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 <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
70  Microsoft library. Some macros provided below to replace these functions */
71 #ifndef __ABSOFT_WIN
72 #include <sys/types.h>
73 #endif
74 #include <limits.h>
75 #include <time.h>
76 
77 #include <errno.h>
78 
79 #include "kmp_os.h"
80 
81 #include "kmp_safe_c_api.h"
82 
83 #if KMP_STATS_ENABLED
84 class kmp_stats_list;
85 #endif
86 
87 #if KMP_USE_HIER_SCHED
88 // Only include hierarchical scheduling if affinity is supported
89 #undef KMP_USE_HIER_SCHED
90 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
91 #endif
92 
93 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
94 #include "hwloc.h"
95 #ifndef HWLOC_OBJ_NUMANODE
96 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
97 #endif
98 #ifndef HWLOC_OBJ_PACKAGE
99 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
100 #endif
101 #if HWLOC_API_VERSION >= 0x00020000
102 // hwloc 2.0 changed type of depth of object from unsigned to int
103 typedef int kmp_hwloc_depth_t;
104 #else
105 typedef unsigned int kmp_hwloc_depth_t;
106 #endif
107 #endif
108 
109 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
110 #include <xmmintrin.h>
111 #endif
112 
113 #include "kmp_debug.h"
114 #include "kmp_lock.h"
115 #include "kmp_version.h"
116 #if USE_DEBUGGER
117 #include "kmp_debugger.h"
118 #endif
119 #include "kmp_i18n.h"
120 
121 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
122 
123 #include "kmp_wrapper_malloc.h"
124 #if KMP_OS_UNIX
125 #include <unistd.h>
126 #if !defined NSIG && defined _NSIG
127 #define NSIG _NSIG
128 #endif
129 #endif
130 
131 #if KMP_OS_LINUX
132 #pragma weak clock_gettime
133 #endif
134 
135 #if OMPT_SUPPORT
136 #include "ompt-internal.h"
137 #endif
138 
139 #ifndef UNLIKELY
140 #define UNLIKELY(x) (x)
141 #endif
142 
143 // Affinity format function
144 #include "kmp_str.h"
145 
146 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
147 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
148 // free lists of limited size.
149 #ifndef USE_FAST_MEMORY
150 #define USE_FAST_MEMORY 3
151 #endif
152 
153 #ifndef KMP_NESTED_HOT_TEAMS
154 #define KMP_NESTED_HOT_TEAMS 0
155 #define USE_NESTED_HOT_ARG(x)
156 #else
157 #if KMP_NESTED_HOT_TEAMS
158 #define USE_NESTED_HOT_ARG(x) , x
159 #else
160 #define USE_NESTED_HOT_ARG(x)
161 #endif
162 #endif
163 
164 // Assume using BGET compare_exchange instruction instead of lock by default.
165 #ifndef USE_CMP_XCHG_FOR_BGET
166 #define USE_CMP_XCHG_FOR_BGET 1
167 #endif
168 
169 // Test to see if queuing lock is better than bootstrap lock for bget
170 // #ifndef USE_QUEUING_LOCK_FOR_BGET
171 // #define USE_QUEUING_LOCK_FOR_BGET
172 // #endif
173 
174 #define KMP_NSEC_PER_SEC 1000000000L
175 #define KMP_USEC_PER_SEC 1000000L
176 
185 enum {
190  /* 0x04 is no longer used */
199  KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
200  KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
201  KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
202 
203  KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
204  KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
205 
217  KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
218  KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
219  KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
220  KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
221 };
222 
226 typedef struct ident {
227  kmp_int32 reserved_1;
228  kmp_int32 flags;
230  kmp_int32 reserved_2;
231 #if USE_ITT_BUILD
232 /* but currently used for storing region-specific ITT */
233 /* contextual information. */
234 #endif /* USE_ITT_BUILD */
235  kmp_int32 reserved_3;
236  char const *psource;
240 } ident_t;
245 // Some forward declarations.
246 typedef union kmp_team kmp_team_t;
247 typedef struct kmp_taskdata kmp_taskdata_t;
248 typedef union kmp_task_team kmp_task_team_t;
249 typedef union kmp_team kmp_team_p;
250 typedef union kmp_info kmp_info_p;
251 typedef union kmp_root kmp_root_p;
252 
253 #ifdef __cplusplus
254 extern "C" {
255 #endif
256 
257 /* ------------------------------------------------------------------------ */
258 
259 /* Pack two 32-bit signed integers into a 64-bit signed integer */
260 /* ToDo: Fix word ordering for big-endian machines. */
261 #define KMP_PACK_64(HIGH_32, LOW_32) \
262  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
263 
264 // Generic string manipulation macros. Assume that _x is of type char *
265 #define SKIP_WS(_x) \
266  { \
267  while (*(_x) == ' ' || *(_x) == '\t') \
268  (_x)++; \
269  }
270 #define SKIP_DIGITS(_x) \
271  { \
272  while (*(_x) >= '0' && *(_x) <= '9') \
273  (_x)++; \
274  }
275 #define SKIP_TOKEN(_x) \
276  { \
277  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
278  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
279  (_x)++; \
280  }
281 #define SKIP_TO(_x, _c) \
282  { \
283  while (*(_x) != '\0' && *(_x) != (_c)) \
284  (_x)++; \
285  }
286 
287 /* ------------------------------------------------------------------------ */
288 
289 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
290 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
291 
292 /* ------------------------------------------------------------------------ */
293 /* Enumeration types */
294 
295 enum kmp_state_timer {
296  ts_stop,
297  ts_start,
298  ts_pause,
299 
300  ts_last_state
301 };
302 
303 enum dynamic_mode {
304  dynamic_default,
305 #ifdef USE_LOAD_BALANCE
306  dynamic_load_balance,
307 #endif /* USE_LOAD_BALANCE */
308  dynamic_random,
309  dynamic_thread_limit,
310  dynamic_max
311 };
312 
313 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
314  * not include it here */
315 #ifndef KMP_SCHED_TYPE_DEFINED
316 #define KMP_SCHED_TYPE_DEFINED
317 typedef enum kmp_sched {
318  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
319  // Note: need to adjust __kmp_sch_map global array in case enum is changed
320  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
321  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
322  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
323  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
324  kmp_sched_upper_std = 5, // upper bound for standard schedules
325  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
326  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
327 #if KMP_STATIC_STEAL_ENABLED
328  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
329 #endif
330  kmp_sched_upper,
331  kmp_sched_default = kmp_sched_static, // default scheduling
332  kmp_sched_monotonic = 0x80000000
333 } kmp_sched_t;
334 #endif
335 
340 enum sched_type : kmp_int32 {
342  kmp_sch_static_chunked = 33,
344  kmp_sch_dynamic_chunked = 35,
346  kmp_sch_runtime = 37,
348  kmp_sch_trapezoidal = 39,
349 
350  /* accessible only through KMP_SCHEDULE environment variable */
351  kmp_sch_static_greedy = 40,
352  kmp_sch_static_balanced = 41,
353  /* accessible only through KMP_SCHEDULE environment variable */
354  kmp_sch_guided_iterative_chunked = 42,
355  kmp_sch_guided_analytical_chunked = 43,
356  /* accessible only through KMP_SCHEDULE environment variable */
357  kmp_sch_static_steal = 44,
358 
359  /* static with chunk adjustment (e.g., simd) */
360  kmp_sch_static_balanced_chunked = 45,
364  /* accessible only through KMP_SCHEDULE environment variable */
368  kmp_ord_static_chunked = 65,
370  kmp_ord_dynamic_chunked = 67,
371  kmp_ord_guided_chunked = 68,
372  kmp_ord_runtime = 69,
374  kmp_ord_trapezoidal = 71,
377  /* Schedules for Distribute construct */
381  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
382  single iteration/chunk, even if the loop is serialized. For the schedule
383  types listed above, the entire iteration vector is returned if the loop is
384  serialized. This doesn't work for gcc/gcomp sections. */
385  kmp_nm_lower = 160,
387  kmp_nm_static_chunked =
388  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
390  kmp_nm_dynamic_chunked = 163,
392  kmp_nm_runtime = 165,
393  kmp_nm_auto = 166,
394  kmp_nm_trapezoidal = 167,
395 
396  /* accessible only through KMP_SCHEDULE environment variable */
397  kmp_nm_static_greedy = 168,
398  kmp_nm_static_balanced = 169,
399  /* accessible only through KMP_SCHEDULE environment variable */
400  kmp_nm_guided_iterative_chunked = 170,
401  kmp_nm_guided_analytical_chunked = 171,
402  kmp_nm_static_steal =
403  172, /* accessible only through OMP_SCHEDULE environment variable */
404 
405  kmp_nm_ord_static_chunked = 193,
407  kmp_nm_ord_dynamic_chunked = 195,
408  kmp_nm_ord_guided_chunked = 196,
409  kmp_nm_ord_runtime = 197,
411  kmp_nm_ord_trapezoidal = 199,
414  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
415  we need to distinguish the three possible cases (no modifier, monotonic
416  modifier, nonmonotonic modifier), we need separate bits for each modifier.
417  The absence of monotonic does not imply nonmonotonic, especially since 4.5
418  says that the behaviour of the "no modifier" case is implementation defined
419  in 4.5, but will become "nonmonotonic" in 5.0.
420 
421  Since we're passing a full 32 bit value, we can use a couple of high bits
422  for these flags; out of paranoia we avoid the sign bit.
