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 #define KMP_NSEC_PER_USEC 1000L
184 
193 enum {
198  /* 0x04 is no longer used */
207  KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
208  KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
209  KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
210 
211  KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
212  KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
213 
225  KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
226  KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
227  KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
228  KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
229  KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
230 };
231 
235 typedef struct ident {
236  kmp_int32 reserved_1;
237  kmp_int32 flags;
239  kmp_int32 reserved_2;
240 #if USE_ITT_BUILD
241 /* but currently used for storing region-specific ITT */
242 /* contextual information. */
243 #endif /* USE_ITT_BUILD */
244  kmp_int32 reserved_3;
245  char const *psource;
249  // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
250  kmp_int32 get_openmp_version() {
251  return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
252  }
258 // Some forward declarations.
259 typedef union kmp_team kmp_team_t;
260 typedef struct kmp_taskdata kmp_taskdata_t;
261 typedef union kmp_task_team kmp_task_team_t;
262 typedef union kmp_team kmp_team_p;
263 typedef union kmp_info kmp_info_p;
264 typedef union kmp_root kmp_root_p;
265 
266 template <bool C = false, bool S = true> class kmp_flag_32;
267 template <bool C = false, bool S = true> class kmp_flag_64;
268 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
269 class kmp_flag_oncore;
270 
271 #ifdef __cplusplus
272 extern "C" {
273 #endif
274 
275 /* ------------------------------------------------------------------------ */
276 
277 /* Pack two 32-bit signed integers into a 64-bit signed integer */
278 /* ToDo: Fix word ordering for big-endian machines. */
279 #define KMP_PACK_64(HIGH_32, LOW_32) \
280  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
281 
282 // Generic string manipulation macros. Assume that _x is of type char *
283 #define SKIP_WS(_x) \
284  { \
285  while (*(_x) == ' ' || *(_x) == '\t') \
286  (_x)++; \
287  }
288 #define SKIP_DIGITS(_x) \
289  { \
290  while (*(_x) >= '0' && *(_x) <= '9') \
291  (_x)++; \
292  }
293 #define SKIP_TOKEN(_x) \
294  { \
295  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
296  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
297  (_x)++; \
298  }
299 #define SKIP_TO(_x, _c) \
300  { \
301  while (*(_x) != '\0' && *(_x) != (_c)) \
302  (_x)++; \
303  }
304 
305 /* ------------------------------------------------------------------------ */
306 
307 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
308 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
309 
310 /* ------------------------------------------------------------------------ */
311 /* Enumeration types */
312 
313 enum kmp_state_timer {
314  ts_stop,
315  ts_start,
316  ts_pause,
317 
318  ts_last_state
319 };
320 
321 enum dynamic_mode {
322  dynamic_default,
323 #ifdef USE_LOAD_BALANCE
324  dynamic_load_balance,
325 #endif /* USE_LOAD_BALANCE */
326  dynamic_random,
327  dynamic_thread_limit,
328  dynamic_max
329 };
330 
331 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
332  * not include it here */
333 #ifndef KMP_SCHED_TYPE_DEFINED
334 #define KMP_SCHED_TYPE_DEFINED
335 typedef enum kmp_sched {
336  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
337  // Note: need to adjust __kmp_sch_map global array in case enum is changed
338  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
339  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
340  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
341  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
342  kmp_sched_upper_std = 5, // upper bound for standard schedules
343  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
344  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
345 #if KMP_STATIC_STEAL_ENABLED
346  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
347 #endif
348  kmp_sched_upper,
349  kmp_sched_default = kmp_sched_static, // default scheduling
350  kmp_sched_monotonic = 0x80000000
351 } kmp_sched_t;
352 #endif
353 
358 enum sched_type : kmp_int32 {
360  kmp_sch_static_chunked = 33,
362  kmp_sch_dynamic_chunked = 35,
364  kmp_sch_runtime = 37,
366  kmp_sch_trapezoidal = 39,
367 
368  /* accessible only through KMP_SCHEDULE environment variable */
369  kmp_sch_static_greedy = 40,
370  kmp_sch_static_balanced = 41,
371  /* accessible only through KMP_SCHEDULE environment variable */
372  kmp_sch_guided_iterative_chunked = 42,
373  kmp_sch_guided_analytical_chunked = 43,
374  /* accessible only through KMP_SCHEDULE environment variable */
375  kmp_sch_static_steal = 44,
376 
377  /* static with chunk adjustment (e.g., simd) */
378  kmp_sch_static_balanced_chunked = 45,
382  /* accessible only through KMP_SCHEDULE environment variable */
386  kmp_ord_static_chunked = 65,
388  kmp_ord_dynamic_chunked = 67,
389  kmp_ord_guided_chunked = 68,
390  kmp_ord_runtime = 69,
392  kmp_ord_trapezoidal = 71,
395  /* Schedules for Distribute construct */
399  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
400  single iteration/chunk, even if the loop is serialized. For the schedule
401  types listed above, the entire iteration vector is returned if the loop is
402  serialized. This doesn't work for gcc/gcomp sections. */
403  kmp_nm_lower = 160,
405  kmp_nm_static_chunked =
406  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
408  kmp_nm_dynamic_chunked = 163,
410  kmp_nm_runtime = 165,
411  kmp_nm_auto = 166,
412  kmp_nm_trapezoidal = 167,
413 
414  /* accessible only through KMP_SCHEDULE environment variable */
415  kmp_nm_static_greedy = 168,
416  kmp_nm_static_balanced = 169,
417  /* accessible only through KMP_SCHEDULE environment variable */
418  kmp_nm_guided_iterative_chunked = 170,
419  kmp_nm_guided_analytical_chunked = 171,
420  kmp_nm_static_steal =
421  172, /* accessible only through OMP_SCHEDULE environment variable */
422 
423  kmp_nm_ord_static_chunked = 193,
425  kmp_nm_ord_dynamic_chunked = 195,
426  kmp_nm_ord_guided_chunked = 196,
427  kmp_nm_ord_runtime = 197,
429  kmp_nm_ord_trapezoidal = 199,
432  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
433  we need to distinguish the three possible cases (no modifier, monotonic
434  modifier, nonmonotonic modifier), we need separate bits for each modifier.
435  The absence of monotonic does not imply nonmonotonic, especially since 4.5
436  says that the behaviour of the "no modifier" case is implementation defined
437  in 4.5, but will become "nonmonotonic" in 5.0.
438 
439  Since we're passing a full 32 bit value, we can use a couple of high bits
440  for these flags; out of paranoia we avoid the sign bit.
441 
442  These modifiers can be or-ed into non-static schedules by the compiler to
443  pass the additional information. They will be stripped early in the
444  processing in __kmp_dispatch_init when setting up schedules, so most of the
445  code won't ever see schedules with these bits set. */
447  (1 << 29),
449  (1 << 30),
451 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
452  (enum sched_type)( \
454 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
455 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
456 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
457  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
458 #define SCHEDULE_GET_MODIFIERS(s) \
459  ((enum sched_type)( \
460  (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
461 #define SCHEDULE_SET_MODIFIERS(s, m) \
462  (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
463 #define SCHEDULE_NONMONOTONIC 0
464 #define SCHEDULE_MONOTONIC 1
465 
467 };
468 
469 // Apply modifiers on internal kind to standard kind
470 static inline void
471 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
472  enum sched_type internal_kind) {
473  if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
474  *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
475  }
476 }
477 
478 // Apply modifiers on standard kind to internal kind
479 static inline void
480 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
481  enum sched_type *internal_kind) {
482  if ((int)kind & (int)kmp_sched_monotonic) {
483  *internal_kind = (enum sched_type)((int)*internal_kind |
485  }
486 }
487 
488 // Get standard schedule without modifiers
489 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
490  return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
491 }
492 
493 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
494 typedef union kmp_r_sched {
495  struct {
496  enum sched_type r_sched_type;
497  int chunk;
498  };
499  kmp_int64 sched;
500 } kmp_r_sched_t;
501 
502 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
503 // internal schedule types
504 
505 enum library_type {
506  library_none,
507  library_serial,
508  library_turnaround,
509  library_throughput
510 };
511 
512 #if KMP_OS_LINUX
513 enum clock_function_type {
514  clock_function_gettimeofday,
515  clock_function_clock_gettime
516 };
517 #endif /* KMP_OS_LINUX */
518 
519 #if KMP_MIC_SUPPORTED
520 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
521 #endif
522 
523 /* -- fast reduction stuff ------------------------------------------------ */
524 
525 #undef KMP_FAST_REDUCTION_BARRIER
526 #define KMP_FAST_REDUCTION_BARRIER 1
527 
528 #undef KMP_FAST_REDUCTION_CORE_DUO
529 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
530 #define KMP_FAST_REDUCTION_CORE_DUO 1
531 #endif
532 
533 enum _reduction_method {
534  reduction_method_not_defined = 0,
535  critical_reduce_block = (1 << 8),
536  atomic_reduce_block = (2 << 8),
537  tree_reduce_block = (3 << 8),
538  empty_reduce_block = (4 << 8)
539 };
540 
541 // Description of the packed_reduction_method variable:
542 // The packed_reduction_method variable consists of two enum types variables
543 // that are packed together into 0-th byte and 1-st byte:
544 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
545 // barrier that will be used in fast reduction: bs_plain_barrier or
546 // bs_reduction_barrier
547 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
548 // be used in fast reduction;
549 // Reduction method is of 'enum _reduction_method' type and it's defined the way
550 // so that the bits of 0-th byte are empty, so no need to execute a shift
551 // instruction while packing/unpacking
552 
553 #if KMP_FAST_REDUCTION_BARRIER
554 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
555  ((reduction_method) | (barrier_type))
556 
557 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
558  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
559 
560 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
561  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
562 #else
563 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
564  (reduction_method)
565 
566 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
567  (packed_reduction_method)
568 
569 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
570 #endif
571 
572 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
573  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
574  (which_reduction_block))
575 
576 #if KMP_FAST_REDUCTION_BARRIER
577 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
578  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
579 
580 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
581  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
582 #endif
583 
584 typedef int PACKED_REDUCTION_METHOD_T;
585 
586 /* -- end of fast reduction stuff ----------------------------------------- */
587 
588 #if KMP_OS_WINDOWS
589 #define USE_CBLKDATA
590 #if KMP_MSVC_COMPAT
591 #pragma warning(push)
592 #pragma warning(disable : 271 310)
593 #endif
594 #include <windows.h>
595 #if KMP_MSVC_COMPAT
596 #pragma warning(pop)
597 #endif
598 #endif
599 
600 #if KMP_OS_UNIX
601 #include <dlfcn.h>
602 #include <pthread.h>
603 #endif
604 
605 enum kmp_hw_t : int {
606  KMP_HW_UNKNOWN = -1,
607  KMP_HW_SOCKET = 0,
608  KMP_HW_PROC_GROUP,
609  KMP_HW_NUMA,
610  KMP_HW_DIE,
611  KMP_HW_LLC,
612  KMP_HW_L3,
613  KMP_HW_TILE,
614  KMP_HW_MODULE,
615  KMP_HW_L2,
616  KMP_HW_L1,
617  KMP_HW_CORE,
618  KMP_HW_THREAD,
619  KMP_HW_LAST
620 };
621 
622 typedef enum kmp_hw_core_type_t {
623  KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
624 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
625  KMP_HW_CORE_TYPE_ATOM = 0x20,
626  KMP_HW_CORE_TYPE_CORE = 0x40,
627  KMP_HW_MAX_NUM_CORE_TYPES = 3,
628 #else
629  KMP_HW_MAX_NUM_CORE_TYPES = 1,
630 #endif
631 } kmp_hw_core_type_t;
632 
633 #define KMP_HW_MAX_NUM_CORE_EFFS 8
634 
635 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
636  KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
637 #define KMP_ASSERT_VALID_HW_TYPE(type) \
638  KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
639 
640 #define KMP_FOREACH_HW_TYPE(type) \
641  for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
642  type = (kmp_hw_t)((int)type + 1))
643 
644 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
645 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
646 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
647 
648 /* Only Linux* OS and Windows* OS support thread affinity. */
649 #if KMP_AFFINITY_SUPPORTED
650 
651 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
652 #if KMP_OS_WINDOWS
653 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
654 typedef struct GROUP_AFFINITY {
655  KAFFINITY Mask;
656  WORD Group;
657  WORD Reserved[3];
658 } GROUP_AFFINITY;
659 #endif /* _MSC_VER < 1600 */
660 #if KMP_GROUP_AFFINITY
661 extern int __kmp_num_proc_groups;
662 #else
663 static const int __kmp_num_proc_groups = 1;
664 #endif /* KMP_GROUP_AFFINITY */
665 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
666 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
667 
668 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
669 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
670 
671 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
672 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
673 
674 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
675  GROUP_AFFINITY *);
676 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
677 #endif /* KMP_OS_WINDOWS */
678 
679 #if KMP_USE_HWLOC
680 extern hwloc_topology_t __kmp_hwloc_topology;
681 extern int __kmp_hwloc_error;
682 #endif
683 
684 extern size_t __kmp_affin_mask_size;
685 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
686 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
687 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
688 #define KMP_CPU_SET_ITERATE(i, mask) \
689  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
690 #define KMP_CPU_SET(i, mask) (mask)->set(i)
691 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
692 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
693 #define KMP_CPU_ZERO(mask) (mask)->zero()
694 #define KMP_CPU_ISEMPTY(mask) (mask)->empty()
695 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
696 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
697 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
698 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
699 #define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
700 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
701 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
702 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
703 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
704 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
705 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
706 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
707 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
708  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
709 #define KMP_CPU_FREE_ARRAY(arr, n) \
710  __kmp_affinity_dispatch->deallocate_mask_array(arr)
711 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
712 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
713 #define __kmp_get_system_affinity(mask, abort_bool) \
714  (mask)->get_system_affinity(abort_bool)
715 #define __kmp_set_system_affinity(mask, abort_bool) \
716  (mask)->set_system_affinity(abort_bool)
717 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
718 
719 class KMPAffinity {
720 public:
721  class Mask {
722  public:
723  void *operator new(size_t n);
724  void operator delete(void *p);
725  void *operator new[](size_t n);
726  void operator delete[](void *p);
727  virtual ~Mask() {}
728  // Set bit i to 1
729  virtual void set(int i) {}
730  // Return bit i
731  virtual bool is_set(int i) const { return false; }
732  // Set bit i to 0
733  virtual void clear(int i) {}
734  // Zero out entire mask
735  virtual void zero() {}
736  // Check whether mask is empty
737  virtual bool empty() const { return true; }
738  // Copy src into this mask
739  virtual void copy(const Mask *src) {}
740  // this &= rhs
741  virtual void bitwise_and(const Mask *rhs) {}
742  // this |= rhs
743  virtual void bitwise_or(const Mask *rhs) {}
744  // this = ~this
745  virtual void bitwise_not() {}
746  // this == rhs
747  virtual bool is_equal(const Mask *rhs) const { return false; }
748  // API for iterating over an affinity mask
749  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
750  virtual int begin() const { return 0; }
751  virtual int end() const { return 0; }
752  virtual int next(int previous) const { return 0; }
753 #if KMP_OS_WINDOWS
754  virtual int set_process_affinity(bool abort_on_error) const { return -1; }
755 #endif
756  // Set the system's affinity to this affinity mask's value
757  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
758  // Set this affinity mask to the current system affinity
759  virtual int get_system_affinity(bool abort_on_error) { return -1; }
760  // Only 1 DWORD in the mask should have any procs set.
761  // Return the appropriate index, or -1 for an invalid mask.
762  virtual int get_proc_group() const { return -1; }
763  int get_max_cpu() const {
764  int cpu;
765  int max_cpu = -1;
766  KMP_CPU_SET_ITERATE(cpu, this) {
767  if (cpu > max_cpu)
768  max_cpu = cpu;
769  }
770  return max_cpu;
771  }
772  };
773  void *operator new(size_t n);
774  void operator delete(void *p);
775  // Need virtual destructor
776  virtual ~KMPAffinity() = default;
777  // Determine if affinity is capable
778  virtual void determine_capable(const char *env_var) {}
779  // Bind the current thread to os proc
780  virtual void bind_thread(int proc) {}
781  // Factory functions to allocate/deallocate a mask
782  virtual Mask *allocate_mask() { return nullptr; }
783  virtual void deallocate_mask(Mask *m) {}
784  virtual Mask *allocate_mask_array(int num) { return nullptr; }
785  virtual void deallocate_mask_array(Mask *m) {}
786  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
787  static void pick_api();
788  static void destroy_api();
789  enum api_type {
790  NATIVE_OS
791 #if KMP_USE_HWLOC
792  ,
793  HWLOC
794 #endif
795  };
796  virtual api_type get_api_type() const {
797  KMP_ASSERT(0);
798  return NATIVE_OS;
799  }
800 
801 private:
802  static bool picked_api;
803 };
804 
805 typedef KMPAffinity::Mask kmp_affin_mask_t;
806 extern KMPAffinity *__kmp_affinity_dispatch;
807 
808 class kmp_affinity_raii_t {
809  kmp_affin_mask_t *mask;
810  bool restored;
811 
812 public:
813  kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
814  : restored(false) {
815  if (KMP_AFFINITY_CAPABLE()) {
816  KMP_CPU_ALLOC(mask);
817  KMP_ASSERT(mask != NULL);
818  __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
819  if (new_mask)
820  __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
821  }
822  }
823  void restore() {
824  if (!restored && KMP_AFFINITY_CAPABLE()) {
825  __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
826  KMP_CPU_FREE(mask);
827  }
828  restored = true;
829  }
830  ~kmp_affinity_raii_t() { restore(); }
831 };
832 
833 // Declare local char buffers with this size for printing debug and info
834 // messages, using __kmp_affinity_print_mask().
