
__kmpc_critical_with_hint
KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, kmp_critical_name *, uint32_t hint)
Definition: kmp_csupport.cpp:1478
__kmpc_sections_init
KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid)
Definition: kmp_dispatch.cpp:2303
kmp_sch_lower
@ kmp_sch_lower
Definition: kmp.h:358
kmpc_micro
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1664
kmp_ord_lower
@ kmp_ord_lower
Definition: kmp.h:384
__kmpc_end_critical
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
Definition: kmp_csupport.cpp:1610
__kmpc_taskred_modifier_init
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
Definition: kmp_tasking.cpp:2837
kmp_distribute_static_chunked
@ kmp_distribute_static_chunked
Definition: kmp.h:395
kmp_nm_ord_static
@ kmp_nm_ord_static
Definition: kmp.h:423
kmp_safe_raii_file_t::try_open
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4544
__kmpc_barrier_master_nowait
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:1767
kmp_ord_auto
@ kmp_ord_auto
Definition: kmp.h:390
__kmpc_copyprivate
KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *), kmp_int32 didit)
Definition: kmp_csupport.cpp:2178
kmp_sch_static
@ kmp_sch_static
Definition: kmp.h:360
__kmpc_end_sections
KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid)
Definition: kmp_dispatch.cpp:2476
__kmpc_dispatch_fini_4
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
Definition: kmp_dispatch.cpp:2825
kmpc_cctor_vec
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition: kmp.h:1717
__kmpc_dispatch_init_8
void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int64 lb, kmp_int64 ub, kmp_int64 st, kmp_int64 chunk)
Definition: kmp_dispatch.cpp:2663
__kmpc_omp_wait_deps
KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
Definition: kmp_taskdeps.cpp:873
__kmpc_threadprivate_cached
KMP_EXPORT void * __kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, void *data, size_t size, void ***cache)
Definition: kmp_threadprivate.cpp:614
kmp_nm_guided_chunked
@ kmp_nm_guided_chunked
Definition: kmp.h:408
__kmpc_dispatch_init_4u
void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint32 lb, kmp_uint32 ub, kmp_int32 st, kmp_int32 chunk)
Definition: kmp_dispatch.cpp:2650
ident::flags
kmp_int32 flags
Definition: kmp.h:236
KMP_IDENT_ATOMIC_HINT_MASK
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition: kmp.h:223
__kmpc_push_num_teams
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
Definition: kmp_csupport.cpp:375
KMP_IDENT_ATOMIC_REDUCE
@ KMP_IDENT_ATOMIC_REDUCE
Definition: kmp.h:201
__kmpc_dispatch_next_8u
int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint64 *p_lb, kmp_uint64 *p_ub, kmp_int64 *p_st)
Definition: kmp_dispatch.cpp:2805
__kmpc_serialized_parallel
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:522
kmp_ord_upper
@ kmp_ord_upper
Definition: kmp.h:392
__kmpc_for_static_fini
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:1910
__kmpc_barrier_master
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
Definition: kmp_csupport.cpp:1707
kmp_nm_auto
@ kmp_nm_auto
Definition: kmp.h:410
__kmpc_omp_reg_task_with_affinity
KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, kmp_task_affinity_info_t *affin_list)
Definition: kmp_tasking.cpp:1729