423 
424  These modifiers can be or-ed into non-static schedules by the compiler to
425  pass the additional information. They will be stripped early in the
426  processing in __kmp_dispatch_init when setting up schedules, so most of the
427  code won't ever see schedules with these bits set. */
429  (1 << 29),
431  (1 << 30),
433 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
434  (enum sched_type)( \
436 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
437 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
438 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
439  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
440 #define SCHEDULE_GET_MODIFIERS(s) \
441  ((enum sched_type)( \
442  (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
443 #define SCHEDULE_SET_MODIFIERS(s, m) \
444  (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
445 #define SCHEDULE_NONMONOTONIC 0
446 #define SCHEDULE_MONOTONIC 1
447 
449 };
450 
451 // Apply modifiers on internal kind to standard kind
452 static inline void
453 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
454  enum sched_type internal_kind) {
455  if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
456  *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
457  }
458 }
459 
460 // Apply modifiers on standard kind to internal kind
461 static inline void
462 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
463  enum sched_type *internal_kind) {
464  if ((int)kind & (int)kmp_sched_monotonic) {
465  *internal_kind = (enum sched_type)((int)*internal_kind |
467  }
468 }
469 
470 // Get standard schedule without modifiers
471 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
472  return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
473 }
474 
475 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
476 typedef union kmp_r_sched {
477  struct {
478  enum sched_type r_sched_type;
479  int chunk;
480  };
481  kmp_int64 sched;
482 } kmp_r_sched_t;
483 
484 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
485 // internal schedule types
486 
487 enum library_type {
488  library_none,
489  library_serial,
490  library_turnaround,
491  library_throughput
492 };
493 
494 #if KMP_OS_LINUX
495 enum clock_function_type {
496  clock_function_gettimeofday,
497  clock_function_clock_gettime
498 };
499 #endif /* KMP_OS_LINUX */
500 
501 #if KMP_MIC_SUPPORTED
502 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
503 #endif
504 
505 /* -- fast reduction stuff ------------------------------------------------ */
506 
507 #undef KMP_FAST_REDUCTION_BARRIER
508 #define KMP_FAST_REDUCTION_BARRIER 1
509 
510 #undef KMP_FAST_REDUCTION_CORE_DUO
511 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
512 #define KMP_FAST_REDUCTION_CORE_DUO 1
513 #endif
514 
515 enum _reduction_method {
516  reduction_method_not_defined = 0,
517  critical_reduce_block = (1 << 8),
518  atomic_reduce_block = (2 << 8),
519  tree_reduce_block = (3 << 8),
520  empty_reduce_block = (4 << 8)
521 };
522 
523 // Description of the packed_reduction_method variable:
524 // The packed_reduction_method variable consists of two enum types variables
525 // that are packed together into 0-th byte and 1-st byte:
526 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
527 // barrier that will be used in fast reduction: bs_plain_barrier or
528 // bs_reduction_barrier
529 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
530 // be used in fast reduction;
531 // Reduction method is of 'enum _reduction_method' type and it's defined the way
532 // so that the bits of 0-th byte are empty, so no need to execute a shift
533 // instruction while packing/unpacking
534 
535 #if KMP_FAST_REDUCTION_BARRIER
536 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
537  ((reduction_method) | (barrier_type))
538 
539 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
540  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
541 
542 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
543  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
544 #else
545 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
546  (reduction_method)
547 
548 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
549  (packed_reduction_method)
550 
551 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
552 #endif
553 
554 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
555  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
556  (which_reduction_block))
557 
558 #if KMP_FAST_REDUCTION_BARRIER
559 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
560  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
561 
562 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
563  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
564 #endif
565 
566 typedef int PACKED_REDUCTION_METHOD_T;
567 
568 /* -- end of fast reduction stuff ----------------------------------------- */
569 
570 #if KMP_OS_WINDOWS
571 #define USE_CBLKDATA
572 #if KMP_MSVC_COMPAT
573 #pragma warning(push)
574 #pragma warning(disable : 271 310)
575 #endif
576 #include <windows.h>
577 #if KMP_MSVC_COMPAT
578 #pragma warning(pop)
579 #endif
580 #endif
581 
582 #if KMP_OS_UNIX
583 #include <dlfcn.h>
584 #include <pthread.h>
585 #endif
586 
587 /* Only Linux* OS and Windows* OS support thread affinity. */
588 #if KMP_AFFINITY_SUPPORTED
589 
590 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
591 #if KMP_OS_WINDOWS
592 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
593 typedef struct GROUP_AFFINITY {
594  KAFFINITY Mask;
595  WORD Group;
596  WORD Reserved[3];
597 } GROUP_AFFINITY;
598 #endif /* _MSC_VER < 1600 */
599 #if KMP_GROUP_AFFINITY
600 extern int __kmp_num_proc_groups;
601 #else
602 static const int __kmp_num_proc_groups = 1;
603 #endif /* KMP_GROUP_AFFINITY */
604 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
605 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
606 
607 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
608 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
609 
610 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
611 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
612 
613 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
614  GROUP_AFFINITY *);
615 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
616 #endif /* KMP_OS_WINDOWS */
617 
618 #if KMP_USE_HWLOC
619 extern hwloc_topology_t __kmp_hwloc_topology;
620 extern int __kmp_hwloc_error;
621 extern int __kmp_numa_detected;
622 extern int __kmp_tile_depth;
623 #endif
624 
625 extern size_t __kmp_affin_mask_size;
626 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
627 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
628 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
629 #define KMP_CPU_SET_ITERATE(i, mask) \
630  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
631 #define KMP_CPU_SET(i, mask) (mask)->set(i)
632 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
633 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
634 #define KMP_CPU_ZERO(mask) (mask)->zero()
635 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
636 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
637 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
638 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
639 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
640 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
641 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
642 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
643 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
644 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
645 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
646 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
647  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
648 #define KMP_CPU_FREE_ARRAY(arr, n) \
649  __kmp_affinity_dispatch->deallocate_mask_array(arr)
650 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
651 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
652 #define __kmp_get_system_affinity(mask, abort_bool) \
653  (mask)->get_system_affinity(abort_bool)
654 #define __kmp_set_system_affinity(mask, abort_bool) \
655  (mask)->set_system_affinity(abort_bool)
656 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
657 
658 class KMPAffinity {
659 public:
660  class Mask {
661  public:
662  void *operator new(size_t n);
663  void operator delete(void *p);
664  void *operator new[](size_t n);
665  void operator delete[](void *p);
666  virtual ~Mask() {}
667  // Set bit i to 1
668  virtual void set(int i) {}
669  // Return bit i
670  virtual bool is_set(int i) const { return false; }
671  // Set bit i to 0
672  virtual void clear(int i) {}
673  // Zero out entire mask
674  virtual void zero() {}
675  // Copy src into this mask
676  virtual void copy(const Mask *src) {}
677  // this &= rhs
678  virtual void bitwise_and(const Mask *rhs) {}
679  // this |= rhs
680  virtual void bitwise_or(const Mask *rhs) {}
681  // this = ~this
682  virtual void bitwise_not() {}
683  // API for iterating over an affinity mask
684  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
685  virtual int begin() const { return 0; }
686  virtual int end() const { return 0; }
687  virtual int next(int previous) const { return 0; }
688  // Set the system's affinity to this affinity mask's value
689  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
690  // Set this affinity mask to the current system affinity
691  virtual int get_system_affinity(bool abort_on_error) { return -1; }
692  // Only 1 DWORD in the mask should have any procs set.
693  // Return the appropriate index, or -1 for an invalid mask.
694  virtual int get_proc_group() const { return -1; }
695  };
696  void *operator new(size_t n);
697  void operator delete(void *p);
698  // Need virtual destructor
699  virtual ~KMPAffinity() = default;
700  // Determine if affinity is capable
701  virtual void determine_capable(const char *env_var) {}
702  // Bind the current thread to os proc
703  virtual void bind_thread(int proc) {}
704  // Factory functions to allocate/deallocate a mask
705  virtual Mask *allocate_mask() { return nullptr; }
706  virtual void deallocate_mask(Mask *m) {}
707  virtual Mask *allocate_mask_array(int num) { return nullptr; }
708  virtual void deallocate_mask_array(Mask *m) {}
709  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
710  static void pick_api();
711  static void destroy_api();
712  enum api_type {
713  NATIVE_OS
714 #if KMP_USE_HWLOC
715  ,
716  HWLOC
717 #endif
718  };
719  virtual api_type get_api_type() const {
720  KMP_ASSERT(0);
721  return NATIVE_OS;
722  }
723 
724 private:
725  static bool picked_api;
726 };
727 
728 typedef KMPAffinity::Mask kmp_affin_mask_t;
729 extern KMPAffinity *__kmp_affinity_dispatch;
730 
731 // Declare local char buffers with this size for printing debug and info
732 // messages, using __kmp_affinity_print_mask().
733 #define KMP_AFFIN_MASK_PRINT_LEN 1024
734 
735 enum affinity_type {
736  affinity_none = 0,
737  affinity_physical,
738  affinity_logical,
739  affinity_compact,
740  affinity_scatter,
741  affinity_explicit,
742  affinity_balanced,
743  affinity_disabled, // not used outsize the env var parser
744  affinity_default
745 };
746 
747 enum affinity_gran {
748  affinity_gran_fine = 0,
749  affinity_gran_thread,
750  affinity_gran_core,
751  affinity_gran_tile,
752  affinity_gran_numa,
753  affinity_gran_package,
754  affinity_gran_node,
755 #if KMP_GROUP_AFFINITY
756  // The "group" granularity isn't necesssarily coarser than all of the
757  // other levels, but we put it last in the enum.
758  affinity_gran_group,
759 #endif /* KMP_GROUP_AFFINITY */
760  affinity_gran_default
761 };
762 
763 enum affinity_top_method {
764  affinity_top_method_all = 0, // try all (supported) methods, in order
765 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
766  affinity_top_method_apicid,
767  affinity_top_method_x2apicid,
768 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
769  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
770 #if KMP_GROUP_AFFINITY
771  affinity_top_method_group,
772 #endif /* KMP_GROUP_AFFINITY */
773  affinity_top_method_flat,
774 #if KMP_USE_HWLOC
775  affinity_top_method_hwloc,
776 #endif
777  affinity_top_method_default
778 };
779 
780 #define affinity_respect_mask_default (-1)
781 
782 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
783 extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */
784 extern int __kmp_affinity_gran_levels; /* corresponding int value */
785 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
786 extern enum affinity_top_method __kmp_affinity_top_method;
787 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
788 extern int __kmp_affinity_offset; /* Affinity offset value */
789 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
790 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
791 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
792 extern char *__kmp_affinity_proclist; /* proc ID list */
793 extern kmp_affin_mask_t *__kmp_affinity_masks;
794 extern unsigned __kmp_affinity_num_masks;
795 extern void __kmp_affinity_bind_thread(int which);
796 
797 extern kmp_affin_mask_t *__kmp_affin_fullMask;
798 extern char *__kmp_cpuinfo_file;
799 
800 #endif /* KMP_AFFINITY_SUPPORTED */
801 
802 // This needs to be kept in sync with the values in omp.h !!!
803 typedef enum kmp_proc_bind_t {
804  proc_bind_false = 0,
805  proc_bind_true,
806  proc_bind_master,
807  proc_bind_close,
808  proc_bind_spread,
809  proc_bind_intel, // use KMP_AFFINITY interface
810  proc_bind_default
811 } kmp_proc_bind_t;
812 
813 typedef struct kmp_nested_proc_bind_t {
814  kmp_proc_bind_t *bind_types;
815  int size;
816  int used;
817 } kmp_nested_proc_bind_t;
818 
819 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
820 
821 extern int __kmp_display_affinity;
822 extern char *__kmp_affinity_format;
823 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
824 
825 #if KMP_AFFINITY_SUPPORTED
826 #define KMP_PLACE_ALL (-1)
827 #define KMP_PLACE_UNDEFINED (-2)
828 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
829 #define KMP_AFFINITY_NON_PROC_BIND \
830  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
831  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
832  (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
833 #endif /* KMP_AFFINITY_SUPPORTED */
834 
835 extern int __kmp_affinity_num_places;
836 
837 typedef enum kmp_cancel_kind_t {
838  cancel_noreq = 0,
839  cancel_parallel = 1,
840  cancel_loop = 2,
841  cancel_sections = 3,
842  cancel_taskgroup = 4
843 } kmp_cancel_kind_t;
844 
845 // KMP_HW_SUBSET support:
846 typedef struct kmp_hws_item {
847  int num;
848  int offset;
849 } kmp_hws_item_t;
850 
851 extern kmp_hws_item_t __kmp_hws_socket;
852 extern kmp_hws_item_t __kmp_hws_node;
853 extern kmp_hws_item_t __kmp_hws_tile;
854 extern kmp_hws_item_t __kmp_hws_core;
855 extern kmp_hws_item_t __kmp_hws_proc;
856 extern int __kmp_hws_requested;
857 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
858 
859 /* ------------------------------------------------------------------------ */
860 
861 #define KMP_PAD(type, sz) \
862  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
863 
864 // We need to avoid using -1 as a GTID as +1 is added to the gtid
865 // when storing it in a lock, and the value 0 is reserved.