835 #define KMP_AFFIN_MASK_PRINT_LEN 1024
836 
837 enum affinity_type {
838  affinity_none = 0,
839  affinity_physical,
840  affinity_logical,
841  affinity_compact,
842  affinity_scatter,
843  affinity_explicit,
844  affinity_balanced,
845  affinity_disabled, // not used outsize the env var parser
846  affinity_default
847 };
848 
849 enum affinity_top_method {
850  affinity_top_method_all = 0, // try all (supported) methods, in order
851 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
852  affinity_top_method_apicid,
853  affinity_top_method_x2apicid,
854  affinity_top_method_x2apicid_1f,
855 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
856  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
857 #if KMP_GROUP_AFFINITY
858  affinity_top_method_group,
859 #endif /* KMP_GROUP_AFFINITY */
860  affinity_top_method_flat,
861 #if KMP_USE_HWLOC
862  affinity_top_method_hwloc,
863 #endif
864  affinity_top_method_default
865 };
866 
867 #define affinity_respect_mask_default (2)
868 
869 typedef struct kmp_affinity_flags_t {
870  unsigned dups : 1;
871  unsigned verbose : 1;
872  unsigned warnings : 1;
873  unsigned respect : 2;
874  unsigned reset : 1;
875  unsigned initialized : 1;
876  unsigned core_types_gran : 1;
877  unsigned core_effs_gran : 1;
878  unsigned omp_places : 1;
879  unsigned reserved : 22;
880 } kmp_affinity_flags_t;
881 KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
882 
883 typedef struct kmp_affinity_ids_t {
884  int ids[KMP_HW_LAST];
885  int operator[](size_t idx) const { return ids[idx]; }
886  int &operator[](size_t idx) { return ids[idx]; }
887  kmp_affinity_ids_t &operator=(const kmp_affinity_ids_t &rhs) {
888  for (int i = 0; i < KMP_HW_LAST; ++i)
889  ids[i] = rhs[i];
890  return *this;
891  }
892 } kmp_affinity_ids_t;
893 
894 typedef struct kmp_affinity_attrs_t {
895  int core_type : 8;
896  int core_eff : 8;
897  unsigned valid : 1;
898  unsigned reserved : 15;
899 } kmp_affinity_attrs_t;
900 #define KMP_AFFINITY_ATTRS_UNKNOWN \
901  { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
902 
903 typedef struct kmp_affinity_t {
904  char *proclist;
905  enum affinity_type type;
906  kmp_hw_t gran;
907  int gran_levels;
908  kmp_affinity_attrs_t core_attr_gran;
909  int compact;
910  int offset;
911  kmp_affinity_flags_t flags;
912  unsigned num_masks;
913  kmp_affin_mask_t *masks;
914  kmp_affinity_ids_t *ids;
915  kmp_affinity_attrs_t *attrs;
916  unsigned num_os_id_masks;
917  kmp_affin_mask_t *os_id_masks;
918  const char *env_var;
919 } kmp_affinity_t;
920 
921 #define KMP_AFFINITY_INIT(env) \
922  { \
923  nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
924  0, 0, \
925  {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE, \
926  FALSE, FALSE, FALSE}, \
927  0, nullptr, nullptr, nullptr, 0, nullptr, env \
928  }
929 
930 extern enum affinity_top_method __kmp_affinity_top_method;
931 extern kmp_affinity_t __kmp_affinity;
932 extern kmp_affinity_t __kmp_hh_affinity;
933 extern kmp_affinity_t *__kmp_affinities[2];
934 
935 extern void __kmp_affinity_bind_thread(int which);
936 
937 extern kmp_affin_mask_t *__kmp_affin_fullMask;
938 extern kmp_affin_mask_t *__kmp_affin_origMask;
939 extern char *__kmp_cpuinfo_file;
940 
941 #endif /* KMP_AFFINITY_SUPPORTED */
942 
943 // This needs to be kept in sync with the values in omp.h !!!
944 typedef enum kmp_proc_bind_t {
945  proc_bind_false = 0,
946  proc_bind_true,
947  proc_bind_primary,
948  proc_bind_close,
949  proc_bind_spread,
950  proc_bind_intel, // use KMP_AFFINITY interface
951  proc_bind_default
952 } kmp_proc_bind_t;
953 
954 typedef struct kmp_nested_proc_bind_t {
955  kmp_proc_bind_t *bind_types;
956  int size;
957  int used;
958 } kmp_nested_proc_bind_t;
959 
960 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
961 extern kmp_proc_bind_t __kmp_teams_proc_bind;
962 
963 extern int __kmp_display_affinity;
964 extern char *__kmp_affinity_format;
965 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
966 #if OMPT_SUPPORT
967 extern int __kmp_tool;
968 extern char *__kmp_tool_libraries;
969 #endif // OMPT_SUPPORT
970 
971 #if KMP_AFFINITY_SUPPORTED
972 #define KMP_PLACE_ALL (-1)
973 #define KMP_PLACE_UNDEFINED (-2)
974 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
975 #define KMP_AFFINITY_NON_PROC_BIND \
976  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
977  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
978  (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
979 #endif /* KMP_AFFINITY_SUPPORTED */
980 
981 extern int __kmp_affinity_num_places;
982 
983 typedef enum kmp_cancel_kind_t {
984  cancel_noreq = 0,
985  cancel_parallel = 1,
986  cancel_loop = 2,
987  cancel_sections = 3,
988  cancel_taskgroup = 4
989 } kmp_cancel_kind_t;
990 
991 // KMP_HW_SUBSET support:
992 typedef struct kmp_hws_item {
993  int num;
994  int offset;
995 } kmp_hws_item_t;
996 
997 extern kmp_hws_item_t __kmp_hws_socket;
998 extern kmp_hws_item_t __kmp_hws_die;
999 extern kmp_hws_item_t __kmp_hws_node;
1000 extern kmp_hws_item_t __kmp_hws_tile;
1001 extern kmp_hws_item_t __kmp_hws_core;
1002 extern kmp_hws_item_t __kmp_hws_proc;
1003 extern int __kmp_hws_requested;
1004 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1005 
1006 /* ------------------------------------------------------------------------ */
1007 
1008 #define KMP_PAD(type, sz) \
1009  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1010 
1011 // We need to avoid using -1 as a GTID as +1 is added to the gtid
1012 // when storing it in a lock, and the value 0 is reserved.
1013 #define KMP_GTID_DNE (-2) /* Does not exist */
1014 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1015 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1016 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
1017 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1018 
1019 /* OpenMP 5.0 Memory Management support */
1020 
1021 #ifndef __OMP_H
1022 // Duplicate type definitions from omp.h
1023 typedef uintptr_t omp_uintptr_t;
1024 
1025 typedef enum {
1026  omp_atk_sync_hint = 1,
1027  omp_atk_alignment = 2,
1028  omp_atk_access = 3,
1029  omp_atk_pool_size = 4,
1030  omp_atk_fallback = 5,
1031  omp_atk_fb_data = 6,
1032  omp_atk_pinned = 7,
1033  omp_atk_partition = 8
1034 } omp_alloctrait_key_t;
1035 
1036 typedef enum {
1037  omp_atv_false = 0,
1038  omp_atv_true = 1,
1039  omp_atv_contended = 3,
1040  omp_atv_uncontended = 4,
1041  omp_atv_serialized = 5,
1042  omp_atv_sequential = omp_atv_serialized, // (deprecated)
1043  omp_atv_private = 6,
1044  omp_atv_all = 7,
1045  omp_atv_thread = 8,
1046  omp_atv_pteam = 9,
1047  omp_atv_cgroup = 10,
1048  omp_atv_default_mem_fb = 11,
1049  omp_atv_null_fb = 12,
1050  omp_atv_abort_fb = 13,
1051  omp_atv_allocator_fb = 14,
1052  omp_atv_environment = 15,
1053  omp_atv_nearest = 16,
1054  omp_atv_blocked = 17,
1055  omp_atv_interleaved = 18
1056 } omp_alloctrait_value_t;
1057 #define omp_atv_default ((omp_uintptr_t)-1)
1058 
1059 typedef void *omp_memspace_handle_t;
1060 extern omp_memspace_handle_t const omp_default_mem_space;
1061 extern omp_memspace_handle_t const omp_large_cap_mem_space;
1062 extern omp_memspace_handle_t const omp_const_mem_space;
1063 extern omp_memspace_handle_t const omp_high_bw_mem_space;
1064 extern omp_memspace_handle_t const omp_low_lat_mem_space;
1065 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1066 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1067 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1068 
1069 typedef struct {
1070  omp_alloctrait_key_t key;
1071  omp_uintptr_t value;
1072 } omp_alloctrait_t;
1073 
1074 typedef void *omp_allocator_handle_t;
1075 extern omp_allocator_handle_t const omp_null_allocator;
1076 extern omp_allocator_handle_t const omp_default_mem_alloc;
1077 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1078 extern omp_allocator_handle_t const omp_const_mem_alloc;
1079 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1080 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1081 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1082 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1083 extern omp_allocator_handle_t const omp_thread_mem_alloc;
1084 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1085 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1086 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1087 extern omp_allocator_handle_t const kmp_max_mem_alloc;
1088 extern omp_allocator_handle_t __kmp_def_allocator;
1089 
1090 // end of duplicate type definitions from omp.h
1091 #endif
1092 
1093 extern int __kmp_memkind_available;
1094 
1095 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1096 
1097 typedef struct kmp_allocator_t {
1098  omp_memspace_handle_t memspace;
1099  void **memkind; // pointer to memkind
1100  size_t alignment;
1101  omp_alloctrait_value_t fb;
1102  kmp_allocator_t *fb_data;
1103  kmp_uint64 pool_size;
1104  kmp_uint64 pool_used;
1105  bool pinned;
1106 } kmp_allocator_t;
1107 
1108 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1109  omp_memspace_handle_t,
1110  int ntraits,
1111  omp_alloctrait_t traits[]);
1112 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1113 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1114 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1115 // external interfaces, may be used by compiler
1116 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1117 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1118  omp_allocator_handle_t al);
1119 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1120  omp_allocator_handle_t al);
1121 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1122  omp_allocator_handle_t al,
1123  omp_allocator_handle_t free_al);
1124 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1125 // internal interfaces, contain real implementation
1126 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1127  omp_allocator_handle_t al);
1128 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1129  omp_allocator_handle_t al);
1130 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1131  omp_allocator_handle_t al,
1132  omp_allocator_handle_t free_al);
1133 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1134 
1135 extern void __kmp_init_memkind();
1136 extern void __kmp_fini_memkind();
1137 extern void __kmp_init_target_mem();
1138 
1139 /* ------------------------------------------------------------------------ */
1140 
1141 #if ENABLE_LIBOMPTARGET
1142 extern void __kmp_init_target_task();
1143 #endif
1144 
1145 /* ------------------------------------------------------------------------ */
1146 
1147 #define KMP_UINT64_MAX \
1148  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1149 
1150 #define KMP_MIN_NTH 1
1151 
1152 #ifndef KMP_MAX_NTH
1153 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1154 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1155 #else
1156 #define KMP_MAX_NTH INT_MAX
1157 #endif
1158 #endif /* KMP_MAX_NTH */
1159 
1160 #ifdef PTHREAD_STACK_MIN
1161 #define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1162 #else
1163 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1164 #endif
1165 
1166 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1167 
1168 #if KMP_ARCH_X86
1169 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1170 #elif KMP_ARCH_X86_64
1171 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1172 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1173 #elif KMP_ARCH_VE
1174 // Minimum stack size for pthread for VE is 4MB.
1175 // https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1176 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1177 #else
1178 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1179 #endif
1180 
1181 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1182 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1183 #define KMP_MAX_MALLOC_POOL_INCR \
1184  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1185 
1186 #define KMP_MIN_STKOFFSET (0)
1187 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1188 #if KMP_OS_DARWIN
1189 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1190 #else
1191 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1192 #endif
1193 
1194 #define KMP_MIN_STKPADDING (0)
1195 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1196 
1197 #define KMP_BLOCKTIME_MULTIPLIER \
1198  (1000000) /* number of blocktime units per second */
1199 #define KMP_MIN_BLOCKTIME (0)
1200 #define KMP_MAX_BLOCKTIME \
1201  (INT_MAX) /* Must be this for "infinite" setting the work */
1202 
1203 /* __kmp_blocktime is in microseconds */
1204 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1205 
1206 #if KMP_USE_MONITOR
1207 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1208 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1209 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1210 
1211 /* Calculate new number of monitor wakeups for a specific block time based on
1212  previous monitor_wakeups. Only allow increasing number of wakeups */
1213 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1214  (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1215  : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1216  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1217  ? (monitor_wakeups) \
1218  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1219 
1220 /* Calculate number of intervals for a specific block time based on
1221  monitor_wakeups */
1222 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1223  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1224  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1225 #else
1226 #define KMP_BLOCKTIME(team, tid) \
1227  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1228 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1229 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1230 extern kmp_uint64 __kmp_ticks_per_msec;
1231 extern kmp_uint64 __kmp_ticks_per_usec;
1232 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1233 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1234 #else
1235 #define KMP_NOW() __kmp_hardware_timestamp()
1236 #endif
1237 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1238  ((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1239 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1240 #else
1241 // System time is retrieved sporadically while blocking.
1242 extern kmp_uint64 __kmp_now_nsec();
1243 #define KMP_NOW() __kmp_now_nsec()
1244 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1245  ((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1246 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1247 #endif
1248 #endif // KMP_USE_MONITOR
1249 
1250 #define KMP_MIN_STATSCOLS 40
1251 #define KMP_MAX_STATSCOLS 4096
1252 #define KMP_DEFAULT_STATSCOLS 80
1253 
1254 #define KMP_MIN_INTERVAL 0
1255 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1256 #define KMP_DEFAULT_INTERVAL 0
1257 
1258 #define KMP_MIN_CHUNK 1
1259 #define KMP_MAX_CHUNK (INT_MAX - 1)
1260 #define KMP_DEFAULT_CHUNK 1
1261 
1262 #define KMP_MIN_DISP_NUM_BUFF 1
1263 #define KMP_DFLT_DISP_NUM_BUFF 7
1264 #define KMP_MAX_DISP_NUM_BUFF 4096
1265 
1266 #define KMP_MAX_ORDERED 8
1267 
1268 #define KMP_MAX_FIELDS 32
1269 
1270 #define KMP_MAX_BRANCH_BITS 31
1271 
1272 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1273 
1274 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1275 
1276 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1277 
1278 /* Minimum number of threads before switch to TLS gtid (experimentally
1279  determined) */
1280 /* josh TODO: what about OS X* tuning? */
1281 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1282 #define KMP_TLS_GTID_MIN 5
1283 #else
1284 #define KMP_TLS_GTID_MIN INT_MAX
1285 #endif
1286 
1287 #define KMP_MASTER_TID(tid) (0 == (tid))
1288 #define KMP_WORKER_TID(tid) (0 != (tid))
1289 
1290 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1291 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1292 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1293 
1294 #ifndef TRUE
1295 #define FALSE 0
1296 #define TRUE (!FALSE)
1297 #endif
1298 
1299 /* NOTE: all of the following constants must be even */
1300 
1301 #if KMP_OS_WINDOWS
1302 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1303 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1304 #elif KMP_OS_LINUX
1305 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1306 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1307 #elif KMP_OS_DARWIN
1308 /* TODO: tune for KMP_OS_DARWIN */
1309 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1310 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1311 #elif KMP_OS_DRAGONFLY
1312 /* TODO: tune for KMP_OS_DRAGONFLY */
1313 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1314 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1315 #elif KMP_OS_FREEBSD
1316 /* TODO: tune for KMP_OS_FREEBSD */
1317 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1318 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1319 #elif KMP_OS_NETBSD
1320 /* TODO: tune for KMP_OS_NETBSD */
1321 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1322 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1323 #elif KMP_OS_HURD
1324 /* TODO: tune for KMP_OS_HURD */
1325 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1326 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1327 #elif KMP_OS_OPENBSD
1328 /* TODO: tune for KMP_OS_OPENBSD */
1329 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1330 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1331 #endif
1332 
1333 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1334 typedef struct kmp_cpuid {
1335  kmp_uint32 eax;
1336  kmp_uint32 ebx;
1337  kmp_uint32 ecx;
1338  kmp_uint32 edx;
1339 } kmp_cpuid_t;
1340 
1341 typedef struct kmp_cpuinfo_flags_t {
1342  unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1343  unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1344  unsigned hybrid : 1;
1345  unsigned reserved : 29; // Ensure size of 32 bits
1346 } kmp_cpuinfo_flags_t;
1347 
1348 typedef struct kmp_cpuinfo {
1349  int initialized; // If 0, other fields are not initialized.