866 #define KMP_GTID_DNE (-2) /* Does not exist */
867 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
868 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
869 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
870 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
871 
872 /* OpenMP 5.0 Memory Management support */
873 
874 #ifndef __OMP_H
875 // Duplicate type definitions from omp.h
876 typedef uintptr_t omp_uintptr_t;
877 
878 typedef enum {
879  omp_atk_threadmodel = 1,
880  omp_atk_alignment = 2,
881  omp_atk_access = 3,
882  omp_atk_pool_size = 4,
883  omp_atk_fallback = 5,
884  omp_atk_fb_data = 6,
885  omp_atk_pinned = 7,
886  omp_atk_partition = 8
887 } omp_alloctrait_key_t;
888 
889 typedef enum {
890  omp_atv_false = 0,
891  omp_atv_true = 1,
892  omp_atv_default = 2,
893  omp_atv_contended = 3,
894  omp_atv_uncontended = 4,
895  omp_atv_sequential = 5,
896  omp_atv_private = 6,
897  omp_atv_all = 7,
898  omp_atv_thread = 8,
899  omp_atv_pteam = 9,
900  omp_atv_cgroup = 10,
901  omp_atv_default_mem_fb = 11,
902  omp_atv_null_fb = 12,
903  omp_atv_abort_fb = 13,
904  omp_atv_allocator_fb = 14,
905  omp_atv_environment = 15,
906  omp_atv_nearest = 16,
907  omp_atv_blocked = 17,
908  omp_atv_interleaved = 18
909 } omp_alloctrait_value_t;
910 
911 typedef void *omp_memspace_handle_t;
912 extern omp_memspace_handle_t const omp_default_mem_space;
913 extern omp_memspace_handle_t const omp_large_cap_mem_space;
914 extern omp_memspace_handle_t const omp_const_mem_space;
915 extern omp_memspace_handle_t const omp_high_bw_mem_space;
916 extern omp_memspace_handle_t const omp_low_lat_mem_space;
917 
918 typedef struct {
919  omp_alloctrait_key_t key;
920  omp_uintptr_t value;
921 } omp_alloctrait_t;
922 
923 typedef void *omp_allocator_handle_t;
924 extern omp_allocator_handle_t const omp_null_allocator;
925 extern omp_allocator_handle_t const omp_default_mem_alloc;
926 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
927 extern omp_allocator_handle_t const omp_const_mem_alloc;
928 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
929 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
930 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
931 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
932 extern omp_allocator_handle_t const omp_thread_mem_alloc;
933 extern omp_allocator_handle_t const kmp_max_mem_alloc;
934 extern omp_allocator_handle_t __kmp_def_allocator;
935 
936 // end of duplicate type definitions from omp.h
937 #endif
938 
939 extern int __kmp_memkind_available;
940 
941 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
942 
943 typedef struct kmp_allocator_t {
944  omp_memspace_handle_t memspace;
945  void **memkind; // pointer to memkind
946  int alignment;
947  omp_alloctrait_value_t fb;
948  kmp_allocator_t *fb_data;
949  kmp_uint64 pool_size;
950  kmp_uint64 pool_used;
951 } kmp_allocator_t;
952 
953 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
954  omp_memspace_handle_t,
955  int ntraits,
956  omp_alloctrait_t traits[]);
957 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
958 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
959 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
960 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
961 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
962 
963 extern void __kmp_init_memkind();
964 extern void __kmp_fini_memkind();
965 
966 /* ------------------------------------------------------------------------ */
967 
968 #define KMP_UINT64_MAX \
969  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
970 
971 #define KMP_MIN_NTH 1
972 
973 #ifndef KMP_MAX_NTH
974 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
975 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
976 #else
977 #define KMP_MAX_NTH INT_MAX
978 #endif
979 #endif /* KMP_MAX_NTH */
980 
981 #ifdef PTHREAD_STACK_MIN
982 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
983 #else
984 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
985 #endif
986 
987 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
988 
989 #if KMP_ARCH_X86
990 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
991 #elif KMP_ARCH_X86_64
992 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
993 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
994 #else
995 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
996 #endif
997 
998 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
999 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1000 #define KMP_MAX_MALLOC_POOL_INCR \
1001  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1002 
1003 #define KMP_MIN_STKOFFSET (0)
1004 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1005 #if KMP_OS_DARWIN
1006 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1007 #else
1008 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1009 #endif
1010 
1011 #define KMP_MIN_STKPADDING (0)
1012 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1013 
1014 #define KMP_BLOCKTIME_MULTIPLIER \
1015  (1000) /* number of blocktime units per second */
1016 #define KMP_MIN_BLOCKTIME (0)
1017 #define KMP_MAX_BLOCKTIME \
1018  (INT_MAX) /* Must be this for "infinite" setting the work */
1019 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
1020 
1021 #if KMP_USE_MONITOR
1022 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1023 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1024 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1025 
1026 /* Calculate new number of monitor wakeups for a specific block time based on
1027  previous monitor_wakeups. Only allow increasing number of wakeups */
1028 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1029  (((blocktime) == KMP_MAX_BLOCKTIME) \
1030  ? (monitor_wakeups) \
1031  : ((blocktime) == KMP_MIN_BLOCKTIME) \
1032  ? KMP_MAX_MONITOR_WAKEUPS \
1033  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1034  ? (monitor_wakeups) \
1035  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1036 
1037 /* Calculate number of intervals for a specific block time based on
1038  monitor_wakeups */
1039 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1040  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1041  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1042 #else
1043 #define KMP_BLOCKTIME(team, tid) \
1044  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1045 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1046 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1047 extern kmp_uint64 __kmp_ticks_per_msec;
1048 #if KMP_COMPILER_ICC
1049 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1050 #else
1051 #define KMP_NOW() __kmp_hardware_timestamp()
1052 #endif
1053 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1054 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1055  (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1056 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1057 #else
1058 // System time is retrieved sporadically while blocking.
1059 extern kmp_uint64 __kmp_now_nsec();
1060 #define KMP_NOW() __kmp_now_nsec()
1061 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1062 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1063  (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1064 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1065 #endif
1066 #endif // KMP_USE_MONITOR
1067 
1068 #define KMP_MIN_STATSCOLS 40
1069 #define KMP_MAX_STATSCOLS 4096
1070 #define KMP_DEFAULT_STATSCOLS 80
1071 
1072 #define KMP_MIN_INTERVAL 0
1073 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1074 #define KMP_DEFAULT_INTERVAL 0
1075 
1076 #define KMP_MIN_CHUNK 1
1077 #define KMP_MAX_CHUNK (INT_MAX - 1)
1078 #define KMP_DEFAULT_CHUNK 1
1079 
1080 #define KMP_DFLT_DISP_NUM_BUFF 7
1081 #define KMP_MAX_ORDERED 8
1082 
1083 #define KMP_MAX_FIELDS 32
1084 
1085 #define KMP_MAX_BRANCH_BITS 31
1086 
1087 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1088 
1089 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1090 
1091 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1092 
1093 /* Minimum number of threads before switch to TLS gtid (experimentally
1094  determined) */
1095 /* josh TODO: what about OS X* tuning? */
1096 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1097 #define KMP_TLS_GTID_MIN 5
1098 #else
1099 #define KMP_TLS_GTID_MIN INT_MAX
1100 #endif
1101 
1102 #define KMP_MASTER_TID(tid) ((tid) == 0)
1103 #define KMP_WORKER_TID(tid) ((tid) != 0)
1104 
1105 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1106 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1107 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1108 
1109 #ifndef TRUE
1110 #define FALSE 0
1111 #define TRUE (!FALSE)
1112 #endif
1113 
1114 /* NOTE: all of the following constants must be even */
1115 
1116 #if KMP_OS_WINDOWS
1117 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1118 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1119 #elif KMP_OS_LINUX
1120 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1121 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1122 #elif KMP_OS_DARWIN
1123 /* TODO: tune for KMP_OS_DARWIN */
1124 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1125 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1126 #elif KMP_OS_DRAGONFLY
1127 /* TODO: tune for KMP_OS_DRAGONFLY */
1128 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1129 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1130 #elif KMP_OS_FREEBSD
1131 /* TODO: tune for KMP_OS_FREEBSD */
1132 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1133 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1134 #elif KMP_OS_NETBSD
1135 /* TODO: tune for KMP_OS_NETBSD */
1136 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1137 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1138 #elif KMP_OS_HURD
1139 /* TODO: tune for KMP_OS_HURD */
1140 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1141 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1142 #elif KMP_OS_OPENBSD
1143 /* TODO: tune for KMP_OS_OPENBSD */
1144 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1145 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1146 #endif
1147 
1148 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1149 typedef struct kmp_cpuid {
1150  kmp_uint32 eax;
1151  kmp_uint32 ebx;
1152  kmp_uint32 ecx;
1153  kmp_uint32 edx;
1154 } kmp_cpuid_t;
1155 
1156 typedef struct kmp_cpuinfo {
1157  int initialized; // If 0, other fields are not initialized.
1158  int signature; // CPUID(1).EAX
1159  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1160  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1161  // Model << 4 ) + Model)
1162  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1163  int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1164  int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1165  int cpu_stackoffset;
1166  int apic_id;
1167  int physical_id;
1168  int logical_id;
1169  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1170  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1171 } kmp_cpuinfo_t;
1172 
1173 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1174 
1175 #if KMP_OS_UNIX
1176 // subleaf is only needed for cache and topology discovery and can be set to
1177 // zero in most cases
1178 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1179  __asm__ __volatile__("cpuid"
1180  : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1181  : "a"(leaf), "c"(subleaf));
1182 }
1183 // Load p into FPU control word
1184 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1185  __asm__ __volatile__("fldcw %0" : : "m"(*p));
1186 }
1187 // Store FPU control word into p
1188 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1189  __asm__ __volatile__("fstcw %0" : "=m"(*p));
1190 }
1191 static inline void __kmp_clear_x87_fpu_status_word() {
1192 #if KMP_MIC
1193  // 32-bit protected mode x87 FPU state
1194  struct x87_fpu_state {
1195  unsigned cw;
1196  unsigned sw;
1197  unsigned tw;
1198  unsigned fip;
1199  unsigned fips;
1200  unsigned fdp;
1201  unsigned fds;
1202  };
1203  struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1204  __asm__ __volatile__("fstenv %0\n\t" // store FP env
1205  "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1206  "fldenv %0\n\t" // load FP env back
1207  : "+m"(fpu_state), "+m"(fpu_state.sw));
1208 #else
1209  __asm__ __volatile__("fnclex");
1210 #endif // KMP_MIC
1211 }
1212 #if __SSE__
1213 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1214 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1215 #else
1216 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1217 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1218 #endif
1219 #else
1220 // Windows still has these as external functions in assembly file
1221 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1222 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1223 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1224 extern void __kmp_clear_x87_fpu_status_word();
1225 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1226 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1227 #endif // KMP_OS_UNIX
1228 
1229 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1230 
1231 #if KMP_ARCH_X86
1232 extern void __kmp_x86_pause(void);
1233 #elif KMP_MIC
1234 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1235 // regression after removal of extra PAUSE from spin loops. Changing
1236 // the delay from 100 to 300 showed even better performance than double PAUSE
1237 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1238 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1239 #else
1240 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1241 #endif
1242 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1243 #elif KMP_ARCH_PPC64
1244 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1245 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1246 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1247 #define KMP_CPU_PAUSE() \
1248  do { \
1249  KMP_PPC64_PRI_LOW(); \
1250  KMP_PPC64_PRI_MED(); \
1251  KMP_PPC64_PRI_LOC_MB(); \
1252  } while (0)
1253 #else
1254 #define KMP_CPU_PAUSE() /* nothing to do */
1255 #endif
1256 
1257 #define KMP_INIT_YIELD(count) \
1258  { (count) = __kmp_yield_init; }
1259 
1260 #define KMP_OVERSUBSCRIBED \
1261  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1262 
1263 #define KMP_TRY_YIELD \
1264  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1265 
1266 #define KMP_TRY_YIELD_OVERSUB \
1267  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1268 
1269 #define KMP_YIELD(cond) \
1270  { \
1271  KMP_CPU_PAUSE(); \
1272  if ((cond) && (KMP_TRY_YIELD)) \
1273  __kmp_yield(); \
1274  }
1275 
1276 #define KMP_YIELD_OVERSUB() \
1277  { \
1278  KMP_CPU_PAUSE(); \
1279  if ((KMP_TRY_YIELD_OVERSUB)) \
1280  __kmp_yield(); \
1281  }
1282 
1283 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1284 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1285 #define KMP_YIELD_SPIN(count) \
1286  { \
1287  KMP_CPU_PAUSE(); \
1288  if (KMP_TRY_YIELD) { \
1289  (count) -= 2; \
1290  if (!(count)) { \
1291  __kmp_yield(); \
1292  (count) = __kmp_yield_next; \
1293  } \
1294  } \
1295  }
1296 
1297 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1298  { \
1299  KMP_CPU_PAUSE(); \
1300  if ((KMP_TRY_YIELD_OVERSUB)) \
1301  __kmp_yield(); \
1302  else if (__kmp_use_yield == 1) { \
1303  (count) -= 2; \
1304  if (!(count)) { \
1305  __kmp_yield(); \
1306  (count) = __kmp_yield_next; \
1307  } \
1308  } \
1309  }
1310 
1311 /* ------------------------------------------------------------------------ */
1312 /* Support datatypes for the orphaned construct nesting checks. */
1313 /* ------------------------------------------------------------------------ */
1314 
1315 enum cons_type {
1316  ct_none,
1317  ct_parallel,
1318  ct_pdo,
1319  ct_pdo_ordered,
1320  ct_psections,
1321  ct_psingle,
1322  ct_critical,
1323  ct_ordered_in_parallel,
1324  ct_ordered_in_pdo,
1325  ct_master,
1326  ct_reduce,
1327  ct_barrier
1328 };
1329 
1330 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1331 
1332 struct cons_data {
1333  ident_t const *ident;
1334  enum cons_type type;
1335  int prev;
1336  kmp_user_lock_p
1337  name; /* address exclusively for critical section name comparison */
1338 };
1339 
1340 struct cons_header {
1341  int p_top, w_top, s_top;
1342  int stack_size, stack_top;
1343  struct cons_data *stack_data;
1344 };
1345 
1346 struct kmp_region_info {
1347  char *text;
1348  int offset[KMP_MAX_FIELDS];
1349  int length[KMP_MAX_FIELDS];
1350 };
1351 
1352 /* ---------------------------------------------------------------------- */
1353 /* ---------------------------------------------------------------------- */
1354 
1355 #if KMP_OS_WINDOWS
1356 typedef HANDLE kmp_thread_t;
1357 typedef DWORD kmp_key_t;
1358 #endif /* KMP_OS_WINDOWS */
1359 
1360 #if KMP_OS_UNIX
1361 typedef pthread_t kmp_thread_t;
1362 typedef pthread_key_t kmp_key_t;
1363 #endif
1364 
1365 extern kmp_key_t __kmp_gtid_threadprivate_key;
1366 
1367 typedef struct kmp_sys_info {
1368  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1369  long minflt; /* the number of page faults serviced without any I/O */
1370  long majflt; /* the number of page faults serviced that required I/O */
1371  long nswap; /* the number of times a process was "swapped" out of memory */
1372  long inblock; /* the number of times the file system had to perform input */
1373  long oublock; /* the number of times the file system had to perform output */
1374  long nvcsw; /* the number of times a context switch was voluntarily */
1375  long nivcsw; /* the number of times a context switch was forced */
1376 } kmp_sys_info_t;
1377 
1378 #if USE_ITT_BUILD
1379 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1380 // required type here. Later we will check the type meets requirements.
1381 typedef int kmp_itt_mark_t;
1382 #define KMP_ITT_DEBUG 0
1383 #endif /* USE_ITT_BUILD */
1384 
1385 typedef kmp_int32 kmp_critical_name[8];
1386 
1396 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1397 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1398  ...);
1399 
1404 /* ---------------------------------------------------------------------------
1405  */
1406 /* Threadprivate initialization/finalization function declarations */
1407 
1408 /* for non-array objects: __kmpc_threadprivate_register() */
1409 
1414 typedef void *(*kmpc_ctor)(void *);
1415 
1420 typedef void (*kmpc_dtor)(
1421  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1422  compiler */
1427 typedef void *(*kmpc_cctor)(void *, void *);
1428 
1429 /* for array objects: __kmpc_threadprivate_register_vec() */
1430 /* First arg: "this" pointer */
1431 /* Last arg: number of array elements */
1437 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1443 typedef void (*kmpc_dtor_vec)(void *, size_t);
1449 typedef void *(*kmpc_cctor_vec)(void *, void *,
1450  size_t); /* function unused by compiler */
1451 
1456 /* keeps tracked of threadprivate cache allocations for cleanup later */
1457 typedef struct kmp_cached_addr {
1458  void **addr; /* address of allocated cache */
1459  void ***compiler_cache; /* pointer to compiler's cache */
1460  void *data; /* pointer to global data */
1461  struct kmp_cached_addr *next; /* pointer to next cached address */
1462 } kmp_cached_addr_t;
1463 
1464 struct private_data {
1465  struct private_data *next; /* The next descriptor in the list */
1466  void *data; /* The data buffer for this descriptor */
1467  int more; /* The repeat count for this descriptor */
1468  size_t size; /* The data size for this descriptor */
1469 };
1470 
1471 struct private_common {
1472  struct private_common *next;
1473  struct private_common *link;
1474  void *gbl_addr;
1475  void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1476  size_t cmn_size;
1477 };
1478 
1479 struct shared_common {
1480  struct shared_common *next;
1481  struct private_data *pod_init;
1482  void *obj_init;
1483  void *gbl_addr;
1484  union {
1485  kmpc_ctor ctor;
1486  kmpc_ctor_vec ctorv;
1487  } ct;
1488  union {
1489  kmpc_cctor cctor;
1490  kmpc_cctor_vec cctorv;
1491  } cct;
1492  union {
1493  kmpc_dtor dtor;
1494  kmpc_dtor_vec dtorv;
1495  } dt;
1496  size_t vec_len;
1497  int is_vec;
1498  size_t cmn_size;
1499 };
1500 
1501 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1502 #define KMP_HASH_TABLE_SIZE \
1503  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1504 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1505 #define KMP_HASH(x) \
1506  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1507 
1508 struct common_table {
1509  struct private_common *data[KMP_HASH_TABLE_SIZE];
1510 };
1511 
1512 struct shared_table {
1513  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1514 };
1515 
1516 /* ------------------------------------------------------------------------ */
1517 
1518 #if KMP_USE_HIER_SCHED
1519 // Shared barrier data that exists inside a single unit of the scheduling
1520 // hierarchy
1521 typedef struct kmp_hier_private_bdata_t {
1522  kmp_int32 num_active;
1523  kmp_uint64 index;
1524  kmp_uint64 wait_val[2];
1525 } kmp_hier_private_bdata_t;
1526 #endif
1527 
1528 typedef struct kmp_sched_flags {
1529  unsigned ordered : 1;
1530  unsigned nomerge : 1;
1531  unsigned contains_last : 1;
1532 #if KMP_USE_HIER_SCHED
1533  unsigned use_hier : 1;
1534  unsigned unused : 28;
1535 #else
1536  unsigned unused : 29;
1537 #endif
1538 } kmp_sched_flags_t;
1539 
1540 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1541 
1542 #if KMP_STATIC_STEAL_ENABLED
1543 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1544  kmp_int32 count;
1545  kmp_int32 ub;
1546  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1547  kmp_int32 lb;
1548  kmp_int32 st;
1549  kmp_int32 tc;
1550  kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1551  after ub */
1552  kmp_lock_t *th_steal_lock; // lock used for chunk stealing
1553  // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1554  // a) parm3 is properly aligned and
1555  // b) all parm1-4 are in the same cache line.
1556  // Because of parm1-4 are used together, performance seems to be better
1557  // if they are in the same line (not measured though).
1558 
1559  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1560  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1561  kmp_int32 parm2; // make no real change at least while padding is off.
1562  kmp_int32 parm3;
1563  kmp_int32 parm4;
1564  };
1565 
1566  kmp_uint32 ordered_lower;
1567  kmp_uint32 ordered_upper;
1568 #if KMP_OS_WINDOWS
1569  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1570  // 'static_steal_counter'. It would be nice to measure execution times.
1571  // Conditional if/endif can be removed at all.
1572  kmp_int32 last_upper;
1573 #endif /* KMP_OS_WINDOWS */
1574 } dispatch_private_info32_t;
1575 
1576 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1577  kmp_int64 count; // current chunk number for static & static-steal scheduling
1578  kmp_int64 ub; /* upper-bound */
1579  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1580  kmp_int64 lb; /* lower-bound */
1581  kmp_int64 st; /* stride */
1582  kmp_int64 tc; /* trip count (number of iterations) */
1583  kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1584  after ub */
1585  kmp_lock_t *th_steal_lock; // lock used for chunk stealing
1586  /* parm[1-4] are used in different ways by different scheduling algorithms */
1587 
1588  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1589  // a) parm3 is properly aligned and
1590  // b) all parm1-4 are in the same cache line.
1591  // Because of parm1-4 are used together, performance seems to be better
1592  // if they are in the same line (not measured though).
1593 
1594  struct KMP_ALIGN(32) {
1595  kmp_int64 parm1;
1596  kmp_int64 parm2;
1597  kmp_int64 parm3;
1598  kmp_int64 parm4;
1599  };
1600 
1601  kmp_uint64 ordered_lower;
1602  kmp_uint64 ordered_upper;
1603 #if KMP_OS_WINDOWS
1604  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1605  // 'static_steal_counter'. It would be nice to measure execution times.
1606  // Conditional if/endif can be removed at all.
1607  kmp_int64 last_upper;
1608 #endif /* KMP_OS_WINDOWS */
1609 } dispatch_private_info64_t;
1610 #else /* KMP_STATIC_STEAL_ENABLED */
1611 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1612  kmp_int32 lb;
1613  kmp_int32 ub;
1614  kmp_int32 st;
1615  kmp_int32 tc;
1616 
1617  kmp_int32 parm1;
1618  kmp_int32 parm2;
1619  kmp_int32 parm3;
1620  kmp_int32 parm4;
1621 
1622  kmp_int32 count;
1623 
1624  kmp_uint32 ordered_lower;
1625  kmp_uint32 ordered_upper;
1626 #if KMP_OS_WINDOWS
1627  kmp_int32 last_upper;
1628 #endif /* KMP_OS_WINDOWS */
1629 } dispatch_private_info32_t;
1630 
1631 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1632  kmp_int64 lb; /* lower-bound */
1633  kmp_int64 ub; /* upper-bound */
1634  kmp_int64 st; /* stride */
1635  kmp_int64 tc; /* trip count (number of iterations) */
1636 
1637  /* parm[1-4] are used in different ways by different scheduling algorithms */
1638  kmp_int64 parm1;
1639  kmp_int64 parm2;
1640  kmp_int64 parm3;
1641  kmp_int64 parm4;
1642 
1643  kmp_int64 count; /* current chunk number for static scheduling */
1644 
1645  kmp_uint64 ordered_lower;
1646  kmp_uint64 ordered_upper;
1647 #if KMP_OS_WINDOWS
1648  kmp_int64 last_upper;
1649 #endif /* KMP_OS_WINDOWS */
1650 } dispatch_private_info64_t;
1651 #endif /* KMP_STATIC_STEAL_ENABLED */
1652 
1653 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1654  union private_info {
1655  dispatch_private_info32_t p32;
1656  dispatch_private_info64_t p64;
1657  } u;
1658  enum sched_type schedule; /* scheduling algorithm */
1659  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1660  kmp_int32 ordered_bumped;
1661  // To retain the structure size after making ordered_iteration scalar
1662  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1663  // Stack of buffers for nest of serial regions
1664  struct dispatch_private_info *next;
1665  kmp_int32 type_size; /* the size of types in private_info */
1666 #if KMP_USE_HIER_SCHED
1667  kmp_int32 hier_id;
1668  void *parent; /* hierarchical scheduling parent pointer */
1669 #endif
1670  enum cons_type pushed_ws;
1671 } dispatch_private_info_t;
1672 
1673 typedef struct dispatch_shared_info32 {
1674  /* chunk index under dynamic, number of idle threads under static-steal;
1675  iteration index otherwise */
1676  volatile kmp_uint32 iteration;
1677  volatile kmp_uint32 num_done;
1678  volatile kmp_uint32 ordered_iteration;
1679  // Dummy to retain the structure size after making ordered_iteration scalar
1680  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1681 } dispatch_shared_info32_t;
1682 
1683 typedef struct dispatch_shared_info64 {
1684  /* chunk index under dynamic, number of idle threads under static-steal;
1685  iteration index otherwise */
1686  volatile kmp_uint64 iteration;
1687  volatile kmp_uint64 num_done;
1688  volatile kmp_uint64 ordered_iteration;
1689  // Dummy to retain the structure size after making ordered_iteration scalar
1690  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1691 } dispatch_shared_info64_t;
1692 
1693 typedef struct dispatch_shared_info {
1694  union shared_info {
1695  dispatch_shared_info32_t s32;
1696  dispatch_shared_info64_t s64;
1697  } u;
1698  volatile kmp_uint32 buffer_index;
1699  volatile kmp_int32 doacross_buf_idx; // teamwise index
1700  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1701  kmp_int32 doacross_num_done; // count finished threads
1702 #if KMP_USE_HIER_SCHED
1703  void *hier;
1704 #endif
1705 #if KMP_USE_HWLOC
1706  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1707  // machines (> 48 cores). Performance analysis showed that a cache thrash
1708  // was occurring and this padding helps alleviate the problem.