1350  int signature; // CPUID(1).EAX
1351  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1352  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1353  // Model << 4 ) + Model)
1354  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1355  kmp_cpuinfo_flags_t flags;
1356  int apic_id;
1357  int physical_id;
1358  int logical_id;
1359  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1360  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1361 } kmp_cpuinfo_t;
1362 
1363 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1364 
1365 #if KMP_OS_UNIX
1366 // subleaf is only needed for cache and topology discovery and can be set to
1367 // zero in most cases
1368 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1369  __asm__ __volatile__("cpuid"
1370  : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1371  : "a"(leaf), "c"(subleaf));
1372 }
1373 // Load p into FPU control word
1374 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1375  __asm__ __volatile__("fldcw %0" : : "m"(*p));
1376 }
1377 // Store FPU control word into p
1378 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1379  __asm__ __volatile__("fstcw %0" : "=m"(*p));
1380 }
1381 static inline void __kmp_clear_x87_fpu_status_word() {
1382 #if KMP_MIC
1383  // 32-bit protected mode x87 FPU state
1384  struct x87_fpu_state {
1385  unsigned cw;
1386  unsigned sw;
1387  unsigned tw;
1388  unsigned fip;
1389  unsigned fips;
1390  unsigned fdp;
1391  unsigned fds;
1392  };
1393  struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1394  __asm__ __volatile__("fstenv %0\n\t" // store FP env
1395  "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1396  "fldenv %0\n\t" // load FP env back
1397  : "+m"(fpu_state), "+m"(fpu_state.sw));
1398 #else
1399  __asm__ __volatile__("fnclex");
1400 #endif // KMP_MIC
1401 }
1402 #if __SSE__
1403 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1404 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1405 #else
1406 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1407 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1408 #endif
1409 #else
1410 // Windows still has these as external functions in assembly file
1411 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1412 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1413 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1414 extern void __kmp_clear_x87_fpu_status_word();
1415 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1416 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1417 #endif // KMP_OS_UNIX
1418 
1419 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1420 
1421 // User-level Monitor/Mwait
1422 #if KMP_HAVE_UMWAIT
1423 // We always try for UMWAIT first
1424 #if KMP_HAVE_WAITPKG_INTRINSICS
1425 #if KMP_HAVE_IMMINTRIN_H
1426 #include <immintrin.h>
1427 #elif KMP_HAVE_INTRIN_H
1428 #include <intrin.h>
1429 #endif
1430 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1431 
1432 KMP_ATTRIBUTE_TARGET_WAITPKG
1433 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1434 #if !KMP_HAVE_WAITPKG_INTRINSICS
1435  uint32_t timeHi = uint32_t(counter >> 32);
1436  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1437  char flag;
1438  __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1439  "setb %0"
1440  // The "=q" restraint means any register accessible as rl
1441  // in 32-bit mode: a, b, c, and d;
1442  // in 64-bit mode: any integer register
1443  : "=q"(flag)
1444  : "a"(timeLo), "d"(timeHi), "c"(hint)
1445  :);
1446  return flag;
1447 #else
1448  return _tpause(hint, counter);
1449 #endif
1450 }
1451 KMP_ATTRIBUTE_TARGET_WAITPKG
1452 static inline void __kmp_umonitor(void *cacheline) {
1453 #if !KMP_HAVE_WAITPKG_INTRINSICS
1454  __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1455  :
1456  : "a"(cacheline)
1457  :);
1458 #else
1459  _umonitor(cacheline);
1460 #endif
1461 }
1462 KMP_ATTRIBUTE_TARGET_WAITPKG
1463 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1464 #if !KMP_HAVE_WAITPKG_INTRINSICS
1465  uint32_t timeHi = uint32_t(counter >> 32);
1466  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1467  char flag;
1468  __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1469  "setb %0"
1470  // The "=q" restraint means any register accessible as rl
1471  // in 32-bit mode: a, b, c, and d;
1472  // in 64-bit mode: any integer register
1473  : "=q"(flag)
1474  : "a"(timeLo), "d"(timeHi), "c"(hint)
1475  :);
1476  return flag;
1477 #else
1478  return _umwait(hint, counter);
1479 #endif
1480 }
1481 #elif KMP_HAVE_MWAIT
1482 #if KMP_OS_UNIX
1483 #include <pmmintrin.h>
1484 #else
1485 #include <intrin.h>
1486 #endif
1487 #if KMP_OS_UNIX
1488 __attribute__((target("sse3")))
1489 #endif
1490 static inline void
1491 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1492  _mm_monitor(cacheline, extensions, hints);
1493 }
1494 #if KMP_OS_UNIX
1495 __attribute__((target("sse3")))
1496 #endif
1497 static inline void
1498 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1499  _mm_mwait(extensions, hints);
1500 }
1501 #endif // KMP_HAVE_UMWAIT
1502 
1503 #if KMP_ARCH_X86
1504 extern void __kmp_x86_pause(void);
1505 #elif KMP_MIC
1506 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1507 // regression after removal of extra PAUSE from spin loops. Changing
1508 // the delay from 100 to 300 showed even better performance than double PAUSE
1509 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1510 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1511 #else
1512 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1513 #endif
1514 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1515 #elif KMP_ARCH_PPC64
1516 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1517 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1518 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1519 #define KMP_CPU_PAUSE() \
1520  do { \
1521  KMP_PPC64_PRI_LOW(); \
1522  KMP_PPC64_PRI_MED(); \
1523  KMP_PPC64_PRI_LOC_MB(); \
1524  } while (0)
1525 #else
1526 #define KMP_CPU_PAUSE() /* nothing to do */
1527 #endif
1528 
1529 #define KMP_INIT_YIELD(count) \
1530  { (count) = __kmp_yield_init; }
1531 
1532 #define KMP_INIT_BACKOFF(time) \
1533  { (time) = __kmp_pause_init; }
1534 
1535 #define KMP_OVERSUBSCRIBED \
1536  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1537 
1538 #define KMP_TRY_YIELD \
1539  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1540 
1541 #define KMP_TRY_YIELD_OVERSUB \
1542  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1543 
1544 #define KMP_YIELD(cond) \
1545  { \
1546  KMP_CPU_PAUSE(); \
1547  if ((cond) && (KMP_TRY_YIELD)) \
1548  __kmp_yield(); \
1549  }
1550 
1551 #define KMP_YIELD_OVERSUB() \
1552  { \
1553  KMP_CPU_PAUSE(); \
1554  if ((KMP_TRY_YIELD_OVERSUB)) \
1555  __kmp_yield(); \
1556  }
1557 
1558 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1559 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1560 #define KMP_YIELD_SPIN(count) \
1561  { \
1562  KMP_CPU_PAUSE(); \
1563  if (KMP_TRY_YIELD) { \
1564  (count) -= 2; \
1565  if (!(count)) { \
1566  __kmp_yield(); \
1567  (count) = __kmp_yield_next; \
1568  } \
1569  } \
1570  }
1571 
1572 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1573 // (C0.2) state, which improves performance of other SMT threads on the same
1574 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1575 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1576 // available, fall back to the regular CPU pause and yield combination.
1577 #if KMP_HAVE_UMWAIT
1578 #define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1579 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1580  { \
1581  if (__kmp_tpause_enabled) { \
1582  if (KMP_OVERSUBSCRIBED) { \
1583  __kmp_tpause(0, (time)); \
1584  } else { \
1585  __kmp_tpause(__kmp_tpause_hint, (time)); \
1586  } \
1587  (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1588  } else { \
1589  KMP_CPU_PAUSE(); \
1590  if ((KMP_TRY_YIELD_OVERSUB)) { \
1591  __kmp_yield(); \
1592  } else if (__kmp_use_yield == 1) { \
1593  (count) -= 2; \
1594  if (!(count)) { \
1595  __kmp_yield(); \
1596  (count) = __kmp_yield_next; \
1597  } \
1598  } \
1599  } \
1600  }
1601 #else
1602 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1603  { \
1604  KMP_CPU_PAUSE(); \
1605  if ((KMP_TRY_YIELD_OVERSUB)) \
1606  __kmp_yield(); \
1607  else if (__kmp_use_yield == 1) { \
1608  (count) -= 2; \
1609  if (!(count)) { \
1610  __kmp_yield(); \
1611  (count) = __kmp_yield_next; \
1612  } \
1613  } \
1614  }
1615 #endif // KMP_HAVE_UMWAIT
1616 
1617 /* ------------------------------------------------------------------------ */
1618 /* Support datatypes for the orphaned construct nesting checks. */
1619 /* ------------------------------------------------------------------------ */
1620 
1621 /* When adding to this enum, add its corresponding string in cons_text_c[]
1622  * array in kmp_error.cpp */
1623 enum cons_type {
1624  ct_none,
1625  ct_parallel,
1626  ct_pdo,
1627  ct_pdo_ordered,
1628  ct_psections,
1629  ct_psingle,
1630  ct_critical,
1631  ct_ordered_in_parallel,
1632  ct_ordered_in_pdo,
1633  ct_master,
1634  ct_reduce,
1635  ct_barrier,
1636  ct_masked
1637 };
1638 
1639 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1640 
1641 struct cons_data {
1642  ident_t const *ident;
1643  enum cons_type type;
1644  int prev;
1645  kmp_user_lock_p
1646  name; /* address exclusively for critical section name comparison */
1647 };
1648 
1649 struct cons_header {
1650  int p_top, w_top, s_top;
1651  int stack_size, stack_top;
1652  struct cons_data *stack_data;
1653 };
1654 
1655 struct kmp_region_info {
1656  char *text;
1657  int offset[KMP_MAX_FIELDS];
1658  int length[KMP_MAX_FIELDS];
1659 };
1660 
1661 /* ---------------------------------------------------------------------- */
1662 /* ---------------------------------------------------------------------- */
1663 
1664 #if KMP_OS_WINDOWS
1665 typedef HANDLE kmp_thread_t;
1666 typedef DWORD kmp_key_t;
1667 #endif /* KMP_OS_WINDOWS */
1668 
1669 #if KMP_OS_UNIX
1670 typedef pthread_t kmp_thread_t;
1671 typedef pthread_key_t kmp_key_t;
1672 #endif
1673 
1674 extern kmp_key_t __kmp_gtid_threadprivate_key;
1675 
1676 typedef struct kmp_sys_info {
1677  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1678  long minflt; /* the number of page faults serviced without any I/O */
1679  long majflt; /* the number of page faults serviced that required I/O */
1680  long nswap; /* the number of times a process was "swapped" out of memory */
1681  long inblock; /* the number of times the file system had to perform input */
1682  long oublock; /* the number of times the file system had to perform output */
1683  long nvcsw; /* the number of times a context switch was voluntarily */
1684  long nivcsw; /* the number of times a context switch was forced */
1685 } kmp_sys_info_t;
1686 
1687 #if USE_ITT_BUILD
1688 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1689 // required type here. Later we will check the type meets requirements.
1690 typedef int kmp_itt_mark_t;
1691 #define KMP_ITT_DEBUG 0
1692 #endif /* USE_ITT_BUILD */
1693 
1694 typedef kmp_int32 kmp_critical_name[8];
1695 
1705 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1706 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1707  ...);
1708 
1713 /* ---------------------------------------------------------------------------
1714  */
1715 /* Threadprivate initialization/finalization function declarations */
1716 
1717 /* for non-array objects: __kmpc_threadprivate_register() */
1718 
1723 typedef void *(*kmpc_ctor)(void *);
1724 
1729 typedef void (*kmpc_dtor)(
1730  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1731  compiler */
1736 typedef void *(*kmpc_cctor)(void *, void *);
1737 
1738 /* for array objects: __kmpc_threadprivate_register_vec() */
1739 /* First arg: "this" pointer */
1740 /* Last arg: number of array elements */
1746 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1752 typedef void (*kmpc_dtor_vec)(void *, size_t);
1758 typedef void *(*kmpc_cctor_vec)(void *, void *,
1759  size_t); /* function unused by compiler */
1760 
1765 /* keeps tracked of threadprivate cache allocations for cleanup later */
1766 typedef struct kmp_cached_addr {
1767  void **addr; /* address of allocated cache */
1768  void ***compiler_cache; /* pointer to compiler's cache */
1769  void *data; /* pointer to global data */
1770  struct kmp_cached_addr *next; /* pointer to next cached address */
1771 } kmp_cached_addr_t;
1772 
1773 struct private_data {
1774  struct private_data *next; /* The next descriptor in the list */
1775  void *data; /* The data buffer for this descriptor */
1776  int more; /* The repeat count for this descriptor */
1777  size_t size; /* The data size for this descriptor */
1778 };
1779 
1780 struct private_common {
1781  struct private_common *next;
1782  struct private_common *link;
1783  void *gbl_addr;
1784  void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1785  size_t cmn_size;
1786 };
1787 
1788 struct shared_common {
1789  struct shared_common *next;
1790  struct private_data *pod_init;
1791  void *obj_init;
1792  void *gbl_addr;
1793  union {
1794  kmpc_ctor ctor;
1795  kmpc_ctor_vec ctorv;
1796  } ct;
1797  union {
1798  kmpc_cctor cctor;
1799  kmpc_cctor_vec cctorv;
1800  } cct;
1801  union {
1802  kmpc_dtor dtor;
1803  kmpc_dtor_vec dtorv;
1804  } dt;
1805  size_t vec_len;
1806  int is_vec;
1807  size_t cmn_size;
1808 };
1809 
1810 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1811 #define KMP_HASH_TABLE_SIZE \
1812  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1813 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1814 #define KMP_HASH(x) \
1815  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1816 
1817 struct common_table {
1818  struct private_common *data[KMP_HASH_TABLE_SIZE];
1819 };
1820 
1821 struct shared_table {
1822  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1823 };
1824 
1825 /* ------------------------------------------------------------------------ */
1826 
1827 #if KMP_USE_HIER_SCHED
1828 // Shared barrier data that exists inside a single unit of the scheduling
1829 // hierarchy
1830 typedef struct kmp_hier_private_bdata_t {
1831  kmp_int32 num_active;
1832  kmp_uint64 index;
1833  kmp_uint64 wait_val[2];
1834 } kmp_hier_private_bdata_t;
1835 #endif
1836 
1837 typedef struct kmp_sched_flags {
1838  unsigned ordered : 1;
1839  unsigned nomerge : 1;
1840  unsigned contains_last : 1;
1841 #if KMP_USE_HIER_SCHED
1842  unsigned use_hier : 1;
1843  unsigned unused : 28;
1844 #else
1845  unsigned unused : 29;
1846 #endif
1847 } kmp_sched_flags_t;
1848 
1849 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1850 
1851 #if KMP_STATIC_STEAL_ENABLED
1852 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1853  kmp_int32 count;
1854  kmp_int32 ub;
1855  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1856  kmp_int32 lb;
1857  kmp_int32 st;
1858  kmp_int32 tc;
1859  kmp_lock_t *steal_lock; // lock used for chunk stealing
1860  // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1861  // a) parm3 is properly aligned and
1862  // b) all parm1-4 are on the same cache line.
1863  // Because of parm1-4 are used together, performance seems to be better
1864  // if they are on the same cache line (not measured though).
1865 
1866  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1867  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1868  kmp_int32 parm2; // make no real change at least while padding is off.
1869  kmp_int32 parm3;
1870  kmp_int32 parm4;
1871  };
1872 
1873  kmp_uint32 ordered_lower;
1874  kmp_uint32 ordered_upper;
1875 #if KMP_OS_WINDOWS
1876  kmp_int32 last_upper;
1877 #endif /* KMP_OS_WINDOWS */
1878 } dispatch_private_info32_t;
1879 
1880 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1881  kmp_int64 count; // current chunk number for static & static-steal scheduling
1882  kmp_int64 ub; /* upper-bound */
1883  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1884  kmp_int64 lb; /* lower-bound */
1885  kmp_int64 st; /* stride */
1886  kmp_int64 tc; /* trip count (number of iterations) */
1887  kmp_lock_t *steal_lock; // lock used for chunk stealing
1888  /* parm[1-4] are used in different ways by different scheduling algorithms */
1889 
1890  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1891  // a) parm3 is properly aligned and
1892  // b) all parm1-4 are in the same cache line.
1893  // Because of parm1-4 are used together, performance seems to be better
1894  // if they are in the same line (not measured though).
1895 
1896  struct KMP_ALIGN(32) {
1897  kmp_int64 parm1;
1898  kmp_int64 parm2;
1899  kmp_int64 parm3;
1900  kmp_int64 parm4;
1901  };
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 #else /* KMP_STATIC_STEAL_ENABLED */
1910 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1911  kmp_int32 lb;
1912  kmp_int32 ub;
1913  kmp_int32 st;
1914  kmp_int32 tc;
1915 
1916  kmp_int32 parm1;
1917  kmp_int32 parm2;
1918  kmp_int32 parm3;
1919  kmp_int32 parm4;
1920 
1921  kmp_int32 count;
1922 
1923  kmp_uint32 ordered_lower;
1924  kmp_uint32 ordered_upper;
1925 #if KMP_OS_WINDOWS
1926  kmp_int32 last_upper;
1927 #endif /* KMP_OS_WINDOWS */
1928 } dispatch_private_info32_t;
1929 
1930 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1931  kmp_int64 lb; /* lower-bound */
1932  kmp_int64 ub; /* upper-bound */
1933  kmp_int64 st; /* stride */
1934  kmp_int64 tc; /* trip count (number of iterations) */
1935 
1936  /* parm[1-4] are used in different ways by different scheduling algorithms */
1937  kmp_int64 parm1;
1938  kmp_int64 parm2;
1939  kmp_int64 parm3;
1940  kmp_int64 parm4;
1941 
1942  kmp_int64 count; /* current chunk number for static scheduling */
1943 
1944  kmp_uint64 ordered_lower;
1945  kmp_uint64 ordered_upper;
1946 #if KMP_OS_WINDOWS
1947  kmp_int64 last_upper;
1948 #endif /* KMP_OS_WINDOWS */
1949 } dispatch_private_info64_t;
1950 #endif /* KMP_STATIC_STEAL_ENABLED */
1951 
1952 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1953  union private_info {
1954  dispatch_private_info32_t p32;
1955  dispatch_private_info64_t p64;
1956  } u;
1957  enum sched_type schedule; /* scheduling algorithm */
1958  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1959  std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1960  kmp_int32 ordered_bumped;
1961  // Stack of buffers for nest of serial regions
1962  struct dispatch_private_info *next;
1963  kmp_int32 type_size; /* the size of types in private_info */
1964 #if KMP_USE_HIER_SCHED
1965  kmp_int32 hier_id;
1966  void *parent; /* hierarchical scheduling parent pointer */
1967 #endif
1968  enum cons_type pushed_ws;
1969 } dispatch_private_info_t;
1970 
1971 typedef struct dispatch_shared_info32 {
1972  /* chunk index under dynamic, number of idle threads under static-steal;
1973  iteration index otherwise */
1974  volatile kmp_uint32 iteration;
1975  volatile kmp_int32 num_done;
1976  volatile kmp_uint32 ordered_iteration;
1977  // Dummy to retain the structure size after making ordered_iteration scalar
1978  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1979 } dispatch_shared_info32_t;
1980 
1981 typedef struct dispatch_shared_info64 {
1982  /* chunk index under dynamic, number of idle threads under static-steal;
1983  iteration index otherwise */
1984  volatile kmp_uint64 iteration;
1985  volatile kmp_int64 num_done;
1986  volatile kmp_uint64 ordered_iteration;
1987  // Dummy to retain the structure size after making ordered_iteration scalar
1988  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1989 } dispatch_shared_info64_t;
1990 
1991 typedef struct dispatch_shared_info {
1992  union shared_info {
1993  dispatch_shared_info32_t s32;
1994  dispatch_shared_info64_t s64;
1995  } u;
1996  volatile kmp_uint32 buffer_index;
1997  volatile kmp_int32 doacross_buf_idx; // teamwise index
1998  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1999  kmp_int32 doacross_num_done; // count finished threads
2000 #if KMP_USE_HIER_SCHED
2001  void *hier;
2002 #endif
2003 #if KMP_USE_HWLOC
2004  // When linking with libhwloc, the ORDERED EPCC test slows down on big
2005  // machines (> 48 cores). Performance analysis showed that a cache thrash
2006  // was occurring and this padding helps alleviate the problem.