1709  char padding[64];
1710 #endif
1711 } dispatch_shared_info_t;
1712 
1713 typedef struct kmp_disp {
1714  /* Vector for ORDERED SECTION */
1715  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1716  /* Vector for END ORDERED SECTION */
1717  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1718 
1719  dispatch_shared_info_t *th_dispatch_sh_current;
1720  dispatch_private_info_t *th_dispatch_pr_current;
1721 
1722  dispatch_private_info_t *th_disp_buffer;
1723  kmp_int32 th_disp_index;
1724  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1725  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1726  kmp_int64 *th_doacross_info; // info on loop bounds
1727 #if KMP_USE_INTERNODE_ALIGNMENT
1728  char more_padding[INTERNODE_CACHE_LINE];
1729 #endif
1730 } kmp_disp_t;
1731 
1732 /* ------------------------------------------------------------------------ */
1733 /* Barrier stuff */
1734 
1735 /* constants for barrier state update */
1736 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1737 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1738 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1739 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1740 
1741 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1742 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1743 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1744 
1745 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1746 #error "Barrier sleep bit must be smaller than barrier bump bit"
1747 #endif
1748 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1749 #error "Barrier unused bit must be smaller than barrier bump bit"
1750 #endif
1751 
1752 // Constants for release barrier wait state: currently, hierarchical only
1753 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1754 #define KMP_BARRIER_OWN_FLAG \
1755  1 // Normal state; worker waiting on own b_go flag in release
1756 #define KMP_BARRIER_PARENT_FLAG \
1757  2 // Special state; worker waiting on parent's b_go flag in release
1758 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1759  3 // Special state; tells worker to shift from parent to own b_go
1760 #define KMP_BARRIER_SWITCHING \
1761  4 // Special state; worker resets appropriate flag on wake-up
1762 
1763 #define KMP_NOT_SAFE_TO_REAP \
1764  0 // Thread th_reap_state: not safe to reap (tasking)
1765 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1766 
1767 enum barrier_type {
1768  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1769  barriers if enabled) */
1770  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1771 #if KMP_FAST_REDUCTION_BARRIER
1772  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1773 #endif // KMP_FAST_REDUCTION_BARRIER
1774  bs_last_barrier /* Just a placeholder to mark the end */
1775 };
1776 
1777 // to work with reduction barriers just like with plain barriers
1778 #if !KMP_FAST_REDUCTION_BARRIER
1779 #define bs_reduction_barrier bs_plain_barrier
1780 #endif // KMP_FAST_REDUCTION_BARRIER
1781 
1782 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1783  bp_linear_bar =
1784  0, /* Single level (degenerate) tree */
1785  bp_tree_bar =
1786  1, /* Balanced tree with branching factor 2^n */
1787  bp_hyper_bar =
1788  2, /* Hypercube-embedded tree with min branching
1789  factor 2^n */
1790  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1791  bp_last_bar /* Placeholder to mark the end */
1792 } kmp_bar_pat_e;
1793 
1794 #define KMP_BARRIER_ICV_PUSH 1
1795 
1796 /* Record for holding the values of the internal controls stack records */
1797 typedef struct kmp_internal_control {
1798  int serial_nesting_level; /* corresponds to the value of the
1799  th_team_serialized field */
1800  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1801  thread) */
1802  kmp_int8
1803  bt_set; /* internal control for whether blocktime is explicitly set */
1804  int blocktime; /* internal control for blocktime */
1805 #if KMP_USE_MONITOR
1806  int bt_intervals; /* internal control for blocktime intervals */
1807 #endif
1808  int nproc; /* internal control for #threads for next parallel region (per
1809  thread) */
1810  int thread_limit; /* internal control for thread-limit-var */
1811  int max_active_levels; /* internal control for max_active_levels */
1812  kmp_r_sched_t
1813  sched; /* internal control for runtime schedule {sched,chunk} pair */
1814  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1815  kmp_int32 default_device; /* internal control for default device */
1816  struct kmp_internal_control *next;
1817 } kmp_internal_control_t;
1818 
1819 static inline void copy_icvs(kmp_internal_control_t *dst,
1820  kmp_internal_control_t *src) {
1821  *dst = *src;
1822 }
1823 
1824 /* Thread barrier needs volatile barrier fields */
1825 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1826  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1827  // uses of it). It is not explicitly aligned below, because we *don't* want
1828  // it to be padded -- instead, we fit b_go into the same cache line with
1829  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1830  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1831  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1832  // same NGO store
1833  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1834  KMP_ALIGN_CACHE volatile kmp_uint64
1835  b_arrived; // STATE => task reached synch point.
1836  kmp_uint32 *skip_per_level;
1837  kmp_uint32 my_level;
1838  kmp_int32 parent_tid;
1839  kmp_int32 old_tid;
1840  kmp_uint32 depth;
1841  struct kmp_bstate *parent_bar;
1842  kmp_team_t *team;
1843  kmp_uint64 leaf_state;
1844  kmp_uint32 nproc;
1845  kmp_uint8 base_leaf_kids;
1846  kmp_uint8 leaf_kids;
1847  kmp_uint8 offset;
1848  kmp_uint8 wait_flag;
1849  kmp_uint8 use_oncore_barrier;
1850 #if USE_DEBUGGER
1851  // The following field is intended for the debugger solely. Only the worker
1852  // thread itself accesses this field: the worker increases it by 1 when it
1853  // arrives to a barrier.
1854  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1855 #endif /* USE_DEBUGGER */
1856 } kmp_bstate_t;
1857 
1858 union KMP_ALIGN_CACHE kmp_barrier_union {
1859  double b_align; /* use worst case alignment */
1860  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1861  kmp_bstate_t bb;
1862 };
1863 
1864 typedef union kmp_barrier_union kmp_balign_t;
1865 
1866 /* Team barrier needs only non-volatile arrived counter */
1867 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1868  double b_align; /* use worst case alignment */
1869  char b_pad[CACHE_LINE];
1870  struct {
1871  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1872 #if USE_DEBUGGER
1873  // The following two fields are indended for the debugger solely. Only
1874  // master of the team accesses these fields: the first one is increased by
1875  // 1 when master arrives to a barrier, the second one is increased by one
1876  // when all the threads arrived.
1877  kmp_uint b_master_arrived;
1878  kmp_uint b_team_arrived;
1879 #endif
1880  };
1881 };
1882 
1883 typedef union kmp_barrier_team_union kmp_balign_team_t;
1884 
1885 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
1886  threads when a condition changes. This is to workaround an NPTL bug where
1887  padding was added to pthread_cond_t which caused the initialization routine
1888  to write outside of the structure if compiled on pre-NPTL threads. */
1889 #if KMP_OS_WINDOWS
1890 typedef struct kmp_win32_mutex {
1891  /* The Lock */
1892  CRITICAL_SECTION cs;
1893 } kmp_win32_mutex_t;
1894 
1895 typedef struct kmp_win32_cond {
1896  /* Count of the number of waiters. */
1897  int waiters_count_;
1898 
1899  /* Serialize access to <waiters_count_> */
1900  kmp_win32_mutex_t waiters_count_lock_;
1901 
1902  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
1903  int release_count_;
1904 
1905  /* Keeps track of the current "generation" so that we don't allow */
1906  /* one thread to steal all the "releases" from the broadcast. */
1907  int wait_generation_count_;
1908 
1909  /* A manual-reset event that's used to block and release waiting threads. */
1910  HANDLE event_;
1911 } kmp_win32_cond_t;
1912 #endif
1913 
1914 #if KMP_OS_UNIX
1915 
1916 union KMP_ALIGN_CACHE kmp_cond_union {
1917  double c_align;
1918  char c_pad[CACHE_LINE];
1919  pthread_cond_t c_cond;
1920 };
1921 
1922 typedef union kmp_cond_union kmp_cond_align_t;
1923 
1924 union KMP_ALIGN_CACHE kmp_mutex_union {
1925  double m_align;
1926  char m_pad[CACHE_LINE];
1927  pthread_mutex_t m_mutex;
1928 };
1929 
1930 typedef union kmp_mutex_union kmp_mutex_align_t;
1931 
1932 #endif /* KMP_OS_UNIX */
1933 
1934 typedef struct kmp_desc_base {
1935  void *ds_stackbase;
1936  size_t ds_stacksize;
1937  int ds_stackgrow;
1938  kmp_thread_t ds_thread;
1939  volatile int ds_tid;
1940  int ds_gtid;
1941 #if KMP_OS_WINDOWS
1942  volatile int ds_alive;
1943  DWORD ds_thread_id;
1944 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
1945  However, debugger support (libomp_db) cannot work with handles, because they
1946  uncomparable. For example, debugger requests info about thread with handle h.
1947  h is valid within debugger process, and meaningless within debugee process.
1948  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
1949  within debugee process, but it is a *new* handle which does *not* equal to
1950  any other handle in debugee... The only way to compare handles is convert
1951  them to system-wide ids. GetThreadId() function is available only in
1952  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
1953  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
1954  thread id by call to GetCurrentThreadId() from within the thread and save it
1955  to let libomp_db identify threads. */
1956 #endif /* KMP_OS_WINDOWS */
1957 } kmp_desc_base_t;
1958 
1959 typedef union KMP_ALIGN_CACHE kmp_desc {
1960  double ds_align; /* use worst case alignment */
1961  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1962  kmp_desc_base_t ds;
1963 } kmp_desc_t;
1964 
1965 typedef struct kmp_local {
1966  volatile int this_construct; /* count of single's encountered by thread */
1967  void *reduce_data;
1968 #if KMP_USE_BGET
1969  void *bget_data;
1970  void *bget_list;
1971 #if !USE_CMP_XCHG_FOR_BGET
1972 #ifdef USE_QUEUING_LOCK_FOR_BGET
1973  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
1974 #else
1975  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
1976 // bootstrap lock so we can use it at library
1977 // shutdown.
1978 #endif /* USE_LOCK_FOR_BGET */
1979 #endif /* ! USE_CMP_XCHG_FOR_BGET */
1980 #endif /* KMP_USE_BGET */
1981 
1982  PACKED_REDUCTION_METHOD_T
1983  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
1984  __kmpc_end_reduce*() */
1985 
1986 } kmp_local_t;
1987 
1988 #define KMP_CHECK_UPDATE(a, b) \
1989  if ((a) != (b)) \
1990  (a) = (b)
1991 #define KMP_CHECK_UPDATE_SYNC(a, b) \
1992  if ((a) != (b)) \
1993  TCW_SYNC_PTR((a), (b))
1994 
1995 #define get__blocktime(xteam, xtid) \
1996  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
1997 #define get__bt_set(xteam, xtid) \
1998  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
1999 #if KMP_USE_MONITOR
2000 #define get__bt_intervals(xteam, xtid) \
2001  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2002 #endif
2003 
2004 #define get__dynamic_2(xteam, xtid) \
2005  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2006 #define get__nproc_2(xteam, xtid) \
2007  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2008 #define get__sched_2(xteam, xtid) \
2009  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2010 
2011 #define set__blocktime_team(xteam, xtid, xval) \
2012  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2013  (xval))
2014 
2015 #if KMP_USE_MONITOR
2016 #define set__bt_intervals_team(xteam, xtid, xval) \
2017  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2018  (xval))
2019 #endif
2020 
2021 #define set__bt_set_team(xteam, xtid, xval) \
2022  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2023 
2024 #define set__dynamic(xthread, xval) \
2025  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2026 #define get__dynamic(xthread) \
2027  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2028 
2029 #define set__nproc(xthread, xval) \
2030  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2031 
2032 #define set__thread_limit(xthread, xval) \
2033  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2034 
2035 #define set__max_active_levels(xthread, xval) \
2036  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2037 
2038 #define get__max_active_levels(xthread) \
2039  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2040 
2041 #define set__sched(xthread, xval) \
2042  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2043 
2044 #define set__proc_bind(xthread, xval) \
2045  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2046 #define get__proc_bind(xthread) \
2047  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2048 
2049 // OpenMP tasking data structures
2050 
2051 typedef enum kmp_tasking_mode {
2052  tskm_immediate_exec = 0,
2053  tskm_extra_barrier = 1,
2054  tskm_task_teams = 2,
2055  tskm_max = 2
2056 } kmp_tasking_mode_t;
2057 
2058 extern kmp_tasking_mode_t
2059  __kmp_tasking_mode; /* determines how/when to execute tasks */
2060 extern int __kmp_task_stealing_constraint;
2061 extern int __kmp_enable_task_throttling;
2062 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2063 // specified, defaults to 0 otherwise
2064 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2065 extern kmp_int32 __kmp_max_task_priority;
2066 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2067 extern kmp_uint64 __kmp_taskloop_min_tasks;
2068 
2069 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2070  taskdata first */
2071 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2072 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2073 
2074 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2075 // were spawned and queued since the previous barrier release.
2076 #define KMP_TASKING_ENABLED(task_team) \
2077  (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2078 
2085 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2086 
2087 typedef union kmp_cmplrdata {
2088  kmp_int32 priority;
2089  kmp_routine_entry_t
2090  destructors; /* pointer to function to invoke deconstructors of
2091  firstprivate C++ objects */
2092  /* future data */
2093 } kmp_cmplrdata_t;
2094 
2095 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2098 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2099  void *shareds;
2100  kmp_routine_entry_t
2101  routine;
2102  kmp_int32 part_id;
2103  kmp_cmplrdata_t
2104  data1; /* Two known optional additions: destructors and priority */
2105  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2106  /* future data */
2107  /* private vars */
2108 } kmp_task_t;
2109 
2114 typedef struct kmp_taskgroup {
2115  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2116  std::atomic<kmp_int32>
2117  cancel_request; // request for cancellation of this taskgroup
2118  struct kmp_taskgroup *parent; // parent taskgroup
2119  // Block of data to perform task reduction
2120  void *reduce_data; // reduction related info
2121  kmp_int32 reduce_num_data; // number of data items to reduce
2122 } kmp_taskgroup_t;
2123 
2124 // forward declarations
2125 typedef union kmp_depnode kmp_depnode_t;
2126 typedef struct kmp_depnode_list kmp_depnode_list_t;
2127 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2128 
2129 // Compiler sends us this info:
2130 typedef struct kmp_depend_info {
2131  kmp_intptr_t base_addr;
2132  size_t len;
2133  struct {
2134  bool in : 1;
2135  bool out : 1;
2136  bool mtx : 1;
2137  } flags;
2138 } kmp_depend_info_t;
2139 
2140 // Internal structures to work with task dependencies:
2141 struct kmp_depnode_list {
2142  kmp_depnode_t *node;
2143  kmp_depnode_list_t *next;
2144 };
2145 
2146 // Max number of mutexinoutset dependencies per node
2147 #define MAX_MTX_DEPS 4
2148 
2149 typedef struct kmp_base_depnode {
2150  kmp_depnode_list_t *successors; /* used under lock */
2151  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2152  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2153  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2154  kmp_lock_t lock; /* guards shared fields: task, successors */
2155 #if KMP_SUPPORT_GRAPH_OUTPUT
2156  kmp_uint32 id;
2157 #endif
2158  std::atomic<kmp_int32> npredecessors;
2159  std::atomic<kmp_int32> nrefs;
2160 } kmp_base_depnode_t;
2161 
2162 union KMP_ALIGN_CACHE kmp_depnode {
2163  double dn_align; /* use worst case alignment */
2164  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2165  kmp_base_depnode_t dn;
2166 };
2167 
2168 struct kmp_dephash_entry {
2169  kmp_intptr_t addr;
2170  kmp_depnode_t *last_out;
2171  kmp_depnode_list_t *last_ins;
2172  kmp_depnode_list_t *last_mtxs;
2173  kmp_int32 last_flag;
2174  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2175  kmp_dephash_entry_t *next_in_bucket;
2176 };
2177 
2178 typedef struct kmp_dephash {
2179  kmp_dephash_entry_t **buckets;
2180  size_t size;
2181  size_t generation;
2182  kmp_uint32 nelements;
2183  kmp_uint32 nconflicts;
2184 } kmp_dephash_t;
2185 
2186 typedef struct kmp_task_affinity_info {
2187  kmp_intptr_t base_addr;
2188  size_t len;
2189  struct {
2190  bool flag1 : 1;
2191  bool flag2 : 1;
2192  kmp_int32 reserved : 30;
2193  } flags;
2194 } kmp_task_affinity_info_t;
2195 
2196 typedef enum kmp_event_type_t {
2197  KMP_EVENT_UNINITIALIZED = 0,
2198  KMP_EVENT_ALLOW_COMPLETION = 1
2199 } kmp_event_type_t;
2200 
2201 typedef struct {
2202  kmp_event_type_t type;
2203  kmp_tas_lock_t lock;
2204  union {
2205  kmp_task_t *task;
2206  } ed;
2207 } kmp_event_t;
2208 
2209 #ifdef BUILD_TIED_TASK_STACK
2210 
2211 /* Tied Task stack definitions */
2212 typedef struct kmp_stack_block {
2213  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2214  struct kmp_stack_block *sb_next;
2215  struct kmp_stack_block *sb_prev;
2216 } kmp_stack_block_t;
2217 
2218 typedef struct kmp_task_stack {
2219  kmp_stack_block_t ts_first_block; // first block of stack entries
2220  kmp_taskdata_t **ts_top; // pointer to the top of stack
2221  kmp_int32 ts_entries; // number of entries on the stack
2222 } kmp_task_stack_t;
2223 
2224 #endif // BUILD_TIED_TASK_STACK
2225 
2226 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2227  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2228  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2229  unsigned final : 1; /* task is final(1) so execute immediately */
2230  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2231  code path */
2232  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2233  invoke destructors from the runtime */
2234  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2235  context of the RTL) */
2236  unsigned priority_specified : 1; /* set if the compiler provides priority
2237  setting for the task */
2238  unsigned detachable : 1; /* 1 == can detach */
2239  unsigned reserved : 9; /* reserved for compiler use */
2240 
2241  /* Library flags */ /* Total library flags must be 16 bits */
2242  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2243  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2244  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2245  // (1) or may be deferred (0)
2246  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2247  // (0) [>= 2 threads]
2248  /* If either team_serial or tasking_ser is set, task team may be NULL */
2249  /* Task State Flags: */
2250  unsigned started : 1; /* 1==started, 0==not started */
2251  unsigned executing : 1; /* 1==executing, 0==not executing */
2252  unsigned complete : 1; /* 1==complete, 0==not complete */
2253  unsigned freed : 1; /* 1==freed, 0==allocated */
2254  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2255  unsigned reserved31 : 7; /* reserved for library use */
2256 
2257 } kmp_tasking_flags_t;
2258 
2259 struct kmp_taskdata { /* aligned during dynamic allocation */
2260  kmp_int32 td_task_id; /* id, assigned by debugger */
2261  kmp_tasking_flags_t td_flags; /* task flags */
2262  kmp_team_t *td_team; /* team for this task */
2263  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2264  /* Currently not used except for perhaps IDB */
2265  kmp_taskdata_t *td_parent; /* parent task */
2266  kmp_int32 td_level; /* task nesting level */
2267  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2268  ident_t *td_ident; /* task identifier */
2269  // Taskwait data.
2270  ident_t *td_taskwait_ident;
2271  kmp_uint32 td_taskwait_counter;
2272  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2273  KMP_ALIGN_CACHE kmp_internal_control_t
2274  td_icvs; /* Internal control variables for the task */
2275  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2276  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2277  deallocated */
2278  std::atomic<kmp_int32>
2279  td_incomplete_child_tasks; /* Child tasks not yet complete */
2280  kmp_taskgroup_t
2281  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2282  kmp_dephash_t
2283  *td_dephash; // Dependencies for children tasks are tracked from here
2284  kmp_depnode_t
2285  *td_depnode; // Pointer to graph node if this task has dependencies
2286  kmp_task_team_t *td_task_team;
2287  kmp_int32 td_size_alloc; // The size of task structure, including shareds etc.
2288 #if defined(KMP_GOMP_COMPAT)
2289  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2290  kmp_int32 td_size_loop_bounds;
2291 #endif
2292  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2293 #if defined(KMP_GOMP_COMPAT)
2294  // GOMP sends in a copy function for copy constructors
2295  void (*td_copy_func)(void *, void *);
2296 #endif
2297  kmp_event_t td_allow_completion_event;
2298 #if OMPT_SUPPORT
2299  ompt_task_info_t ompt_task_info;
2300 #endif
2301 }; // struct kmp_taskdata
2302 
2303 // Make sure padding above worked
2304 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2305 
2306 // Data for task team but per thread
2307 typedef struct kmp_base_thread_data {
2308  kmp_info_p *td_thr; // Pointer back to thread info
2309  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2310  // queued?
2311  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2312  kmp_taskdata_t *
2313  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2314  kmp_int32 td_deque_size; // Size of deck
2315  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2316  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2317  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2318  // GEH: shouldn't this be volatile since used in while-spin?
2319  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2320 #ifdef BUILD_TIED_TASK_STACK
2321  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2322 // scheduling constraint
2323 #endif // BUILD_TIED_TASK_STACK
2324 } kmp_base_thread_data_t;
2325 
2326 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2327 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2328 
2329 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2330 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2331 
2332 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2333  kmp_base_thread_data_t td;
2334  double td_align; /* use worst case alignment */
2335  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2336 } kmp_thread_data_t;
2337 
2338 // Data for task teams which are used when tasking is enabled for the team
2339 typedef struct kmp_base_task_team {
2340  kmp_bootstrap_lock_t
2341  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2342  /* must be bootstrap lock since used at library shutdown*/
2343  kmp_task_team_t *tt_next; /* For linking the task team free list */
2344  kmp_thread_data_t
2345  *tt_threads_data; /* Array of per-thread structures for task team */
2346  /* Data survives task team deallocation */
2347  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2348  executing this team? */
2349  /* TRUE means tt_threads_data is set up and initialized */
2350  kmp_int32 tt_nproc; /* #threads in team */
2351  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2352  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2353  kmp_int32 tt_untied_task_encountered;
2354 
2355  KMP_ALIGN_CACHE
2356  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2357 
2358  KMP_ALIGN_CACHE
2359  volatile kmp_uint32
2360  tt_active; /* is the team still actively executing tasks */
2361 } kmp_base_task_team_t;
2362 
2363 union KMP_ALIGN_CACHE kmp_task_team {
2364  kmp_base_task_team_t tt;
2365  double tt_align; /* use worst case alignment */
2366  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2367 };
2368 
2369 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2370 // Free lists keep same-size free memory slots for fast memory allocation
2371 // routines
2372 typedef struct kmp_free_list {
2373  void *th_free_list_self; // Self-allocated tasks free list
2374  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2375  // threads
2376  void *th_free_list_other; // Non-self free list (to be returned to owner's
2377  // sync list)
2378 } kmp_free_list_t;
2379 #endif
2380 #if KMP_NESTED_HOT_TEAMS
2381 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2382 // are not put in teams pool, and they don't put threads in threads pool.