2007  char padding[64];
2008 #endif
2009 } dispatch_shared_info_t;
2010 
2011 typedef struct kmp_disp {
2012  /* Vector for ORDERED SECTION */
2013  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2014  /* Vector for END ORDERED SECTION */
2015  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2016 
2017  dispatch_shared_info_t *th_dispatch_sh_current;
2018  dispatch_private_info_t *th_dispatch_pr_current;
2019 
2020  dispatch_private_info_t *th_disp_buffer;
2021  kmp_uint32 th_disp_index;
2022  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2023  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2024  kmp_int64 *th_doacross_info; // info on loop bounds
2025 #if KMP_USE_INTERNODE_ALIGNMENT
2026  char more_padding[INTERNODE_CACHE_LINE];
2027 #endif
2028 } kmp_disp_t;
2029 
2030 /* ------------------------------------------------------------------------ */
2031 /* Barrier stuff */
2032 
2033 /* constants for barrier state update */
2034 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2035 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2036 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2037 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2038 
2039 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2040 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2041 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2042 
2043 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2044 #error "Barrier sleep bit must be smaller than barrier bump bit"
2045 #endif
2046 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2047 #error "Barrier unused bit must be smaller than barrier bump bit"
2048 #endif
2049 
2050 // Constants for release barrier wait state: currently, hierarchical only
2051 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2052 #define KMP_BARRIER_OWN_FLAG \
2053  1 // Normal state; worker waiting on own b_go flag in release
2054 #define KMP_BARRIER_PARENT_FLAG \
2055  2 // Special state; worker waiting on parent's b_go flag in release
2056 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2057  3 // Special state; tells worker to shift from parent to own b_go
2058 #define KMP_BARRIER_SWITCHING \
2059  4 // Special state; worker resets appropriate flag on wake-up
2060 
2061 #define KMP_NOT_SAFE_TO_REAP \
2062  0 // Thread th_reap_state: not safe to reap (tasking)
2063 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2064 
2065 // The flag_type describes the storage used for the flag.
2066 enum flag_type {
2067  flag32,
2068  flag64,
2069  atomic_flag64,
2070  flag_oncore,
2071  flag_unset
2072 };
2073 
2074 enum barrier_type {
2075  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2076  barriers if enabled) */
2077  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2078 #if KMP_FAST_REDUCTION_BARRIER
2079  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2080 #endif // KMP_FAST_REDUCTION_BARRIER
2081  bs_last_barrier /* Just a placeholder to mark the end */
2082 };
2083 
2084 // to work with reduction barriers just like with plain barriers
2085 #if !KMP_FAST_REDUCTION_BARRIER
2086 #define bs_reduction_barrier bs_plain_barrier
2087 #endif // KMP_FAST_REDUCTION_BARRIER
2088 
2089 typedef enum kmp_bar_pat { /* Barrier communication patterns */
2090  bp_linear_bar =
2091  0, /* Single level (degenerate) tree */
2092  bp_tree_bar =
2093  1, /* Balanced tree with branching factor 2^n */
2094  bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2095  branching factor 2^n */
2096  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2097  bp_dist_bar = 4, /* Distributed barrier */
2098  bp_last_bar /* Placeholder to mark the end */
2099 } kmp_bar_pat_e;
2100 
2101 #define KMP_BARRIER_ICV_PUSH 1
2102 
2103 /* Record for holding the values of the internal controls stack records */
2104 typedef struct kmp_internal_control {
2105  int serial_nesting_level; /* corresponds to the value of the
2106  th_team_serialized field */
2107  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2108  thread) */
2109  kmp_int8
2110  bt_set; /* internal control for whether blocktime is explicitly set */
2111  int blocktime; /* internal control for blocktime */
2112 #if KMP_USE_MONITOR
2113  int bt_intervals; /* internal control for blocktime intervals */
2114 #endif
2115  int nproc; /* internal control for #threads for next parallel region (per
2116  thread) */
2117  int thread_limit; /* internal control for thread-limit-var */
2118  int task_thread_limit; /* internal control for thread-limit-var of a task*/
2119  int max_active_levels; /* internal control for max_active_levels */
2120  kmp_r_sched_t
2121  sched; /* internal control for runtime schedule {sched,chunk} pair */
2122  kmp_proc_bind_t proc_bind; /* internal control for affinity */
2123  kmp_int32 default_device; /* internal control for default device */
2124  struct kmp_internal_control *next;
2125 } kmp_internal_control_t;
2126 
2127 static inline void copy_icvs(kmp_internal_control_t *dst,
2128  kmp_internal_control_t *src) {
2129  *dst = *src;
2130 }
2131 
2132 /* Thread barrier needs volatile barrier fields */
2133 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2134  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2135  // uses of it). It is not explicitly aligned below, because we *don't* want
2136  // it to be padded -- instead, we fit b_go into the same cache line with
2137  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2138  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2139  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2140  // same NGO store
2141  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2142  KMP_ALIGN_CACHE volatile kmp_uint64
2143  b_arrived; // STATE => task reached synch point.
2144  kmp_uint32 *skip_per_level;
2145  kmp_uint32 my_level;
2146  kmp_int32 parent_tid;
2147  kmp_int32 old_tid;
2148  kmp_uint32 depth;
2149  struct kmp_bstate *parent_bar;
2150  kmp_team_t *team;
2151  kmp_uint64 leaf_state;
2152  kmp_uint32 nproc;
2153  kmp_uint8 base_leaf_kids;
2154  kmp_uint8 leaf_kids;
2155  kmp_uint8 offset;
2156  kmp_uint8 wait_flag;
2157  kmp_uint8 use_oncore_barrier;
2158 #if USE_DEBUGGER
2159  // The following field is intended for the debugger solely. Only the worker
2160  // thread itself accesses this field: the worker increases it by 1 when it
2161  // arrives to a barrier.
2162  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2163 #endif /* USE_DEBUGGER */
2164 } kmp_bstate_t;
2165 
2166 union KMP_ALIGN_CACHE kmp_barrier_union {
2167  double b_align; /* use worst case alignment */
2168  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2169  kmp_bstate_t bb;
2170 };
2171 
2172 typedef union kmp_barrier_union kmp_balign_t;
2173 
2174 /* Team barrier needs only non-volatile arrived counter */
2175 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2176  double b_align; /* use worst case alignment */
2177  char b_pad[CACHE_LINE];
2178  struct {
2179  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2180 #if USE_DEBUGGER
2181  // The following two fields are indended for the debugger solely. Only
2182  // primary thread of the team accesses these fields: the first one is
2183  // increased by 1 when the primary thread arrives to a barrier, the second
2184  // one is increased by one when all the threads arrived.
2185  kmp_uint b_master_arrived;
2186  kmp_uint b_team_arrived;
2187 #endif
2188  };
2189 };
2190 
2191 typedef union kmp_barrier_team_union kmp_balign_team_t;
2192 
2193 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2194  threads when a condition changes. This is to workaround an NPTL bug where
2195  padding was added to pthread_cond_t which caused the initialization routine
2196  to write outside of the structure if compiled on pre-NPTL threads. */
2197 #if KMP_OS_WINDOWS
2198 typedef struct kmp_win32_mutex {
2199  /* The Lock */
2200  CRITICAL_SECTION cs;
2201 } kmp_win32_mutex_t;
2202 
2203 typedef struct kmp_win32_cond {
2204  /* Count of the number of waiters. */
2205  int waiters_count_;
2206 
2207  /* Serialize access to <waiters_count_> */
2208  kmp_win32_mutex_t waiters_count_lock_;
2209 
2210  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2211  int release_count_;
2212 
2213  /* Keeps track of the current "generation" so that we don't allow */
2214  /* one thread to steal all the "releases" from the broadcast. */
2215  int wait_generation_count_;
2216 
2217  /* A manual-reset event that's used to block and release waiting threads. */
2218  HANDLE event_;
2219 } kmp_win32_cond_t;
2220 #endif
2221 
2222 #if KMP_OS_UNIX
2223 
2224 union KMP_ALIGN_CACHE kmp_cond_union {
2225  double c_align;
2226  char c_pad[CACHE_LINE];
2227  pthread_cond_t c_cond;
2228 };
2229 
2230 typedef union kmp_cond_union kmp_cond_align_t;
2231 
2232 union KMP_ALIGN_CACHE kmp_mutex_union {
2233  double m_align;
2234  char m_pad[CACHE_LINE];
2235  pthread_mutex_t m_mutex;
2236 };
2237 
2238 typedef union kmp_mutex_union kmp_mutex_align_t;
2239 
2240 #endif /* KMP_OS_UNIX */
2241 
2242 typedef struct kmp_desc_base {
2243  void *ds_stackbase;
2244  size_t ds_stacksize;
2245  int ds_stackgrow;
2246  kmp_thread_t ds_thread;
2247  volatile int ds_tid;
2248  int ds_gtid;
2249 #if KMP_OS_WINDOWS
2250  volatile int ds_alive;
2251  DWORD ds_thread_id;
2252 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2253  However, debugger support (libomp_db) cannot work with handles, because they
2254  uncomparable. For example, debugger requests info about thread with handle h.
2255  h is valid within debugger process, and meaningless within debugee process.
2256  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2257  within debugee process, but it is a *new* handle which does *not* equal to
2258  any other handle in debugee... The only way to compare handles is convert
2259  them to system-wide ids. GetThreadId() function is available only in
2260  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2261  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2262  thread id by call to GetCurrentThreadId() from within the thread and save it
2263  to let libomp_db identify threads. */
2264 #endif /* KMP_OS_WINDOWS */
2265 } kmp_desc_base_t;
2266 
2267 typedef union KMP_ALIGN_CACHE kmp_desc {
2268  double ds_align; /* use worst case alignment */
2269  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2270  kmp_desc_base_t ds;
2271 } kmp_desc_t;
2272 
2273 typedef struct kmp_local {
2274  volatile int this_construct; /* count of single's encountered by thread */
2275  void *reduce_data;
2276 #if KMP_USE_BGET
2277  void *bget_data;
2278  void *bget_list;
2279 #if !USE_CMP_XCHG_FOR_BGET
2280 #ifdef USE_QUEUING_LOCK_FOR_BGET
2281  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2282 #else
2283  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2284 // bootstrap lock so we can use it at library
2285 // shutdown.
2286 #endif /* USE_LOCK_FOR_BGET */
2287 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2288 #endif /* KMP_USE_BGET */
2289 
2290  PACKED_REDUCTION_METHOD_T
2291  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2292  __kmpc_end_reduce*() */
2293 
2294 } kmp_local_t;
2295 
2296 #define KMP_CHECK_UPDATE(a, b) \
2297  if ((a) != (b)) \
2298  (a) = (b)
2299 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2300  if ((a) != (b)) \
2301  TCW_SYNC_PTR((a), (b))
2302 
2303 #define get__blocktime(xteam, xtid) \
2304  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2305 #define get__bt_set(xteam, xtid) \
2306  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2307 #if KMP_USE_MONITOR
2308 #define get__bt_intervals(xteam, xtid) \
2309  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2310 #endif
2311 
2312 #define get__dynamic_2(xteam, xtid) \
2313  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2314 #define get__nproc_2(xteam, xtid) \
2315  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2316 #define get__sched_2(xteam, xtid) \
2317  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2318 
2319 #define set__blocktime_team(xteam, xtid, xval) \
2320  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2321  (xval))
2322 
2323 #if KMP_USE_MONITOR
2324 #define set__bt_intervals_team(xteam, xtid, xval) \
2325  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2326  (xval))
2327 #endif
2328 
2329 #define set__bt_set_team(xteam, xtid, xval) \
2330  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2331 
2332 #define set__dynamic(xthread, xval) \
2333  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2334 #define get__dynamic(xthread) \
2335  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2336 
2337 #define set__nproc(xthread, xval) \
2338  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2339 
2340 #define set__thread_limit(xthread, xval) \
2341  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2342 
2343 #define set__max_active_levels(xthread, xval) \
2344  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2345 
2346 #define get__max_active_levels(xthread) \
2347  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2348 
2349 #define set__sched(xthread, xval) \
2350  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2351 
2352 #define set__proc_bind(xthread, xval) \
2353  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2354 #define get__proc_bind(xthread) \
2355  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2356 
2357 // OpenMP tasking data structures
2358 
2359 typedef enum kmp_tasking_mode {
2360  tskm_immediate_exec = 0,
2361  tskm_extra_barrier = 1,
2362  tskm_task_teams = 2,
2363  tskm_max = 2
2364 } kmp_tasking_mode_t;
2365 
2366 extern kmp_tasking_mode_t
2367  __kmp_tasking_mode; /* determines how/when to execute tasks */
2368 extern int __kmp_task_stealing_constraint;
2369 extern int __kmp_enable_task_throttling;
2370 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2371 // specified, defaults to 0 otherwise
2372 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2373 extern kmp_int32 __kmp_max_task_priority;
2374 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2375 extern kmp_uint64 __kmp_taskloop_min_tasks;
2376 
2377 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2378  taskdata first */
2379 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2380 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2381 
2382 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2383 // were spawned and queued since the previous barrier release.
2384 #define KMP_TASKING_ENABLED(task_team) \
2385  (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2393 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2394 
2395 typedef union kmp_cmplrdata {
2396  kmp_int32 priority;
2397  kmp_routine_entry_t
2398  destructors; /* pointer to function to invoke deconstructors of
2399  firstprivate C++ objects */
2400  /* future data */
2401 } kmp_cmplrdata_t;
2402 
2403 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2406 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2407  void *shareds;
2408  kmp_routine_entry_t
2409  routine;
2410  kmp_int32 part_id;
2411  kmp_cmplrdata_t
2412  data1; /* Two known optional additions: destructors and priority */
2413  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2414  /* future data */
2415  /* private vars */
2416 } kmp_task_t;
2417 
2422 typedef struct kmp_taskgroup {
2423  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2424  std::atomic<kmp_int32>
2425  cancel_request; // request for cancellation of this taskgroup
2426  struct kmp_taskgroup *parent; // parent taskgroup
2427  // Block of data to perform task reduction
2428  void *reduce_data; // reduction related info
2429  kmp_int32 reduce_num_data; // number of data items to reduce
2430  uintptr_t *gomp_data; // gomp reduction data
2431 } kmp_taskgroup_t;
2432 
2433 // forward declarations
2434 typedef union kmp_depnode kmp_depnode_t;
2435 typedef struct kmp_depnode_list kmp_depnode_list_t;
2436 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2437 
2438 // macros for checking dep flag as an integer
2439 #define KMP_DEP_IN 0x1
2440 #define KMP_DEP_OUT 0x2
2441 #define KMP_DEP_INOUT 0x3
2442 #define KMP_DEP_MTX 0x4
2443 #define KMP_DEP_SET 0x8
2444 #define KMP_DEP_ALL 0x80
2445 // Compiler sends us this info:
2446 typedef struct kmp_depend_info {
2447  kmp_intptr_t base_addr;
2448  size_t len;
2449  union {
2450  kmp_uint8 flag; // flag as an unsigned char
2451  struct { // flag as a set of 8 bits
2452  unsigned in : 1;
2453  unsigned out : 1;
2454  unsigned mtx : 1;
2455  unsigned set : 1;
2456  unsigned unused : 3;
2457  unsigned all : 1;
2458  } flags;
2459  };
2460 } kmp_depend_info_t;
2461 
2462 // Internal structures to work with task dependencies:
2463 struct kmp_depnode_list {
2464  kmp_depnode_t *node;
2465  kmp_depnode_list_t *next;
2466 };
2467 
2468 // Max number of mutexinoutset dependencies per node
2469 #define MAX_MTX_DEPS 4
2470 
2471 typedef struct kmp_base_depnode {
2472  kmp_depnode_list_t *successors; /* used under lock */
2473  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2474  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2475  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2476  kmp_lock_t lock; /* guards shared fields: task, successors */
2477 #if KMP_SUPPORT_GRAPH_OUTPUT
2478  kmp_uint32 id;
2479 #endif
2480  std::atomic<kmp_int32> npredecessors;
2481  std::atomic<kmp_int32> nrefs;
2482 } kmp_base_depnode_t;
2483 
2484 union KMP_ALIGN_CACHE kmp_depnode {
2485  double dn_align; /* use worst case alignment */
2486  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2487  kmp_base_depnode_t dn;
2488 };
2489 
2490 struct kmp_dephash_entry {
2491  kmp_intptr_t addr;
2492  kmp_depnode_t *last_out;
2493  kmp_depnode_list_t *last_set;
2494  kmp_depnode_list_t *prev_set;
2495  kmp_uint8 last_flag;
2496  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2497  kmp_dephash_entry_t *next_in_bucket;
2498 };
2499 
2500 typedef struct kmp_dephash {
2501  kmp_dephash_entry_t **buckets;
2502  size_t size;
2503  kmp_depnode_t *last_all;
2504  size_t generation;
2505  kmp_uint32 nelements;
2506  kmp_uint32 nconflicts;
2507 } kmp_dephash_t;
2508 
2509 typedef struct kmp_task_affinity_info {
2510  kmp_intptr_t base_addr;
2511  size_t len;
2512  struct {
2513  bool flag1 : 1;
2514  bool flag2 : 1;
2515  kmp_int32 reserved : 30;
2516  } flags;
2517 } kmp_task_affinity_info_t;
2518 
2519 typedef enum kmp_event_type_t {
2520  KMP_EVENT_UNINITIALIZED = 0,
2521  KMP_EVENT_ALLOW_COMPLETION = 1
2522 } kmp_event_type_t;
2523 
2524 typedef struct {
2525  kmp_event_type_t type;
2526  kmp_tas_lock_t lock;
2527  union {
2528  kmp_task_t *task;
2529  } ed;
2530 } kmp_event_t;
2531 
2532 #if OMPX_TASKGRAPH
2533 // Initial number of allocated nodes while recording
2534 #define INIT_MAPSIZE 50
2535 
2536 typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2537  unsigned nowait : 1;
2538  unsigned re_record : 1;
2539  unsigned reserved : 30;
2540 } kmp_taskgraph_flags_t;
2541 
2543 typedef struct kmp_node_info {
2544  kmp_task_t *task; // Pointer to the actual task
2545  kmp_int32 *successors; // Array of the succesors ids
2546  kmp_int32 nsuccessors; // Number of succesors of the node
2547  std::atomic<kmp_int32>
2548  npredecessors_counter; // Number of predessors on the fly
2549  kmp_int32 npredecessors; // Total number of predecessors
2550  kmp_int32 successors_size; // Number of allocated succesors ids
2551  kmp_taskdata_t *parent_task; // Parent implicit task
2552 } kmp_node_info_t;
2553 
2555 typedef enum kmp_tdg_status {
2556  KMP_TDG_NONE = 0,
2557  KMP_TDG_RECORDING = 1,
2558  KMP_TDG_READY = 2
2559 } kmp_tdg_status_t;
2560 
2562 typedef struct kmp_tdg_info {
2563  kmp_int32 tdg_id; // Unique idenfifier of the TDG
2564  kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2565  kmp_int32 map_size; // Number of allocated TDG nodes
2566  kmp_int32 num_roots; // Number of roots tasks int the TDG
2567  kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2568  kmp_node_info_t *record_map; // Array of TDG nodes
2569  kmp_tdg_status_t tdg_status =
2570  KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2571  std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2572  kmp_bootstrap_lock_t
2573  graph_lock; // Protect graph attributes when updated via taskloop_recur
2574  // Taskloop reduction related
2575  void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2576  // __kmpc_taskred_init
2577  kmp_int32 rec_num_taskred;
2578 } kmp_tdg_info_t;
2579 
2580 extern int __kmp_tdg_dot;
2581 extern kmp_int32 __kmp_max_tdgs;
2582 extern kmp_tdg_info_t **__kmp_global_tdgs;
2583 extern kmp_int32 __kmp_curr_tdg_idx;
2584 extern kmp_int32 __kmp_successors_size;
2585 extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2586 extern kmp_int32 __kmp_num_tdg;
2587 #endif
2588 
2589 #ifdef BUILD_TIED_TASK_STACK
2590 
2591 /* Tied Task stack definitions */
2592 typedef struct kmp_stack_block {
2593  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2594  struct kmp_stack_block *sb_next;
2595  struct kmp_stack_block *sb_prev;
2596 } kmp_stack_block_t;
2597 
2598 typedef struct kmp_task_stack {
2599  kmp_stack_block_t ts_first_block; // first block of stack entries
2600  kmp_taskdata_t **ts_top; // pointer to the top of stack
2601  kmp_int32 ts_entries; // number of entries on the stack
2602 } kmp_task_stack_t;
2603 
2604 #endif // BUILD_TIED_TASK_STACK
2605 
2606 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2607  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2608  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2609  unsigned final : 1; /* task is final(1) so execute immediately */
2610  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2611  code path */
2612  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2613  invoke destructors from the runtime */
2614  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2615  context of the RTL) */
2616  unsigned priority_specified : 1; /* set if the compiler provides priority
2617  setting for the task */
2618  unsigned detachable : 1; /* 1 == can detach */
2619  unsigned hidden_helper : 1; /* 1 == hidden helper task */
2620  unsigned reserved : 8; /* reserved for compiler use */
2621 
2622  /* Library flags */ /* Total library flags must be 16 bits */
2623  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2624  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2625  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2626  // (1) or may be deferred (0)
2627  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2628  // (0) [>= 2 threads]
2629  /* If either team_serial or tasking_ser is set, task team may be NULL */
2630  /* Task State Flags: */
2631  unsigned started : 1; /* 1==started, 0==not started */
2632  unsigned executing : 1; /* 1==executing, 0==not executing */
2633  unsigned complete : 1; /* 1==complete, 0==not complete */
2634  unsigned freed : 1; /* 1==freed, 0==allocated */
2635  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2636 #if OMPX_TASKGRAPH
2637  unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2638  unsigned reserved31 : 6; /* reserved for library use */
2639 #else
2640  unsigned reserved31 : 7; /* reserved for library use */
2641 #endif
2642 
2643 } kmp_tasking_flags_t;
2644 
2645 typedef struct kmp_target_data {
2646  void *async_handle; // libomptarget async handle for task completion query
2647 } kmp_target_data_t;
2648 
2649 struct kmp_taskdata { /* aligned during dynamic allocation */
2650  kmp_int32 td_task_id; /* id, assigned by debugger */
2651  kmp_tasking_flags_t td_flags; /* task flags */
2652  kmp_team_t *td_team; /* team for this task */
2653  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2654  /* Currently not used except for perhaps IDB */
2655  kmp_taskdata_t *td_parent; /* parent task */
2656  kmp_int32 td_level; /* task nesting level */
2657  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2658  ident_t *td_ident; /* task identifier */
2659  // Taskwait data.
2660  ident_t *td_taskwait_ident;
2661  kmp_uint32 td_taskwait_counter;
2662  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2663  KMP_ALIGN_CACHE kmp_internal_control_t
2664  td_icvs; /* Internal control variables for the task */
2665  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2666  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2667  deallocated */
2668  std::atomic<kmp_int32>
2669  td_incomplete_child_tasks; /* Child tasks not yet complete */
2670  kmp_taskgroup_t
2671  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2672  kmp_dephash_t
2673  *td_dephash; // Dependencies for children tasks are tracked from here
2674  kmp_depnode_t
2675  *td_depnode; // Pointer to graph node if this task has dependencies
2676  kmp_task_team_t *td_task_team;
2677  size_t td_size_alloc; // Size of task structure, including shareds etc.
2678 #if defined(KMP_GOMP_COMPAT)
2679  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2680  kmp_int32 td_size_loop_bounds;
2681 #endif
2682  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2683 #if defined(KMP_GOMP_COMPAT)
2684  // GOMP sends in a copy function for copy constructors
2685  void (*td_copy_func)(void *, void *);
2686 #endif
2687  kmp_event_t td_allow_completion_event;
2688 #if OMPT_SUPPORT
2689  ompt_task_info_t ompt_task_info;
2690 #endif
2691 #if OMPX_TASKGRAPH
2692  bool is_taskgraph = 0; // whether the task is within a TDG
2693  kmp_tdg_info_t *tdg; // used to associate task with a TDG
2694 #endif
2695  kmp_target_data_t td_target_data;
2696 }; // struct kmp_taskdata
2697 
2698 // Make sure padding above worked
2699 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2700 
2701 // Data for task team but per thread
2702 typedef struct kmp_base_thread_data {
2703  kmp_info_p *td_thr; // Pointer back to thread info
2704  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2705  // queued?
2706  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2707  kmp_taskdata_t *
2708  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2709  kmp_int32 td_deque_size; // Size of deck
2710  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2711  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2712  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2713  // GEH: shouldn't this be volatile since used in while-spin?
2714  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2715 #ifdef BUILD_TIED_TASK_STACK
2716  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2717 // scheduling constraint
2718 #endif // BUILD_TIED_TASK_STACK
2719 } kmp_base_thread_data_t;
2720 
2721 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2722 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2723 
2724 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2725 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2726 
2727 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2728  kmp_base_thread_data_t td;
2729  double td_align; /* use worst case alignment */
2730  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2731 } kmp_thread_data_t;
2732 
2733 typedef struct kmp_task_pri {
2734  kmp_thread_data_t td;
2735  kmp_int32 priority;
2736  kmp_task_pri *next;
2737 } kmp_task_pri_t;
2738 
2739 // Data for task teams which are used when tasking is enabled for the team
2740 typedef struct kmp_base_task_team {
2741  kmp_bootstrap_lock_t
2742  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2743  /* must be bootstrap lock since used at library shutdown*/
2744 
2745  // TODO: check performance vs kmp_tas_lock_t
2746  kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2747  kmp_task_pri_t *tt_task_pri_list;
2748 
2749  kmp_task_team_t *tt_next; /* For linking the task team free list */
2750  kmp_thread_data_t
2751  *tt_threads_data; /* Array of per-thread structures for task team */
2752  /* Data survives task team deallocation */
2753  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2754  executing this team? */
2755  /* TRUE means tt_threads_data is set up and initialized */
2756  kmp_int32 tt_nproc; /* #threads in team */
2757  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2758  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2759  kmp_int32 tt_untied_task_encountered;
2760  std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2761  // There is hidden helper thread encountered in this task team so that we must
2762  // wait when waiting on task team
2763  kmp_int32 tt_hidden_helper_task_encountered;
2764 
2765  KMP_ALIGN_CACHE
2766  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2767 
2768  KMP_ALIGN_CACHE
2769  volatile kmp_uint32
2770  tt_active; /* is the team still actively executing tasks */
2771 } kmp_base_task_team_t;
2772 
2773 union KMP_ALIGN_CACHE kmp_task_team {
2774  kmp_base_task_team_t tt;
2775  double tt_align; /* use worst case alignment */
2776  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2777 };
2778 
2779 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2780 // Free lists keep same-size free memory slots for fast memory allocation
2781 // routines
2782 typedef struct kmp_free_list {
2783  void *th_free_list_self; // Self-allocated tasks free list
2784  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2785  // threads
2786  void *th_free_list_other; // Non-self free list (to be returned to owner's
2787  // sync list)
2788 } kmp_free_list_t;
2789 #endif
2790 #if KMP_NESTED_HOT_TEAMS
2791 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2792 // are not put in teams pool, and they don't put threads in threads pool.
2793 typedef struct kmp_hot_team_ptr {
2794  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2795  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2796 } kmp_hot_team_ptr_t;
2797 #endif
2798 typedef struct kmp_teams_size {
2799  kmp_int32 nteams; // number of teams in a league
2800  kmp_int32 nth; // number of threads in each team of the league
2801 } kmp_teams_size_t;
2802 
2803 // This struct stores a thread that acts as a "root" for a contention
2804 // group. Contention groups are rooted at kmp_root threads, but also at
2805 // each primary thread of each team created in the teams construct.
2806 // This struct therefore also stores a thread_limit associated with
2807 // that contention group, and a counter to track the number of threads
2808 // active in that contention group. Each thread has a list of these: CG
2809 // root threads have an entry in their list in which cg_root refers to
2810 // the thread itself, whereas other workers in the CG will have a
2811 // single entry where cg_root is same as the entry containing their CG
2812 // root. When a thread encounters a teams construct, it will add a new
2813 // entry to the front of its list, because it now roots a new CG.
2814 typedef struct kmp_cg_root {
2815  kmp_info_p *cg_root; // "root" thread for a contention group
2816  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2817  // thread_limit clause for teams primary threads
2818  kmp_int32 cg_thread_limit;
2819  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2820  struct kmp_cg_root *up; // pointer to higher level CG root in list
2821 } kmp_cg_root_t;
2822 
2823 // OpenMP thread data structures
2824 
2825 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2826  /* Start with the readonly data which is cache aligned and padded. This is
2827  written before the thread starts working by the primary thread. Uber
2828  masters may update themselves later. Usage does not consider serialized
2829  regions. */
2830  kmp_desc_t th_info;
2831  kmp_team_p *th_team; /* team we belong to */
2832  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2833  kmp_info_p *th_next_pool; /* next available thread in the pool */
2834  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2835  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2836 
2837  /* The following are cached from the team info structure */
2838  /* TODO use these in more places as determined to be needed via profiling */
2839  int th_team_nproc; /* number of threads in a team */
2840  kmp_info_p *th_team_master; /* the team's primary thread */
2841  int th_team_serialized; /* team is serialized */
2842  microtask_t th_teams_microtask; /* save entry address for teams construct */
2843  int th_teams_level; /* save initial level of teams construct */
2844 /* it is 0 on device but may be any on host */
2845 
2846 /* The blocktime info is copied from the team struct to the thread struct */
2847 /* at the start of a barrier, and the values stored in the team are used */
2848 /* at points in the code where the team struct is no longer guaranteed */
2849 /* to exist (from the POV of worker threads). */
2850 #if KMP_USE_MONITOR
2851  int th_team_bt_intervals;
2852  int th_team_bt_set;
2853 #else
2854  kmp_uint64 th_team_bt_intervals;
2855 #endif
2856 
2857 #if KMP_AFFINITY_SUPPORTED
2858  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2859  kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2860  kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2861 #endif
2862  omp_allocator_handle_t th_def_allocator; /* default allocator */
2863  /* The data set by the primary thread at reinit, then R/W by the worker */
2864  KMP_ALIGN_CACHE int
2865  th_set_nproc; /* if > 0, then only use this request for the next fork */
2866 #if KMP_NESTED_HOT_TEAMS
2867  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2868 #endif
2869  kmp_proc_bind_t
2870  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2871  kmp_teams_size_t
2872  th_teams_size; /* number of teams/threads in teams construct */
2873 #if KMP_AFFINITY_SUPPORTED
2874  int th_current_place; /* place currently bound to */
2875  int th_new_place; /* place to bind to in par reg */
2876  int th_first_place; /* first place in partition */
2877  int th_last_place; /* last place in partition */
2878 #endif
2879  int th_prev_level; /* previous level for affinity format */
2880  int th_prev_num_threads; /* previous num_threads for affinity format */
2881 #if USE_ITT_BUILD
2882  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2883  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2884  kmp_uint64 th_frame_time; /* frame timestamp */
2885 #endif /* USE_ITT_BUILD */
2886  kmp_local_t th_local;
2887  struct private_common *th_pri_head;
2888 
2889  /* Now the data only used by the worker (after initial allocation) */
2890  /* TODO the first serial team should actually be stored in the info_t
2891  structure. this will help reduce initial allocation overhead */
2892  KMP_ALIGN_CACHE kmp_team_p
2893  *th_serial_team; /*serialized team held in reserve*/
2894 
2895 #if OMPT_SUPPORT
2896  ompt_thread_info_t ompt_thread_info;
2897 #endif
2898 
2899  /* The following are also read by the primary thread during reinit */
2900  struct common_table *th_pri_common;
2901 
2902  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2903  /* while awaiting queuing lock acquire */
2904 
2905  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2906  flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2907 
2908  ident_t *th_ident;
2909  unsigned th_x; // Random number generator data
2910  unsigned th_a; // Random number generator data
2911 
2912  /* Tasking-related data for the thread */
2913  kmp_task_team_t *th_task_team; // Task team struct
2914  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2915  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2916  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2917  // at nested levels
2918  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2919  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2920  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2921  // tasking, thus safe to reap
2922 
2923  /* More stuff for keeping track of active/sleeping threads (this part is
2924  written by the worker thread) */
2925  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2926  int th_active; // ! sleeping; 32 bits for TCR/TCW
2927  std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2928  // 0 = not used in team; 1 = used in team;
2929  // 2 = transitioning to not used in team; 3 = transitioning to used in team
2930  struct cons_header *th_cons; // used for consistency check
2931 #if KMP_USE_HIER_SCHED
2932  // used for hierarchical scheduling
2933  kmp_hier_private_bdata_t *th_hier_bar_data;
2934 #endif
2935 
2936  /* Add the syncronizing data which is cache aligned and padded. */
2937  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2938 
2939  KMP_ALIGN_CACHE volatile kmp_int32
2940  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2941 
2942 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2943 #define NUM_LISTS 4
2944  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2945 // allocation routines
2946 #endif
2947 
2948 #if KMP_OS_WINDOWS
2949  kmp_win32_cond_t th_suspend_cv;
2950  kmp_win32_mutex_t th_suspend_mx;
2951  std::atomic<int> th_suspend_init;
2952 #endif
2953 #if KMP_OS_UNIX
2954  kmp_cond_align_t th_suspend_cv;
2955  kmp_mutex_align_t th_suspend_mx;
2956  std::atomic<int> th_suspend_init_count;
2957 #endif
2958 
2959 #if USE_ITT_BUILD
2960  kmp_itt_mark_t th_itt_mark_single;
2961 // alignment ???
2962 #endif /* USE_ITT_BUILD */
2963 #if KMP_STATS_ENABLED
2964  kmp_stats_list *th_stats;
2965 #endif
2966 #if KMP_OS_UNIX
2967  std::atomic<bool> th_blocking;
2968 #endif
2969  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2970 } kmp_base_info_t;
2971 
2972 typedef union KMP_ALIGN_CACHE kmp_info {
2973  double th_align; /* use worst case alignment */
2974  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2975  kmp_base_info_t th;
2976 } kmp_info_t;
2977 
2978 // OpenMP thread team data structures
2979 
2980 typedef struct kmp_base_data {
2981  volatile kmp_uint32 t_value;
2982 } kmp_base_data_t;
2983 
2984 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2985  double dt_align; /* use worst case alignment */
2986  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2987  kmp_base_data_t dt;
2988 } kmp_sleep_team_t;
2989 
2990 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2991  double dt_align; /* use worst case alignment */
2992  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2993  kmp_base_data_t dt;
2994 } kmp_ordered_team_t;
2995 
2996 typedef int (*launch_t)(int gtid);
2997 
2998 /* Minimum number of ARGV entries to malloc if necessary */
2999 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3000 
3001 // Set up how many argv pointers will fit in cache lines containing
3002 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3003 // larger value for more space between the primary write/worker read section and
3004 // read/write by all section seems to buy more performance on EPCC PARALLEL.
3005 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3006 #define KMP_INLINE_ARGV_BYTES \
3007  (4 * CACHE_LINE - \
3008  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
3009  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
3010  CACHE_LINE))
3011 #else
3012 #define KMP_INLINE_ARGV_BYTES \
3013  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3014 #endif
3015 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3016 
3017 typedef struct KMP_ALIGN_CACHE kmp_base_team {
3018  // Synchronization Data
3019  // ---------------------------------------------------------------------------
3020  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3021  kmp_balign_team_t t_bar[bs_last_barrier];
3022  std::atomic<int> t_construct; // count of single directive encountered by team
3023  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3024 
3025  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3026  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3027  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3028 
3029  // Primary thread only
3030  // ---------------------------------------------------------------------------
3031  KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3032  int t_master_this_cons; // "this_construct" single counter of primary thread
3033  // in parent team
3034  ident_t *t_ident; // if volatile, have to change too much other crud to
3035  // volatile too
3036  kmp_team_p *t_parent; // parent team
3037  kmp_team_p *t_next_pool; // next free team in the team pool
3038  kmp_disp_t *t_dispatch; // thread's dispatch data
3039  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3040  kmp_proc_bind_t t_proc_bind; // bind type for par region
3041 #if USE_ITT_BUILD
3042  kmp_uint64 t_region_time; // region begin timestamp
3043 #endif /* USE_ITT_BUILD */
3044 
3045  // Primary thread write, workers read
3046  // --------------------------------------------------------------------------
3047  KMP_ALIGN_CACHE void **t_argv;
3048  int t_argc;
3049  int t_nproc; // number of threads in team
3050  microtask_t t_pkfn;
3051  launch_t t_invoke; // procedure to launch the microtask
3052 
3053 #if OMPT_SUPPORT
3054  ompt_team_info_t ompt_team_info;
3055  ompt_lw_taskteam_t *ompt_serialized_team_info;
3056 #endif
3057 
3058 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3059  kmp_int8 t_fp_control_saved;
3060  kmp_int8 t_pad2b;
3061  kmp_int16 t_x87_fpu_control_word; // FP control regs
3062  kmp_uint32 t_mxcsr;
3063 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3064 
3065  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3066 
3067  KMP_ALIGN_CACHE kmp_info_t **t_threads;
3068  kmp_taskdata_t
3069  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3070  int t_level; // nested parallel level
3071 
3072  KMP_ALIGN_CACHE int t_max_argc;
3073  int t_max_nproc; // max threads this team can handle (dynamically expandable)
3074  int t_serialized; // levels deep of serialized teams
3075  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3076  int t_id; // team's id, assigned by debugger.