2383 typedef struct kmp_hot_team_ptr {
2384  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2385  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2386 } kmp_hot_team_ptr_t;
2387 #endif
2388 typedef struct kmp_teams_size {
2389  kmp_int32 nteams; // number of teams in a league
2390  kmp_int32 nth; // number of threads in each team of the league
2391 } kmp_teams_size_t;
2392 
2393 // This struct stores a thread that acts as a "root" for a contention
2394 // group. Contention groups are rooted at kmp_root threads, but also at
2395 // each master thread of each team created in the teams construct.
2396 // This struct therefore also stores a thread_limit associated with
2397 // that contention group, and a counter to track the number of threads
2398 // active in that contention group. Each thread has a list of these: CG
2399 // root threads have an entry in their list in which cg_root refers to
2400 // the thread itself, whereas other workers in the CG will have a
2401 // single entry where cg_root is same as the entry containing their CG
2402 // root. When a thread encounters a teams construct, it will add a new
2403 // entry to the front of its list, because it now roots a new CG.
2404 typedef struct kmp_cg_root {
2405  kmp_info_p *cg_root; // "root" thread for a contention group
2406  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2407  // thread_limit clause for teams masters
2408  kmp_int32 cg_thread_limit;
2409  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2410  struct kmp_cg_root *up; // pointer to higher level CG root in list
2411 } kmp_cg_root_t;
2412 
2413 // OpenMP thread data structures
2414 
2415 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2416  /* Start with the readonly data which is cache aligned and padded. This is
2417  written before the thread starts working by the master. Uber masters may
2418  update themselves later. Usage does not consider serialized regions. */
2419  kmp_desc_t th_info;
2420  kmp_team_p *th_team; /* team we belong to */
2421  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2422  kmp_info_p *th_next_pool; /* next available thread in the pool */
2423  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2424  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2425 
2426  /* The following are cached from the team info structure */
2427  /* TODO use these in more places as determined to be needed via profiling */
2428  int th_team_nproc; /* number of threads in a team */
2429  kmp_info_p *th_team_master; /* the team's master thread */
2430  int th_team_serialized; /* team is serialized */
2431  microtask_t th_teams_microtask; /* save entry address for teams construct */
2432  int th_teams_level; /* save initial level of teams construct */
2433 /* it is 0 on device but may be any on host */
2434 
2435 /* The blocktime info is copied from the team struct to the thread struct */
2436 /* at the start of a barrier, and the values stored in the team are used */
2437 /* at points in the code where the team struct is no longer guaranteed */
2438 /* to exist (from the POV of worker threads). */
2439 #if KMP_USE_MONITOR
2440  int th_team_bt_intervals;
2441  int th_team_bt_set;
2442 #else
2443  kmp_uint64 th_team_bt_intervals;
2444 #endif
2445 
2446 #if KMP_AFFINITY_SUPPORTED
2447  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2448 #endif
2449  omp_allocator_handle_t th_def_allocator; /* default allocator */
2450  /* The data set by the master at reinit, then R/W by the worker */
2451  KMP_ALIGN_CACHE int
2452  th_set_nproc; /* if > 0, then only use this request for the next fork */
2453 #if KMP_NESTED_HOT_TEAMS
2454  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2455 #endif
2456  kmp_proc_bind_t
2457  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2458  kmp_teams_size_t
2459  th_teams_size; /* number of teams/threads in teams construct */
2460 #if KMP_AFFINITY_SUPPORTED
2461  int th_current_place; /* place currently bound to */
2462  int th_new_place; /* place to bind to in par reg */
2463  int th_first_place; /* first place in partition */
2464  int th_last_place; /* last place in partition */
2465 #endif
2466  int th_prev_level; /* previous level for affinity format */
2467  int th_prev_num_threads; /* previous num_threads for affinity format */
2468 #if USE_ITT_BUILD
2469  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2470  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2471  kmp_uint64 th_frame_time; /* frame timestamp */
2472 #endif /* USE_ITT_BUILD */
2473  kmp_local_t th_local;
2474  struct private_common *th_pri_head;
2475 
2476  /* Now the data only used by the worker (after initial allocation) */
2477  /* TODO the first serial team should actually be stored in the info_t
2478  structure. this will help reduce initial allocation overhead */
2479  KMP_ALIGN_CACHE kmp_team_p
2480  *th_serial_team; /*serialized team held in reserve*/
2481 
2482 #if OMPT_SUPPORT
2483  ompt_thread_info_t ompt_thread_info;
2484 #endif
2485 
2486  /* The following are also read by the master during reinit */
2487  struct common_table *th_pri_common;
2488 
2489  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2490  /* while awaiting queuing lock acquire */
2491 
2492  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2493 
2494  ident_t *th_ident;
2495  unsigned th_x; // Random number generator data
2496  unsigned th_a; // Random number generator data
2497 
2498  /* Tasking-related data for the thread */
2499  kmp_task_team_t *th_task_team; // Task team struct
2500  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2501  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2502  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2503  // at nested levels
2504  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2505  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2506  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2507  // tasking, thus safe to reap
2508 
2509  /* More stuff for keeping track of active/sleeping threads (this part is
2510  written by the worker thread) */
2511  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2512  int th_active; // ! sleeping; 32 bits for TCR/TCW
2513  struct cons_header *th_cons; // used for consistency check
2514 #if KMP_USE_HIER_SCHED
2515  // used for hierarchical scheduling
2516  kmp_hier_private_bdata_t *th_hier_bar_data;
2517 #endif
2518 
2519  /* Add the syncronizing data which is cache aligned and padded. */
2520  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2521 
2522  KMP_ALIGN_CACHE volatile kmp_int32
2523  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2524 
2525 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2526 #define NUM_LISTS 4
2527  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2528 // allocation routines
2529 #endif
2530 
2531 #if KMP_OS_WINDOWS
2532  kmp_win32_cond_t th_suspend_cv;
2533  kmp_win32_mutex_t th_suspend_mx;
2534  std::atomic<int> th_suspend_init;
2535 #endif
2536 #if KMP_OS_UNIX
2537  kmp_cond_align_t th_suspend_cv;
2538  kmp_mutex_align_t th_suspend_mx;
2539  std::atomic<int> th_suspend_init_count;
2540 #endif
2541 
2542 #if USE_ITT_BUILD
2543  kmp_itt_mark_t th_itt_mark_single;
2544 // alignment ???
2545 #endif /* USE_ITT_BUILD */
2546 #if KMP_STATS_ENABLED
2547  kmp_stats_list *th_stats;
2548 #endif
2549 #if KMP_OS_UNIX
2550  std::atomic<bool> th_blocking;
2551 #endif
2552  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2553 } kmp_base_info_t;
2554 
2555 typedef union KMP_ALIGN_CACHE kmp_info {
2556  double th_align; /* use worst case alignment */
2557  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2558  kmp_base_info_t th;
2559 } kmp_info_t;
2560 
2561 // OpenMP thread team data structures
2562 
2563 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2564 
2565 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2566  double dt_align; /* use worst case alignment */
2567  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2568  kmp_base_data_t dt;
2569 } kmp_sleep_team_t;
2570 
2571 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2572  double dt_align; /* use worst case alignment */
2573  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2574  kmp_base_data_t dt;
2575 } kmp_ordered_team_t;
2576 
2577 typedef int (*launch_t)(int gtid);
2578 
2579 /* Minimum number of ARGV entries to malloc if necessary */
2580 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2581 
2582 // Set up how many argv pointers will fit in cache lines containing
2583 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2584 // larger value for more space between the master write/worker read section and
2585 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2586 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2587 #define KMP_INLINE_ARGV_BYTES \
2588  (4 * CACHE_LINE - \
2589  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2590  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2591  CACHE_LINE))
2592 #else
2593 #define KMP_INLINE_ARGV_BYTES \
2594  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2595 #endif
2596 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2597 
2598 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2599  // Synchronization Data
2600  // ---------------------------------------------------------------------------
2601  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2602  kmp_balign_team_t t_bar[bs_last_barrier];
2603  std::atomic<int> t_construct; // count of single directive encountered by team
2604  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2605 
2606  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2607  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2608  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2609 
2610  // Master only
2611  // ---------------------------------------------------------------------------
2612  KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2613  int t_master_this_cons; // "this_construct" single counter of master in parent
2614  // team
2615  ident_t *t_ident; // if volatile, have to change too much other crud to
2616  // volatile too
2617  kmp_team_p *t_parent; // parent team
2618  kmp_team_p *t_next_pool; // next free team in the team pool
2619  kmp_disp_t *t_dispatch; // thread's dispatch data
2620  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2621  kmp_proc_bind_t t_proc_bind; // bind type for par region
2622 #if USE_ITT_BUILD
2623  kmp_uint64 t_region_time; // region begin timestamp
2624 #endif /* USE_ITT_BUILD */
2625 
2626  // Master write, workers read
2627  // --------------------------------------------------------------------------
2628  KMP_ALIGN_CACHE void **t_argv;
2629  int t_argc;
2630  int t_nproc; // number of threads in team
2631  microtask_t t_pkfn;
2632  launch_t t_invoke; // procedure to launch the microtask
2633 
2634 #if OMPT_SUPPORT
2635  ompt_team_info_t ompt_team_info;
2636  ompt_lw_taskteam_t *ompt_serialized_team_info;
2637 #endif
2638 
2639 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2640  kmp_int8 t_fp_control_saved;
2641  kmp_int8 t_pad2b;
2642  kmp_int16 t_x87_fpu_control_word; // FP control regs
2643  kmp_uint32 t_mxcsr;
2644 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2645 
2646  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2647 
2648  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2649  kmp_taskdata_t
2650  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2651  int t_level; // nested parallel level
2652 
2653  KMP_ALIGN_CACHE int t_max_argc;
2654  int t_max_nproc; // max threads this team can handle (dynamically expandable)
2655  int t_serialized; // levels deep of serialized teams
2656  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2657  int t_id; // team's id, assigned by debugger.
2658  int t_active_level; // nested active parallel level
2659  kmp_r_sched_t t_sched; // run-time schedule for the team
2660 #if KMP_AFFINITY_SUPPORTED
2661  int t_first_place; // first & last place in parent thread's partition.
2662  int t_last_place; // Restore these values to master after par region.
2663 #endif // KMP_AFFINITY_SUPPORTED
2664  int t_display_affinity;
2665  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2666  // omp_set_num_threads() call
2667  omp_allocator_handle_t t_def_allocator; /* default allocator */
2668 
2669 // Read/write by workers as well
2670 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2671  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2672  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2673  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2674  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2675  char dummy_padding[1024];
2676 #endif
2677  // Internal control stack for additional nested teams.