3077  int t_active_level; // nested active parallel level
3078  kmp_r_sched_t t_sched; // run-time schedule for the team
3079 #if KMP_AFFINITY_SUPPORTED
3080  int t_first_place; // first & last place in parent thread's partition.
3081  int t_last_place; // Restore these values to primary thread after par region.
3082 #endif // KMP_AFFINITY_SUPPORTED
3083  int t_display_affinity;
3084  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3085  // omp_set_num_threads() call
3086  omp_allocator_handle_t t_def_allocator; /* default allocator */
3087 
3088 // Read/write by workers as well
3089 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3090  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3091  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3092  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3093  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3094  char dummy_padding[1024];
3095 #endif
3096  // Internal control stack for additional nested teams.
3097  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3098  // for SERIALIZED teams nested 2 or more levels deep
3099  // typed flag to store request state of cancellation
3100  std::atomic<kmp_int32> t_cancel_request;
3101  int t_master_active; // save on fork, restore on join
3102  void *t_copypriv_data; // team specific pointer to copyprivate data array
3103 #if KMP_OS_WINDOWS
3104  std::atomic<kmp_uint32> t_copyin_counter;
3105 #endif
3106 #if USE_ITT_BUILD
3107  void *t_stack_id; // team specific stack stitching id (for ittnotify)
3108 #endif /* USE_ITT_BUILD */
3109  distributedBarrier *b; // Distributed barrier data associated with team
3110 } kmp_base_team_t;
3111 
3112 union KMP_ALIGN_CACHE kmp_team {
3113  kmp_base_team_t t;
3114  double t_align; /* use worst case alignment */
3115  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3116 };
3117 
3118 typedef union KMP_ALIGN_CACHE kmp_time_global {
3119  double dt_align; /* use worst case alignment */
3120  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3121  kmp_base_data_t dt;
3122 } kmp_time_global_t;
3123 
3124 typedef struct kmp_base_global {
3125  /* cache-aligned */
3126  kmp_time_global_t g_time;
3127 
3128  /* non cache-aligned */
3129  volatile int g_abort;
3130  volatile int g_done;
3131 
3132  int g_dynamic;
3133  enum dynamic_mode g_dynamic_mode;
3134 } kmp_base_global_t;
3135 
3136 typedef union KMP_ALIGN_CACHE kmp_global {
3137  kmp_base_global_t g;
3138  double g_align; /* use worst case alignment */
3139  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3140 } kmp_global_t;
3141 
3142 typedef struct kmp_base_root {
3143  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3144  // (r_in_parallel>= 0)
3145  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3146  // the synch overhead or keeping r_active
3147  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3148  // keeps a count of active parallel regions per root
3149  std::atomic<int> r_in_parallel;
3150  // GEH: This is misnamed, should be r_active_levels
3151  kmp_team_t *r_root_team;
3152  kmp_team_t *r_hot_team;
3153  kmp_info_t *r_uber_thread;
3154  kmp_lock_t r_begin_lock;
3155  volatile int r_begin;
3156  int r_blocktime; /* blocktime for this root and descendants */
3157 #if KMP_AFFINITY_SUPPORTED
3158  int r_affinity_assigned;
3159 #endif // KMP_AFFINITY_SUPPORTED
3160 } kmp_base_root_t;
3161 
3162 typedef union KMP_ALIGN_CACHE kmp_root {
3163  kmp_base_root_t r;
3164  double r_align; /* use worst case alignment */
3165  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3166 } kmp_root_t;
3167 
3168 struct fortran_inx_info {
3169  kmp_int32 data;
3170 };
3171 
3172 // This list type exists to hold old __kmp_threads arrays so that
3173 // old references to them may complete while reallocation takes place when
3174 // expanding the array. The items in this list are kept alive until library
3175 // shutdown.
3176 typedef struct kmp_old_threads_list_t {
3177  kmp_info_t **threads;
3178  struct kmp_old_threads_list_t *next;
3179 } kmp_old_threads_list_t;
3180 
3181 /* ------------------------------------------------------------------------ */
3182 
3183 extern int __kmp_settings;
3184 extern int __kmp_duplicate_library_ok;
3185 #if USE_ITT_BUILD
3186 extern int __kmp_forkjoin_frames;
3187 extern int __kmp_forkjoin_frames_mode;
3188 #endif
3189 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3190 extern int __kmp_determ_red;
3191 
3192 #ifdef KMP_DEBUG
3193 extern int kmp_a_debug;
3194 extern int kmp_b_debug;
3195 extern int kmp_c_debug;
3196 extern int kmp_d_debug;
3197 extern int kmp_e_debug;
3198 extern int kmp_f_debug;
3199 #endif /* KMP_DEBUG */
3200 
3201 /* For debug information logging using rotating buffer */
3202 #define KMP_DEBUG_BUF_LINES_INIT 512
3203 #define KMP_DEBUG_BUF_LINES_MIN 1
3204 
3205 #define KMP_DEBUG_BUF_CHARS_INIT 128
3206 #define KMP_DEBUG_BUF_CHARS_MIN 2
3207 
3208 extern int
3209  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3210 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3211 extern int
3212  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3213 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3214  entry pointer */
3215 
3216 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3217 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3218  printed in buffer so far */
3219 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3220  recommended in warnings */
3221 /* end rotating debug buffer */
3222 
3223 #ifdef KMP_DEBUG
3224 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3225 
3226 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3227 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3228 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3229 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3230 extern int __kmp_par_range_lb;
3231 extern int __kmp_par_range_ub;
3232 #endif
3233 
3234 /* For printing out dynamic storage map for threads and teams */
3235 extern int
3236  __kmp_storage_map; /* True means print storage map for threads and teams */
3237 extern int __kmp_storage_map_verbose; /* True means storage map includes
3238  placement info */
3239 extern int __kmp_storage_map_verbose_specified;
3240 
3241 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3242 extern kmp_cpuinfo_t __kmp_cpuinfo;
3243 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3244 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3245 static inline bool __kmp_is_hybrid_cpu() { return true; }
3246 #else
3247 static inline bool __kmp_is_hybrid_cpu() { return false; }
3248 #endif
3249 
3250 extern volatile int __kmp_init_serial;
3251 extern volatile int __kmp_init_gtid;
3252 extern volatile int __kmp_init_common;
3253 extern volatile int __kmp_need_register_serial;
3254 extern volatile int __kmp_init_middle;
3255 extern volatile int __kmp_init_parallel;
3256 #if KMP_USE_MONITOR
3257 extern volatile int __kmp_init_monitor;
3258 #endif
3259 extern volatile int __kmp_init_user_locks;
3260 extern volatile int __kmp_init_hidden_helper_threads;
3261 extern int __kmp_init_counter;
3262 extern int __kmp_root_counter;
3263 extern int __kmp_version;
3264 
3265 /* list of address of allocated caches for commons */
3266 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3267 
3268 /* Barrier algorithm types and options */
3269 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3270 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3271 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3272 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3273 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3274 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3275 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3276 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3277 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3278 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3279 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3280 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3281 
3282 /* Global Locks */
3283 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3284 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3285 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3286 extern kmp_bootstrap_lock_t
3287  __kmp_exit_lock; /* exit() is not always thread-safe */
3288 #if KMP_USE_MONITOR
3289 extern kmp_bootstrap_lock_t
3290  __kmp_monitor_lock; /* control monitor thread creation */
3291 #endif
3292 extern kmp_bootstrap_lock_t
3293  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3294  __kmp_threads expansion to co-exist */
3295 
3296 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3297 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3298 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3299 
3300 extern enum library_type __kmp_library;
3301 
3302 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3303 extern enum sched_type __kmp_static; /* default static scheduling method */
3304 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3305 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3306 extern int __kmp_chunk; /* default runtime chunk size */
3307 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3308 
3309 extern size_t __kmp_stksize; /* stack size per thread */
3310 #if KMP_USE_MONITOR
3311 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3312 #endif
3313 extern size_t __kmp_stkoffset; /* stack offset per thread */
3314 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3315 
3316 extern size_t
3317  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3318 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3319 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3320 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3321 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3322 extern int __kmp_generate_warnings; /* should we issue warnings? */
3323 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3324 
3325 #ifdef DEBUG_SUSPEND
3326 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3327 #endif
3328 
3329 extern kmp_int32 __kmp_use_yield;
3330 extern kmp_int32 __kmp_use_yield_exp_set;
3331 extern kmp_uint32 __kmp_yield_init;
3332 extern kmp_uint32 __kmp_yield_next;
3333 extern kmp_uint64 __kmp_pause_init;
3334 
3335 /* ------------------------------------------------------------------------- */
3336 extern int __kmp_allThreadsSpecified;
3337 
3338 extern size_t __kmp_align_alloc;
3339 /* following data protected by initialization routines */
3340 extern int __kmp_xproc; /* number of processors in the system */
3341 extern int __kmp_avail_proc; /* number of processors available to the process */
3342 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3343 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3344 // maximum total number of concurrently-existing threads on device
3345 extern int __kmp_max_nth;
3346 // maximum total number of concurrently-existing threads in a contention group
3347 extern int __kmp_cg_max_nth;
3348 extern int __kmp_task_max_nth; // max threads used in a task
3349 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3350 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3351  __kmp_root */
3352 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3353  region a la OMP_NUM_THREADS */
3354 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3355  initialization */
3356 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3357  used (fixed) */
3358 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3359  (__kmpc_threadprivate_cached()) */
3360 extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3361  blocking (env setting) */
3362 extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3363 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3364 
3365 // Convert raw blocktime from ms to us if needed.
3366 static inline void __kmp_aux_convert_blocktime(int *bt) {
3367  if (__kmp_blocktime_units == 'm') {
3368  if (*bt > INT_MAX / 1000) {
3369  *bt = INT_MAX / 1000;
3370  KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3371  }
3372  *bt = *bt * 1000;
3373  }
3374 }
3375 
3376 #if KMP_USE_MONITOR
3377 extern int
3378  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3379 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3380  blocking */
3381 #endif
3382 #ifdef KMP_ADJUST_BLOCKTIME
3383 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3384 #endif /* KMP_ADJUST_BLOCKTIME */
3385 #ifdef KMP_DFLT_NTH_CORES
3386 extern int __kmp_ncores; /* Total number of cores for threads placement */
3387 #endif
3388 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3389 extern int __kmp_abort_delay;
3390 
3391 extern int __kmp_need_register_atfork_specified;
3392 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3393  to install fork handler */
3394 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3395  0 - not set, will be set at runtime
3396  1 - using stack search
3397  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3398  X*) or TlsGetValue(Windows* OS))
3399  3 - static TLS (__declspec(thread) __kmp_gtid),
3400  Linux* OS .so only. */
3401 extern int
3402  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3403 #ifdef KMP_TDATA_GTID
3404 extern KMP_THREAD_LOCAL int __kmp_gtid;
3405 #endif
3406 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3407 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3408 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3409 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3410 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3411 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3412 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3413 
3414 // max_active_levels for nested parallelism enabled by default via
3415 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3416 extern int __kmp_dflt_max_active_levels;
3417 // Indicates whether value of __kmp_dflt_max_active_levels was already
3418 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3419 extern bool __kmp_dflt_max_active_levels_set;
3420 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3421  concurrent execution per team */
3422 #if KMP_NESTED_HOT_TEAMS
3423 extern int __kmp_hot_teams_mode;
3424 extern int __kmp_hot_teams_max_level;
3425 #endif
3426 
3427 #if KMP_OS_LINUX
3428 extern enum clock_function_type __kmp_clock_function;
3429 extern int __kmp_clock_function_param;
3430 #endif /* KMP_OS_LINUX */
3431 
3432 #if KMP_MIC_SUPPORTED
3433 extern enum mic_type __kmp_mic_type;
3434 #endif
3435 
3436 #ifdef USE_LOAD_BALANCE
3437 extern double __kmp_load_balance_interval; // load balance algorithm interval
3438 #endif /* USE_LOAD_BALANCE */
3439 
3440 // OpenMP 3.1 - Nested num threads array
3441 typedef struct kmp_nested_nthreads_t {
3442  int *nth;
3443  int size;
3444  int used;
3445 } kmp_nested_nthreads_t;
3446 
3447 extern kmp_nested_nthreads_t __kmp_nested_nth;
3448 
3449 #if KMP_USE_ADAPTIVE_LOCKS
3450 
3451 // Parameters for the speculative lock backoff system.
3452 struct kmp_adaptive_backoff_params_t {
3453  // Number of soft retries before it counts as a hard retry.
3454  kmp_uint32 max_soft_retries;
3455  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3456  // the right
3457  kmp_uint32 max_badness;
3458 };
3459 
3460 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3461 
3462 #if KMP_DEBUG_ADAPTIVE_LOCKS
3463 extern const char *__kmp_speculative_statsfile;
3464 #endif
3465 
3466 #endif // KMP_USE_ADAPTIVE_LOCKS
3467 
3468 extern int __kmp_display_env; /* TRUE or FALSE */
3469 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3470 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3471 extern int __kmp_nteams;
3472 extern int __kmp_teams_thread_limit;
3473 
3474 /* ------------------------------------------------------------------------- */
3475 
3476 /* the following are protected by the fork/join lock */
3477 /* write: lock read: anytime */
3478 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3479 /* Holds old arrays of __kmp_threads until library shutdown */
3480 extern kmp_old_threads_list_t *__kmp_old_threads_list;
3481 /* read/write: lock */
3482 extern volatile kmp_team_t *__kmp_team_pool;
3483 extern volatile kmp_info_t *__kmp_thread_pool;
3484 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3485 
3486 // total num threads reachable from some root thread including all root threads
3487 extern volatile int __kmp_nth;
3488 /* total number of threads reachable from some root thread including all root
3489  threads, and those in the thread pool */
3490 extern volatile int __kmp_all_nth;
3491 extern std::atomic<int> __kmp_thread_pool_active_nth;
3492 
3493 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3494 /* end data protected by fork/join lock */
3495 /* ------------------------------------------------------------------------- */
3496 
3497 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3498 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3499 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3500 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3501 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3502 
3503 // AT: Which way is correct?
3504 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3505 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3506 #define __kmp_get_team_num_threads(gtid) \
3507  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3508 
3509 static inline bool KMP_UBER_GTID(int gtid) {
3510  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3511  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3512  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3513  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3514 }
3515 
3516 static inline int __kmp_tid_from_gtid(int gtid) {
3517  KMP_DEBUG_ASSERT(gtid >= 0);
3518  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3519 }
3520 
3521 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3522  KMP_DEBUG_ASSERT(tid >= 0 && team);
3523  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3524 }
3525 
3526 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3527  KMP_DEBUG_ASSERT(thr);
3528  return thr->th.th_info.ds.ds_gtid;
3529 }
3530 
3531 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3532  KMP_DEBUG_ASSERT(gtid >= 0);
3533  return __kmp_threads[gtid];
3534 }
3535 
3536 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3537  KMP_DEBUG_ASSERT(gtid >= 0);
3538  return __kmp_threads[gtid]->th.th_team;
3539 }
3540 
3541 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3542  if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3543  KMP_FATAL(ThreadIdentInvalid);
3544 }
3545 
3546 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3547 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3548 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3549 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3550 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3551 #endif
3552 
3553 #if KMP_HAVE_UMWAIT
3554 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3555 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3556 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3557 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3558 #endif
3559 
3560 /* ------------------------------------------------------------------------- */
3561 
3562 extern kmp_global_t __kmp_global; /* global status */
3563 
3564 extern kmp_info_t __kmp_monitor;
3565 // For Debugging Support Library
3566 extern std::atomic<kmp_int32> __kmp_team_counter;
3567 // For Debugging Support Library
3568 extern std::atomic<kmp_int32> __kmp_task_counter;
3569 
3570 #if USE_DEBUGGER
3571 #define _KMP_GEN_ID(counter) \
3572  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3573 #else
3574 #define _KMP_GEN_ID(counter) (~0)
3575 #endif /* USE_DEBUGGER */
3576 
3577 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3578 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3579 
3580 /* ------------------------------------------------------------------------ */
3581 
3582 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3583  size_t size, char const *format, ...);
3584 
3585 extern void __kmp_serial_initialize(void);
3586 extern void __kmp_middle_initialize(void);
3587 extern void __kmp_parallel_initialize(void);
3588 
3589 extern void __kmp_internal_begin(void);
3590 extern void __kmp_internal_end_library(int gtid);
3591 extern void __kmp_internal_end_thread(int gtid);
3592 extern void __kmp_internal_end_atexit(void);
3593 extern void __kmp_internal_end_dtor(void);
3594 extern void __kmp_internal_end_dest(void *);
3595 
3596 extern int __kmp_register_root(int initial_thread);
3597 extern void __kmp_unregister_root(int gtid);
3598 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3599 
3600 extern int __kmp_ignore_mppbeg(void);
3601 extern int __kmp_ignore_mppend(void);
3602 
3603 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3604 extern void __kmp_exit_single(int gtid);
3605 
3606 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3607 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3608 
3609 #ifdef USE_LOAD_BALANCE
3610 extern int __kmp_get_load_balance(int);
3611 #endif
3612 
3613 extern int __kmp_get_global_thread_id(void);
3614 extern int __kmp_get_global_thread_id_reg(void);
3615 extern void __kmp_exit_thread(int exit_status);
3616 extern void __kmp_abort(char const *format, ...);
3617 extern void __kmp_abort_thread(void);
3618 KMP_NORETURN extern void __kmp_abort_process(void);
3619 extern void __kmp_warn(char const *format, ...);
3620 
3621 extern void __kmp_set_num_threads(int new_nth, int gtid);
3622 
3623 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3624 // registered.
3625 static inline kmp_info_t *__kmp_entry_thread() {
3626  int gtid = __kmp_entry_gtid();
3627 
3628  return __kmp_threads[gtid];
3629 }
3630 
3631 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3632 extern int __kmp_get_max_active_levels(int gtid);
3633 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3634 extern int __kmp_get_team_size(int gtid, int level);
3635 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3636 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3637 
3638 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3639 extern void __kmp_init_random(kmp_info_t *thread);
3640 
3641 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3642 extern void __kmp_adjust_num_threads(int new_nproc);
3643 extern void __kmp_check_stksize(size_t *val);
3644 
3645 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3646 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3647 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3648 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3649 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3650 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3651 
3652 #if USE_FAST_MEMORY
3653 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3654  size_t size KMP_SRC_LOC_DECL);
3655 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3656 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3657 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3658 #define __kmp_fast_allocate(this_thr, size) \
3659  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3660 #define __kmp_fast_free(this_thr, ptr) \
3661  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3662 #endif
3663 
3664 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3665 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3666  size_t elsize KMP_SRC_LOC_DECL);
3667 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3668  size_t size KMP_SRC_LOC_DECL);
3669 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3670 #define __kmp_thread_malloc(th, size) \
3671  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3672 #define __kmp_thread_calloc(th, nelem, elsize) \
3673  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3674 #define __kmp_thread_realloc(th, ptr, size) \
3675  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3676 #define __kmp_thread_free(th, ptr) \
3677  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3678 
3679 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3680 
3681 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3682  kmp_proc_bind_t proc_bind);
3683 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3684  int num_threads);
3685 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3686  int num_teams_ub, int num_threads);
3687 
3688 extern void __kmp_yield();
3689 
3690 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3691  enum sched_type schedule, kmp_int32 lb,
3692  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3693 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3694  enum sched_type schedule, kmp_uint32 lb,
3695  kmp_uint32 ub, kmp_int32 st,
3696  kmp_int32 chunk);
3697 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3698  enum sched_type schedule, kmp_int64 lb,
3699  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3700 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3701  enum sched_type schedule, kmp_uint64 lb,
3702  kmp_uint64 ub, kmp_int64 st,
3703  kmp_int64 chunk);
3704 
3705 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3706  kmp_int32 *p_last, kmp_int32 *p_lb,
3707  kmp_int32 *p_ub, kmp_int32 *p_st);
3708 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3709  kmp_int32 *p_last, kmp_uint32 *p_lb,
3710  kmp_uint32 *p_ub, kmp_int32 *p_st);
3711 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3712  kmp_int32 *p_last, kmp_int64 *p_lb,
3713  kmp_int64 *p_ub, kmp_int64 *p_st);
3714 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3715  kmp_int32 *p_last, kmp_uint64 *p_lb,
3716  kmp_uint64 *p_ub, kmp_int64 *p_st);
3717 
3718 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3719 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3720 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3721 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3722 
3723 #ifdef KMP_GOMP_COMPAT
3724 
3725 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3726  enum sched_type schedule, kmp_int32 lb,
3727  kmp_int32 ub, kmp_int32 st,
3728  kmp_int32 chunk, int push_ws);
3729 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3730  enum sched_type schedule, kmp_uint32 lb,
3731  kmp_uint32 ub, kmp_int32 st,
3732  kmp_int32 chunk, int push_ws);
3733 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3734  enum sched_type schedule, kmp_int64 lb,
3735  kmp_int64 ub, kmp_int64 st,
3736  kmp_int64 chunk, int push_ws);
3737 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3738  enum sched_type schedule, kmp_uint64 lb,
3739  kmp_uint64 ub, kmp_int64 st,
3740  kmp_int64 chunk, int push_ws);
3741 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3742 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3743 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3744 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3745 
3746 #endif /* KMP_GOMP_COMPAT */
3747 
3748 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3749 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3750 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3751 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3752 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3753 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3754  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3755  void *obj);
3756 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3757  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3758 
3759 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3760  int final_spin
3761 #if USE_ITT_BUILD
3762  ,
3763  void *itt_sync_obj
3764 #endif
3765 );
3766 extern void __kmp_release_64(kmp_flag_64<> *flag);
3767 
3768 extern void __kmp_infinite_loop(void);
3769 
3770 extern void __kmp_cleanup(void);
3771 
3772 #if KMP_HANDLE_SIGNALS
3773 extern int __kmp_handle_signals;
3774 extern void __kmp_install_signals(int parallel_init);
3775 extern void __kmp_remove_signals(void);
3776 #endif
3777 
3778 extern void __kmp_clear_system_time(void);
3779 extern void __kmp_read_system_time(double *delta);
3780 
3781 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3782 
3783 extern void __kmp_expand_host_name(char *buffer, size_t size);
3784 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3785 
3786 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3787 extern void
3788 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3789 #endif
3790 
3791 extern void
3792 __kmp_runtime_initialize(void); /* machine specific initialization */
3793 extern void __kmp_runtime_destroy(void);
3794 
3795 #if KMP_AFFINITY_SUPPORTED
3796 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3797  kmp_affin_mask_t *mask);
3798 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3799  kmp_affin_mask_t *mask);
3800 extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3801 extern void __kmp_affinity_uninitialize(void);
3802 extern void __kmp_affinity_set_init_mask(
3803  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3804 void __kmp_affinity_bind_init_mask(int gtid);
3805 extern void __kmp_affinity_bind_place(int gtid);
3806 extern void __kmp_affinity_determine_capable(const char *env_var);
3807 extern int __kmp_aux_set_affinity(void **mask);
3808 extern int __kmp_aux_get_affinity(void **mask);
3809 extern int __kmp_aux_get_affinity_max_proc();
3810 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3811 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3812 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3813 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3814 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3815 extern int kmp_set_thread_affinity_mask_initial(void);
3816 #endif
3817 static inline void __kmp_assign_root_init_mask() {
3818  int gtid = __kmp_entry_gtid();
3819  kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3820  if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3821  __kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3822  __kmp_affinity_bind_init_mask(gtid);
3823  r->r.r_affinity_assigned = TRUE;
3824  }
3825 }
3826 static inline void __kmp_reset_root_init_mask(int gtid) {
3827  if (!KMP_AFFINITY_CAPABLE())
3828  return;
3829  kmp_info_t *th = __kmp_threads[gtid];
3830  kmp_root_t *r = th->th.th_root;
3831  if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3832  __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3833  KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3834  r->r.r_affinity_assigned = FALSE;
3835  }
3836 }
3837 #else /* KMP_AFFINITY_SUPPORTED */
3838 #define __kmp_assign_root_init_mask() /* Nothing */
3839 static inline void __kmp_reset_root_init_mask(int gtid) {}
3840 #endif /* KMP_AFFINITY_SUPPORTED */
3841 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3842 // format string is for affinity, so platforms that do not support
3843 // affinity can still use the other fields, e.g., %n for num_threads
3844 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3845  kmp_str_buf_t *buffer);
3846 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3847 
3848 extern void __kmp_cleanup_hierarchy();
3849 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3850 
3851 #if KMP_USE_FUTEX
3852 
3853 extern int __kmp_futex_determine_capable(void);
3854 
3855 #endif // KMP_USE_FUTEX
3856 
3857 extern void __kmp_gtid_set_specific(int gtid);
3858 extern int __kmp_gtid_get_specific(void);
3859 
3860 extern double __kmp_read_cpu_time(void);
3861 
3862 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3863 
3864 #if KMP_USE_MONITOR
3865 extern void __kmp_create_monitor(kmp_info_t *th);
3866 #endif
3867 
3868 extern void *__kmp_launch_thread(kmp_info_t *thr);
3869 
3870 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3871 
3872 #if KMP_OS_WINDOWS
3873 extern int __kmp_still_running(kmp_info_t *th);
3874 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3875 extern void __kmp_free_handle(kmp_thread_t tHandle);
3876 #endif
3877 
3878 #if KMP_USE_MONITOR
3879 extern void __kmp_reap_monitor(kmp_info_t *th);
3880 #endif
3881 extern void __kmp_reap_worker(kmp_info_t *th);
3882 extern void __kmp_terminate_thread(int gtid);
3883 
3884 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3885 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3886 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3887 
3888 extern void __kmp_elapsed(double *);
3889 extern void __kmp_elapsed_tick(double *);
3890 
3891 extern void __kmp_enable(int old_state);
3892 extern void __kmp_disable(int *old_state);
3893 
3894 extern void __kmp_thread_sleep(int millis);
3895 
3896 extern void __kmp_common_initialize(void);
3897 extern void __kmp_common_destroy(void);
3898 extern void __kmp_common_destroy_gtid(int gtid);
3899 
3900 #if KMP_OS_UNIX
3901 extern void __kmp_register_atfork(void);
3902 #endif
3903 extern void __kmp_suspend_initialize(void);
3904 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3905 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3906 
3907 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3908  int tid);
3909 extern kmp_team_t *
3910 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3911 #if OMPT_SUPPORT
3912  ompt_data_t ompt_parallel_data,
3913 #endif
3914  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3915  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3916 extern void __kmp_free_thread(kmp_info_t *);
3917 extern void __kmp_free_team(kmp_root_t *,
3918  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3919 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3920 
3921 /* ------------------------------------------------------------------------ */
3922 
3923 extern void __kmp_initialize_bget(kmp_info_t *th);
3924 extern void __kmp_finalize_bget(kmp_info_t *th);
3925 
3926 KMP_EXPORT void *kmpc_malloc(size_t size);
3927 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3928 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3929 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3930 KMP_EXPORT void kmpc_free(void *ptr);
3931 
3932 /* declarations for internal use */
3933 
3934 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3935  size_t reduce_size, void *reduce_data,
3936  void (*reduce)(void *, void *));
3937 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3938 extern int __kmp_barrier_gomp_cancel(int gtid);
3939 
3944 enum fork_context_e {
3945  fork_context_gnu,
3947  fork_context_intel,
3948  fork_context_last
3949 };
3950 extern int __kmp_fork_call(ident_t *loc, int gtid,
3951  enum fork_context_e fork_context, kmp_int32 argc,
3952  microtask_t microtask, launch_t invoker,
3953  kmp_va_list ap);
3954 
3955 extern void __kmp_join_call(ident_t *loc, int gtid
3956 #if OMPT_SUPPORT
3957  ,
3958  enum fork_context_e fork_context
3959 #endif
3960  ,
3961  int exit_teams = 0);
3962 
3963 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3964 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3965 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3966 extern int __kmp_invoke_task_func(int gtid);
3967 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3968  kmp_info_t *this_thr,
3969  kmp_team_t *team);
3970 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3971  kmp_info_t *this_thr,
3972  kmp_team_t *team);
3973 
3974 // should never have been exported
3975 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3976 extern int __kmp_invoke_teams_master(int gtid);
3977 extern void __kmp_teams_master(int gtid);
3978 extern int __kmp_aux_get_team_num();
3979 extern int __kmp_aux_get_num_teams();
3980 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3981 extern void __kmp_user_set_library(enum library_type arg);
3982 extern void __kmp_aux_set_library(enum library_type arg);
3983 extern void __kmp_aux_set_stacksize(size_t arg);
3984 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3985 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3986 
3987 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3988 void kmpc_set_blocktime(int arg);
3989 void ompc_set_nested(int flag);
3990 void ompc_set_dynamic(int flag);
3991 void ompc_set_num_threads(int arg);
3992 
3993 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3994  kmp_team_t *team, int tid);
3995 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3996 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3997  kmp_tasking_flags_t *flags,
3998  size_t sizeof_kmp_task_t,
3999  size_t sizeof_shareds,
4000  kmp_routine_entry_t task_entry);
4001 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4002  kmp_team_t *team, int tid,
4003  int set_curr_task);
4004 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4005 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4006 
4007 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4008  int gtid,
4009  kmp_task_t *task);
4010 extern void __kmp_fulfill_event(kmp_event_t *event);
4011 
4012 extern void __kmp_free_task_team(kmp_info_t *thread,
4013  kmp_task_team_t *task_team);
4014 extern void __kmp_reap_task_teams(void);
4015 extern void __kmp_wait_to_unref_task_teams(void);
4016 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
4017  int always);
4018 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4019 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4020 #if USE_ITT_BUILD
4021  ,
4022  void *itt_sync_obj
4023 #endif /* USE_ITT_BUILD */
4024  ,
4025  int wait = 1);
4026 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4027  int gtid);
4028 
4029 extern int __kmp_is_address_mapped(void *addr);
4030 extern kmp_uint64 __kmp_hardware_timestamp(void);
4031 
4032 #if KMP_OS_UNIX
4033 extern int __kmp_read_from_file(char const *path, char const *format, ...);
4034 #endif
4035 
4036 /* ------------------------------------------------------------------------ */
4037 //
4038 // Assembly routines that have no compiler intrinsic replacement
4039 //
4040 
4041 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4042  void *argv[]
4043 #if OMPT_SUPPORT
4044  ,
4045  void **exit_frame_ptr
4046 #endif
4047 );
4048 
4049 /* ------------------------------------------------------------------------ */
4050 
4051 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4052 KMP_EXPORT void __kmpc_end(ident_t *);
4053 
4054 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4055  kmpc_ctor_vec ctor,
4056  kmpc_cctor_vec cctor,
4057  kmpc_dtor_vec dtor,
4058  size_t vector_length);
4059 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4060  kmpc_ctor ctor, kmpc_cctor cctor,
4061  kmpc_dtor dtor);
4062 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4063  void *data, size_t size);
4064 
4065 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4066 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4067 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4068 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4069 
4070 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4071 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4072  kmpc_micro microtask, ...);
4073 KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4074  kmpc_micro microtask, kmp_int32 cond,
4075  void *args);
4076 
4077 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4078 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4079 
4080 KMP_EXPORT void __kmpc_flush(ident_t *);
4081 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4082 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4083 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4084 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4085  kmp_int32 filter);
4086 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4087 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4088 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4089 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4090  kmp_critical_name *);
4091 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4092  kmp_critical_name *);
4093 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4094  kmp_critical_name *, uint32_t hint);
4095 
4096 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4097 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4098 
4099 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4100  kmp_int32 global_tid);
4101 
4102 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4103 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4104 
4105 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4106 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4107  kmp_int32 numberOfSections);
4108 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4109 
4110 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4111  kmp_int32 schedtype, kmp_int32 *plastiter,
4112  kmp_int *plower, kmp_int *pupper,
4113  kmp_int *pstride, kmp_int incr,
4114  kmp_int chunk);
4115 
4116 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4117 
4118 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4119  size_t cpy_size, void *cpy_data,
4120  void (*cpy_func)(void *, void *),
4121  kmp_int32 didit);
4122 
4123 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4124  void *cpy_data);
4125 
4126 extern void KMPC_SET_NUM_THREADS(int arg);
4127 extern void KMPC_SET_DYNAMIC(int flag);
4128 extern void KMPC_SET_NESTED(int flag);
4129 
4130 /* OMP 3.0 tasking interface routines */
4131 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4132  kmp_task_t *new_task);
4133 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4134  kmp_int32 flags,
4135  size_t sizeof_kmp_task_t,
4136  size_t sizeof_shareds,
4137  kmp_routine_entry_t task_entry);
4138 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4139  ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4140  size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4141 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4142  kmp_task_t *task);
4143 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4144  kmp_task_t *task);
4145 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4146  kmp_task_t *new_task);
4147 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4148 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4149  int end_part);
4150 
4151 #if TASK_UNUSED
4152 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4153 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4154  kmp_task_t *task);
4155 #endif // TASK_UNUSED
4156 
4157 /* ------------------------------------------------------------------------ */
4158 
4159 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4160 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4161 
4162 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4163  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4164  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4165  kmp_depend_info_t *noalias_dep_list);
4166 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4167  kmp_int32 ndeps,
4168  kmp_depend_info_t *dep_list,
4169  kmp_int32 ndeps_noalias,
4170  kmp_depend_info_t *noalias_dep_list);
4171 /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4172  * Placeholder for taskwait with nowait clause.*/
4173 KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4174  kmp_int32 ndeps,
4175  kmp_depend_info_t *dep_list,
4176  kmp_int32 ndeps_noalias,
4177  kmp_depend_info_t *noalias_dep_list,
4178  kmp_int32 has_no_wait);
4179 
4180 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4181  bool serialize_immediate);
4182 
4183 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4184  kmp_int32 cncl_kind);
4185 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4186  kmp_int32 cncl_kind);
4187 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4188 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4189 
4190 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4191 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4192 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4193  kmp_int32 if_val, kmp_uint64 *lb,
4194  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4195  kmp_int32 sched, kmp_uint64 grainsize,
4196  void *task_dup);
4197 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4198  kmp_task_t *task, kmp_int32 if_val,
4199  kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4200  kmp_int32 nogroup, kmp_int32 sched,
4201  kmp_uint64 grainsize, kmp_int32 modifier,
4202  void *task_dup);
4203 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4204 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4205 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4206 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4207  int is_ws, int num,
4208  void *data);
4209 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4210  int num, void *data);
4211 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4212  int is_ws);
4213 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4214  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4215  kmp_task_affinity_info_t *affin_list);
4216 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4217 KMP_EXPORT int __kmp_get_max_teams(void);
4218 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4219 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4220 
4221 /* Interface target task integration */
4222 KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4223 KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4224 
4225 /* Lock interface routines (fast versions with gtid passed in) */
4226 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4227  void **user_lock);
4228 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4229  void **user_lock);
4230 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4231  void **user_lock);
4232 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4233  void **user_lock);
4234 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4235 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4236  void **user_lock);
4237 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4238  void **user_lock);
4239 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4240  void **user_lock);
4241 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4242 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4243  void **user_lock);
4244 
4245 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4246  void **user_lock, uintptr_t hint);
4247 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4248  void **user_lock,
4249  uintptr_t hint);
4250 
4251 #if OMPX_TASKGRAPH
4252 // Taskgraph's Record & Replay mechanism
4253 // __kmp_tdg_is_recording: check whether a given TDG is recording
4254 // status: the tdg's current status
4255 static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4256  return status == KMP_TDG_RECORDING;
4257 }
4258 
4259 KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4260  kmp_int32 input_flags,
4261  kmp_int32 tdg_id);
4262 KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4263  kmp_int32 input_flags, kmp_int32 tdg_id);
4264 #endif
4265 /* Interface to fast scalable reduce methods routines */
4266 
4267 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4268  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4269  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4270  kmp_critical_name *lck);
4271 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4272  kmp_critical_name *lck);
4273 KMP_EXPORT kmp_int32 __kmpc_reduce(
4274  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4275  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4276  kmp_critical_name *lck);
4277 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4278  kmp_critical_name *lck);
4279 
4280 /* Internal fast reduction routines */
4281 
4282 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4283  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4284  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4285  kmp_critical_name *lck);
4286 
4287 // this function is for testing set/get/determine reduce method
4288 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4289 
4290 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4291 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4292 
4293 // C++ port
4294 // missing 'extern "C"' declarations
4295 
4296 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4297 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4298 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4299  kmp_int32 num_threads);
4300 
4301 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4302  int proc_bind);
4303 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4304  kmp_int32 num_teams,
4305  kmp_int32 num_threads);
4306 KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4307  kmp_int32 thread_limit);
4308 /* Function for OpenMP 5.1 num_teams clause */
4309 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4310  kmp_int32 num_teams_lb,
4311  kmp_int32 num_teams_ub,
4312  kmp_int32 num_threads);
4313 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4314  kmpc_micro microtask, ...);
4315 struct kmp_dim { // loop bounds info casted to kmp_int64
4316  kmp_int64 lo; // lower
4317  kmp_int64 up; // upper
4318  kmp_int64 st; // stride
4319 };
4320 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4321  kmp_int32 num_dims,
4322  const struct kmp_dim *dims);
4323 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4324  const kmp_int64 *vec);
4325 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4326  const kmp_int64 *vec);
4327 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4328 
4329 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4330  void *data, size_t size,
4331  void ***cache);
4332 
4333 // The routines below are not exported.
4334 // Consider making them 'static' in corresponding source files.
4335 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4336  void *data_addr, size_t pc_size);
4337 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4338  void *data_addr,
4339  size_t pc_size);
4340 void __kmp_threadprivate_resize_cache(int newCapacity);
4341 void __kmp_cleanup_threadprivate_caches();
4342 
4343 // ompc_, kmpc_ entries moved from omp.h.
4344 #if KMP_OS_WINDOWS
4345 #define KMPC_CONVENTION __cdecl
4346 #else
4347 #define KMPC_CONVENTION
4348 #endif
4349 
4350 #ifndef __OMP_H
4351 typedef enum omp_sched_t {
4352  omp_sched_static = 1,
4353  omp_sched_dynamic = 2,
4354  omp_sched_guided = 3,
4355  omp_sched_auto = 4
4356 } omp_sched_t;
4357 typedef void *kmp_affinity_mask_t;
4358 #endif
4359 
4360 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4361 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4362 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4363 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4364 KMP_EXPORT int KMPC_CONVENTION
4365 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4366 KMP_EXPORT int KMPC_CONVENTION
4367 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4368 KMP_EXPORT int KMPC_CONVENTION
4369 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4370 
4371 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4372 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4373 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4374 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4375 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4376 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4377 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4378 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4379 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4380  char const *format);
4381 
4382 enum kmp_target_offload_kind {
4383  tgt_disabled = 0,
4384  tgt_default = 1,
4385  tgt_mandatory = 2
4386 };
4387 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4388 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4389 extern kmp_target_offload_kind_t __kmp_target_offload;
4390 extern int __kmpc_get_target_offload();
4391 
4392 // Constants used in libomptarget
4393 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4394 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4395 
4396 // OMP Pause Resource
4397 
4398 // The following enum is used both to set the status in __kmp_pause_status, and
4399 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4400 typedef enum kmp_pause_status_t {
4401  kmp_not_paused = 0, // status is not paused, or, requesting resume
4402  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4403  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4404 } kmp_pause_status_t;
4405 
4406 // This stores the pause state of the runtime
4407 extern kmp_pause_status_t __kmp_pause_status;
4408 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4409 extern int __kmp_pause_resource(kmp_pause_status_t level);
4410 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4411 extern void __kmp_resume_if_soft_paused();
4412 // Hard resume simply resets the status to not paused. Library will appear to
4413 // be uninitialized after hard pause. Let OMP constructs trigger required
4414 // initializations.
4415 static inline void __kmp_resume_if_hard_paused() {
4416  if (__kmp_pause_status == kmp_hard_paused) {
4417  __kmp_pause_status = kmp_not_paused;
4418  }
4419 }
4420 
4421 extern void __kmp_omp_display_env(int verbose);
4422 
4423 // 1: it is initializing hidden helper team
4424 extern volatile int __kmp_init_hidden_helper;
4425 // 1: the hidden helper team is done
4426 extern volatile int __kmp_hidden_helper_team_done;
4427 // 1: enable hidden helper task
4428 extern kmp_int32 __kmp_enable_hidden_helper;
4429 // Main thread of hidden helper team
4430 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4431 // Descriptors for the hidden helper threads
4432 extern kmp_info_t **__kmp_hidden_helper_threads;
4433 // Number of hidden helper threads
4434 extern kmp_int32 __kmp_hidden_helper_threads_num;
4435 // Number of hidden helper tasks that have not been executed yet
4436 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4437 
4438 extern void __kmp_hidden_helper_initialize();
4439 extern void __kmp_hidden_helper_threads_initz_routine();
4440 extern void __kmp_do_initialize_hidden_helper_threads();
4441 extern void __kmp_hidden_helper_threads_initz_wait();
4442 extern void __kmp_hidden_helper_initz_release();
4443 extern void __kmp_hidden_helper_threads_deinitz_wait();
4444 extern void __kmp_hidden_helper_threads_deinitz_release();
4445 extern void __kmp_hidden_helper_main_thread_wait();
4446 extern void __kmp_hidden_helper_worker_thread_wait();
4447 extern void __kmp_hidden_helper_worker_thread_signal();
4448 extern void __kmp_hidden_helper_main_thread_release();
4449 
4450 // Check whether a given thread is a hidden helper thread
4451 #define KMP_HIDDEN_HELPER_THREAD(gtid) \
4452  ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4453 
4454 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4455  ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4456 
4457 #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4458  ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4459 
4460 #define KMP_HIDDEN_HELPER_TEAM(team) \
4461  (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4462 
4463 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4464 // main thread, is skipped.
4465 #define KMP_GTID_TO_SHADOW_GTID(gtid) \
4466  ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4467 
4468 // Return the adjusted gtid value by subtracting from gtid the number
4469 // of hidden helper threads. This adjusted value is the gtid the thread would
4470 // have received if there were no hidden helper threads.
4471 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4472  int adjusted_gtid = gtid;
4473  if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4474  gtid - __kmp_hidden_helper_threads_num >= 0) {
4475  adjusted_gtid -= __kmp_hidden_helper_threads_num;
4476  }
4477  return adjusted_gtid;
4478 }
4479 
4480 // Support for error directive
4481 typedef enum kmp_severity_t {
4482  severity_warning = 1,
4483  severity_fatal = 2
4484 } kmp_severity_t;
4485 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4486 
4487 // Support for scope directive
4488 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4489 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4490 
4491 #ifdef __cplusplus
4492 }
4493 #endif
4494 
4495 template <bool C, bool S>
4496 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4497 template <bool C, bool S>
4498 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4499 template <bool C, bool S>
4500 extern void __kmp_atomic_suspend_64(int th_gtid,
4501  kmp_atomic_flag_64<C, S> *flag);
4502 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4503 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4504 template <bool C, bool S>
4505 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4506 template <bool C, bool S>
4507 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4508 template <bool C, bool S>
4509 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4510 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4511 #endif
4512 template <bool C, bool S>
4513 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4514 template <bool C, bool S>
4515 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4516 template <bool C, bool S>
4517 extern void __kmp_atomic_resume_64(int target_gtid,
4518  kmp_atomic_flag_64<C, S> *flag);
4519 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4520 
4521 template <bool C, bool S>
4522 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4523  kmp_flag_32<C, S> *flag, int final_spin,
4524  int *thread_finished,
4525 #if USE_ITT_BUILD
4526  void *itt_sync_obj,
4527 #endif /* USE_ITT_BUILD */
4528  kmp_int32 is_constrained);
4529 template <bool C, bool S>
4530 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4531  kmp_flag_64<C, S> *flag, int final_spin,
4532  int *thread_finished,
4533 #if USE_ITT_BUILD
4534  void *itt_sync_obj,
4535 #endif /* USE_ITT_BUILD */
4536  kmp_int32 is_constrained);
4537 template <bool C, bool S>
4538 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4539  kmp_atomic_flag_64<C, S> *flag,
4540  int final_spin, int *thread_finished,
4541 #if USE_ITT_BUILD
4542  void *itt_sync_obj,
4543 #endif /* USE_ITT_BUILD */
4544  kmp_int32 is_constrained);
4545 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4546  kmp_flag_oncore *flag, int final_spin,
4547  int *thread_finished,
4548 #if USE_ITT_BUILD
4549  void *itt_sync_obj,
4550 #endif /* USE_ITT_BUILD */
4551  kmp_int32 is_constrained);
4552 
4553 extern int __kmp_nesting_mode;
4554 extern int __kmp_nesting_mode_nlevels;
4555 extern int *__kmp_nesting_nth_level;
4556 extern void __kmp_init_nesting_mode();
4557 extern void __kmp_set_nesting_mode_threads();
4558 
4566  FILE *f;
4567 
4568  void close() {
4569  if (f && f != stdout && f != stderr) {
4570  fclose(f);
4571  f = nullptr;
4572  }
4573  }
4574 
4575 public:
4576  kmp_safe_raii_file_t() : f(nullptr) {}
4577  kmp_safe_raii_file_t(const char *filename, const char *mode,
4578  const char *env_var = nullptr)
4579  : f(nullptr) {
4580  open(filename, mode, env_var);
4581  }
4582  ~kmp_safe_raii_file_t() { close(); }
4583 
4587  void open(const char *filename, const char *mode,
4588  const char *env_var = nullptr) {
4589  KMP_ASSERT(!f);
4590  f = fopen(filename, mode);
4591  if (!f) {
4592  int code = errno;
4593  if (env_var) {
4594  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4595  KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4596  } else {
4597  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4598  __kmp_msg_null);
4599  }
4600  }
4601  }
4604  int try_open(const char *filename, const char *mode) {
4605  KMP_ASSERT(!f);
4606  f = fopen(filename, mode);
4607  if (!f)
4608  return errno;
4609  return 0;
4610  }
4613  void set_stdout() {
4614  KMP_ASSERT(!f);
4615  f = stdout;
4616  }
4619  void set_stderr() {
4620  KMP_ASSERT(!f);
4621  f = stderr;
4622  }
4623  operator bool() { return bool(f); }
4624  operator FILE *() { return f; }
4625 };
4626 
4627 template <typename SourceType, typename TargetType,
4628  bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4629  bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4630  bool isSourceSigned = std::is_signed<SourceType>::value,
4631  bool isTargetSigned = std::is_signed<TargetType>::value>
4632 struct kmp_convert {};
4633 
4634 // Both types are signed; Source smaller
4635 template <typename SourceType, typename TargetType>
4636 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4637  static TargetType to(SourceType src) { return (TargetType)src; }
4638 };
4639 // Source equal
4640 template <typename SourceType, typename TargetType>
4641 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4642  static TargetType to(SourceType src) { return src; }
4643 };
4644 // Source bigger
4645 template <typename SourceType, typename TargetType>
4646 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4647  static TargetType to(SourceType src) {
4648  KMP_ASSERT(src <= static_cast<SourceType>(
4649  (std::numeric_limits<TargetType>::max)()));
4650  KMP_ASSERT(src >= static_cast<SourceType>(
4651  (std::numeric_limits<TargetType>::min)()));
4652  return (TargetType)src;
4653  }
4654 };
4655 
4656 // Source signed, Target unsigned
4657 // Source smaller
4658 template <typename SourceType, typename TargetType>
4659 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4660  static TargetType to(SourceType src) {
4661  KMP_ASSERT(src >= 0);
4662  return (TargetType)src;
4663  }
4664 };
4665 // Source equal
4666 template <typename SourceType, typename TargetType>
4667 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4668  static TargetType to(SourceType src) {
4669  KMP_ASSERT(src >= 0);
4670  return (TargetType)src;
4671  }
4672 };
4673 // Source bigger
4674 template <typename SourceType, typename TargetType>
4675 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4676  static TargetType to(SourceType src) {
4677  KMP_ASSERT(src >= 0);
4678  KMP_ASSERT(src <= static_cast<SourceType>(
4679  (std::numeric_limits<TargetType>::max)()));
4680  return (TargetType)src;
4681  }
4682 };
4683 
4684 // Source unsigned, Target signed
4685 // Source smaller
4686 template <typename SourceType, typename TargetType>
4687 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4688  static TargetType to(SourceType src) { return (TargetType)src; }
4689 };
4690 // Source equal
4691 template <typename SourceType, typename TargetType>
4692 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4693  static TargetType to(SourceType src) {
4694  KMP_ASSERT(src <= static_cast<SourceType>(
4695  (std::numeric_limits<TargetType>::max)()));
4696  return (TargetType)src;
4697  }
4698 };
4699 // Source bigger
4700 template <typename SourceType, typename TargetType>
4701 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4702  static TargetType to(SourceType src) {
4703  KMP_ASSERT(src <= static_cast<SourceType>(
4704  (std::numeric_limits<TargetType>::max)()));
4705  return (TargetType)src;
4706  }
4707 };
4708 
4709 // Source unsigned, Target unsigned
4710 // Source smaller
4711 template <typename SourceType, typename TargetType>
4712 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4713  static TargetType to(SourceType src) { return (TargetType)src; }
4714 };
4715 // Source equal
4716 template <typename SourceType, typename TargetType>
4717 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4718  static TargetType to(SourceType src) { return src; }
4719 };
4720 // Source bigger
4721 template <typename SourceType, typename TargetType>
4722 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4723  static TargetType to(SourceType src) {
4724  KMP_ASSERT(src <= static_cast<SourceType>(
4725  (std::numeric_limits<TargetType>::max)()));
4726  return (TargetType)src;
4727  }
4728 };
4729 
4730 template <typename T1, typename T2>
4731 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4732  *dest = kmp_convert<T1, T2>::to(src);
4733 }
4734 
4735 #endif /* KMP_H */
void set_stdout()
Definition: kmp.h:4613
void set_stderr()
Definition: kmp.h:4619
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4604
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4587
struct ident ident_t
@ KMP_IDENT_KMPC
Definition: kmp.h:197
@ KMP_IDENT_IMB
Definition: kmp.h:195
@ KMP_IDENT_WORK_LOOP
Definition: kmp.h:215
@ KMP_IDENT_BARRIER_IMPL
Definition: kmp.h:206
@ KMP_IDENT_WORK_SECTIONS
Definition: kmp.h:217
@ KMP_IDENT_AUTOPAR
Definition: kmp.h:200
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition: kmp.h:224
@ KMP_IDENT_WORK_DISTRIBUTE
Definition: kmp.h:219
@ KMP_IDENT_BARRIER_EXPL
Definition: kmp.h:204
@ KMP_IDENT_ATOMIC_REDUCE
Definition: kmp.h:202
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs, kmpc_micro microtask, kmp_int32 cond, void *args)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid, kmp_int32 thread_limit)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1705
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)
KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
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)
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_flush(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
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)
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)
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid)
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
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)
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)
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
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)
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
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)
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
KMP_EXPORT void ** __kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid)
void(* kmpc_dtor)(void *)
Definition: kmp.h:1729
KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor)
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)
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1723
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1746
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1736
KMP_EXPORT void * __kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, void *data, size_t size, void ***cache)
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition: kmp.h:1758
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1752
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)
KMP_EXPORT void * __kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid, void *cpy_data)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
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)
sched_type
Definition: kmp.h:358
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, kmp_critical_name *, uint32_t hint)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, kmp_int32 numberOfSections)
void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims, const struct kmp_dim *dims)
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)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
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)
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid)
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)
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, kmp_int32 filter)
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)
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)
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)
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
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)
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
@ kmp_nm_guided_chunked
Definition: kmp.h:409
@ kmp_sch_runtime_simd
Definition: kmp.h:380
@ kmp_nm_ord_auto
Definition: kmp.h:428
@ kmp_sch_auto
Definition: kmp.h:365
@ kmp_nm_auto
Definition: kmp.h:411
@ kmp_distribute_static_chunked
Definition: kmp.h:396
@ kmp_sch_static
Definition: kmp.h:361
@ kmp_sch_guided_simd
Definition: kmp.h:379
@ kmp_sch_modifier_monotonic
Definition: kmp.h:446
@ kmp_sch_default
Definition: kmp.h:466
@ kmp_sch_modifier_nonmonotonic
Definition: kmp.h:448
@ kmp_nm_ord_static
Definition: kmp.h:424
@ kmp_distribute_static
Definition: kmp.h:397
@ kmp_sch_guided_chunked
Definition: kmp.h:363
@ kmp_nm_static
Definition: kmp.h:407
@ kmp_sch_lower
Definition: kmp.h:359
@ kmp_nm_upper
Definition: kmp.h:430
@ kmp_ord_lower
Definition: kmp.h:385
@ kmp_ord_static
Definition: kmp.h:387
@ kmp_sch_upper
Definition: kmp.h:383
@ kmp_ord_upper
Definition: kmp.h:393
@ kmp_nm_lower
Definition: kmp.h:403
@ kmp_ord_auto
Definition: kmp.h:391
Definition: kmp.h:235
kmp_int32 reserved_1
Definition: kmp.h:236
char const * psource
Definition: kmp.h:245
kmp_int32 reserved_2
Definition: kmp.h:239
kmp_int32 reserved_3
Definition: kmp.h:244
kmp_int32 flags
Definition: kmp.h:237