2678  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2679  // for SERIALIZED teams nested 2 or more levels deep
2680  // typed flag to store request state of cancellation
2681  std::atomic<kmp_int32> t_cancel_request;
2682  int t_master_active; // save on fork, restore on join
2683  void *t_copypriv_data; // team specific pointer to copyprivate data array
2684 #if KMP_OS_WINDOWS
2685  std::atomic<kmp_uint32> t_copyin_counter;
2686 #endif
2687 #if USE_ITT_BUILD
2688  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2689 #endif /* USE_ITT_BUILD */
2690 } kmp_base_team_t;
2691 
2692 union KMP_ALIGN_CACHE kmp_team {
2693  kmp_base_team_t t;
2694  double t_align; /* use worst case alignment */
2695  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2696 };
2697 
2698 typedef union KMP_ALIGN_CACHE kmp_time_global {
2699  double dt_align; /* use worst case alignment */
2700  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2701  kmp_base_data_t dt;
2702 } kmp_time_global_t;
2703 
2704 typedef struct kmp_base_global {
2705  /* cache-aligned */
2706  kmp_time_global_t g_time;
2707 
2708  /* non cache-aligned */
2709  volatile int g_abort;
2710  volatile int g_done;
2711 
2712  int g_dynamic;
2713  enum dynamic_mode g_dynamic_mode;
2714 } kmp_base_global_t;
2715 
2716 typedef union KMP_ALIGN_CACHE kmp_global {
2717  kmp_base_global_t g;
2718  double g_align; /* use worst case alignment */
2719  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2720 } kmp_global_t;
2721 
2722 typedef struct kmp_base_root {
2723  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2724  // (r_in_parallel>= 0)
2725  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2726  // the synch overhead or keeping r_active
2727  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2728  // keeps a count of active parallel regions per root
2729  std::atomic<int> r_in_parallel;
2730  // GEH: This is misnamed, should be r_active_levels
2731  kmp_team_t *r_root_team;
2732  kmp_team_t *r_hot_team;
2733  kmp_info_t *r_uber_thread;
2734  kmp_lock_t r_begin_lock;
2735  volatile int r_begin;
2736  int r_blocktime; /* blocktime for this root and descendants */
2737 } kmp_base_root_t;
2738 
2739 typedef union KMP_ALIGN_CACHE kmp_root {
2740  kmp_base_root_t r;
2741  double r_align; /* use worst case alignment */
2742  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2743 } kmp_root_t;
2744 
2745 struct fortran_inx_info {
2746  kmp_int32 data;
2747 };
2748 
2749 /* ------------------------------------------------------------------------ */
2750 
2751 extern int __kmp_settings;
2752 extern int __kmp_duplicate_library_ok;
2753 #if USE_ITT_BUILD
2754 extern int __kmp_forkjoin_frames;
2755 extern int __kmp_forkjoin_frames_mode;
2756 #endif
2757 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2758 extern int __kmp_determ_red;
2759 
2760 #ifdef KMP_DEBUG
2761 extern int kmp_a_debug;
2762 extern int kmp_b_debug;
2763 extern int kmp_c_debug;
2764 extern int kmp_d_debug;
2765 extern int kmp_e_debug;
2766 extern int kmp_f_debug;
2767 #endif /* KMP_DEBUG */
2768 
2769 /* For debug information logging using rotating buffer */
2770 #define KMP_DEBUG_BUF_LINES_INIT 512
2771 #define KMP_DEBUG_BUF_LINES_MIN 1
2772 
2773 #define KMP_DEBUG_BUF_CHARS_INIT 128
2774 #define KMP_DEBUG_BUF_CHARS_MIN 2
2775 
2776 extern int
2777  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2778 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2779 extern int
2780  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2781 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2782  entry pointer */
2783 
2784 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2785 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2786  printed in buffer so far */
2787 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2788  recommended in warnings */
2789 /* end rotating debug buffer */
2790 
2791 #ifdef KMP_DEBUG
2792 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2793 
2794 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2795 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2796 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2797 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2798 extern int __kmp_par_range_lb;
2799 extern int __kmp_par_range_ub;
2800 #endif
2801 
2802 /* For printing out dynamic storage map for threads and teams */
2803 extern int
2804  __kmp_storage_map; /* True means print storage map for threads and teams */
2805 extern int __kmp_storage_map_verbose; /* True means storage map includes
2806  placement info */
2807 extern int __kmp_storage_map_verbose_specified;
2808 
2809 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2810 extern kmp_cpuinfo_t __kmp_cpuinfo;
2811 #endif
2812 
2813 extern volatile int __kmp_init_serial;
2814 extern volatile int __kmp_init_gtid;
2815 extern volatile int __kmp_init_common;
2816 extern volatile int __kmp_init_middle;
2817 extern volatile int __kmp_init_parallel;
2818 #if KMP_USE_MONITOR
2819 extern volatile int __kmp_init_monitor;
2820 #endif
2821 extern volatile int __kmp_init_user_locks;
2822 extern int __kmp_init_counter;
2823 extern int __kmp_root_counter;
2824 extern int __kmp_version;
2825 
2826 /* list of address of allocated caches for commons */
2827 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2828 
2829 /* Barrier algorithm types and options */
2830 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2831 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2832 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2833 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2834 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2835 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2836 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2837 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2838 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2839 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2840 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2841 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2842 
2843 /* Global Locks */
2844 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2845 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2846 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2847 extern kmp_bootstrap_lock_t
2848  __kmp_exit_lock; /* exit() is not always thread-safe */
2849 #if KMP_USE_MONITOR
2850 extern kmp_bootstrap_lock_t
2851  __kmp_monitor_lock; /* control monitor thread creation */
2852 #endif
2853 extern kmp_bootstrap_lock_t
2854  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2855  __kmp_threads expansion to co-exist */
2856 
2857 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
2858 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
2859 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2860 
2861 extern enum library_type __kmp_library;
2862 
2863 extern enum sched_type __kmp_sched; /* default runtime scheduling */
2864 extern enum sched_type __kmp_static; /* default static scheduling method */
2865 extern enum sched_type __kmp_guided; /* default guided scheduling method */
2866 extern enum sched_type __kmp_auto; /* default auto scheduling method */
2867 extern int __kmp_chunk; /* default runtime chunk size */
2868 
2869 extern size_t __kmp_stksize; /* stack size per thread */
2870 #if KMP_USE_MONITOR
2871 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
2872 #endif
2873 extern size_t __kmp_stkoffset; /* stack offset per thread */
2874 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
2875 
2876 extern size_t
2877  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
2878 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
2879 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
2880 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
2881 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
2882 extern int __kmp_generate_warnings; /* should we issue warnings? */
2883 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
2884 
2885 #ifdef DEBUG_SUSPEND
2886 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
2887 #endif
2888 
2889 extern kmp_int32 __kmp_use_yield;
2890 extern kmp_int32 __kmp_use_yield_exp_set;
2891 extern kmp_uint32 __kmp_yield_init;
2892 extern kmp_uint32 __kmp_yield_next;
2893 
2894 /* ------------------------------------------------------------------------- */
2895 extern int __kmp_allThreadsSpecified;
2896 
2897 extern size_t __kmp_align_alloc;
2898 /* following data protected by initialization routines */
2899 extern int __kmp_xproc; /* number of processors in the system */
2900 extern int __kmp_avail_proc; /* number of processors available to the process */
2901 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
2902 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
2903 // maximum total number of concurrently-existing threads on device
2904 extern int __kmp_max_nth;
2905 // maximum total number of concurrently-existing threads in a contention group
2906 extern int __kmp_cg_max_nth;
2907 extern int __kmp_teams_max_nth; // max threads used in a teams construct
2908 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
2909  __kmp_root */
2910 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
2911  region a la OMP_NUM_THREADS */
2912 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
2913  initialization */
2914 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
2915  used (fixed) */
2916 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
2917  (__kmpc_threadprivate_cached()) */
2918 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
2919  blocking (env setting) */
2920 #if KMP_USE_MONITOR
2921 extern int
2922  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
2923 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
2924  blocking */
2925 #endif
2926 #ifdef KMP_ADJUST_BLOCKTIME
2927 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
2928 #endif /* KMP_ADJUST_BLOCKTIME */
2929 #ifdef KMP_DFLT_NTH_CORES
2930 extern int __kmp_ncores; /* Total number of cores for threads placement */
2931 #endif
2932 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
2933 extern int __kmp_abort_delay;
2934 
2935 extern int __kmp_need_register_atfork_specified;
2936 extern int
2937  __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
2938  install fork handler */
2939 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
2940  0 - not set, will be set at runtime
2941  1 - using stack search
2942  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
2943  X*) or TlsGetValue(Windows* OS))
2944  3 - static TLS (__declspec(thread) __kmp_gtid),
2945  Linux* OS .so only. */
2946 extern int
2947  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
2948 #ifdef KMP_TDATA_GTID
2949 extern KMP_THREAD_LOCAL int __kmp_gtid;
2950 #endif
2951 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
2952 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
2953 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2954 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
2955 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
2956 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
2957 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2958 
2959 // max_active_levels for nested parallelism enabled by default via
2960 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
2961 extern int __kmp_dflt_max_active_levels;
2962 // Indicates whether value of __kmp_dflt_max_active_levels was already
2963 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
2964 extern bool __kmp_dflt_max_active_levels_set;
2965 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
2966  concurrent execution per team */
2967 #if KMP_NESTED_HOT_TEAMS
2968 extern int __kmp_hot_teams_mode;
2969 extern int __kmp_hot_teams_max_level;
2970 #endif
2971 
2972 #if KMP_OS_LINUX
2973 extern enum clock_function_type __kmp_clock_function;
2974 extern int __kmp_clock_function_param;
2975 #endif /* KMP_OS_LINUX */
2976 
2977 #if KMP_MIC_SUPPORTED
2978 extern enum mic_type __kmp_mic_type;
2979 #endif
2980 
2981 #ifdef USE_LOAD_BALANCE
2982 extern double __kmp_load_balance_interval; // load balance algorithm interval
2983 #endif /* USE_LOAD_BALANCE */
2984 
2985 // OpenMP 3.1 - Nested num threads array
2986 typedef struct kmp_nested_nthreads_t {
2987  int *nth;
2988  int size;
2989  int used;
2990 } kmp_nested_nthreads_t;
2991 
2992 extern kmp_nested_nthreads_t __kmp_nested_nth;
2993 
2994 #if KMP_USE_ADAPTIVE_LOCKS
2995 
2996 // Parameters for the speculative lock backoff system.
2997 struct kmp_adaptive_backoff_params_t {
2998  // Number of soft retries before it counts as a hard retry.
2999  kmp_uint32 max_soft_retries;
3000  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3001  // the right
3002  kmp_uint32 max_badness;
3003 };
3004 
3005 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3006 
3007 #if KMP_DEBUG_ADAPTIVE_LOCKS
3008 extern const char *__kmp_speculative_statsfile;
3009 #endif
3010 
3011 #endif // KMP_USE_ADAPTIVE_LOCKS
3012 
3013 extern int __kmp_display_env; /* TRUE or FALSE */
3014 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3015 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3016 
3017 /* ------------------------------------------------------------------------- */
3018 
3019 /* the following are protected by the fork/join lock */
3020 /* write: lock read: anytime */
3021 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3022 /* read/write: lock */
3023 extern volatile kmp_team_t *__kmp_team_pool;
3024 extern volatile kmp_info_t *__kmp_thread_pool;
3025 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3026 
3027 // total num threads reachable from some root thread including all root threads
3028 extern volatile int __kmp_nth;
3029 /* total number of threads reachable from some root thread including all root
3030  threads, and those in the thread pool */
3031 extern volatile int __kmp_all_nth;
3032 extern std::atomic<int> __kmp_thread_pool_active_nth;
3033 
3034 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3035 /* end data protected by fork/join lock */
3036 /* ------------------------------------------------------------------------- */
3037 
3038 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3039 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3040 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3041 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3042 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3043 
3044 // AT: Which way is correct?
3045 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3046 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3047 #define __kmp_get_team_num_threads(gtid) \
3048  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3049 
3050 static inline bool KMP_UBER_GTID(int gtid) {
3051  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3052  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3053  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3054  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3055 }
3056 
3057 static inline int __kmp_tid_from_gtid(int gtid) {
3058  KMP_DEBUG_ASSERT(gtid >= 0);
3059  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3060 }
3061 
3062 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3063  KMP_DEBUG_ASSERT(tid >= 0 && team);
3064  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3065 }
3066 
3067 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3068  KMP_DEBUG_ASSERT(thr);
3069  return thr->th.th_info.ds.ds_gtid;
3070 }
3071 
3072 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3073  KMP_DEBUG_ASSERT(gtid >= 0);
3074  return __kmp_threads[gtid];
3075 }
3076 
3077 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3078  KMP_DEBUG_ASSERT(gtid >= 0);
3079  return __kmp_threads[gtid]->th.th_team;
3080 }
3081 
3082 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3083  if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3084  KMP_FATAL(ThreadIdentInvalid);
3085 }
3086 
3087 /* ------------------------------------------------------------------------- */
3088 
3089 extern kmp_global_t __kmp_global; /* global status */
3090 
3091 extern kmp_info_t __kmp_monitor;
3092 // For Debugging Support Library
3093 extern std::atomic<kmp_int32> __kmp_team_counter;
3094 // For Debugging Support Library
3095 extern std::atomic<kmp_int32> __kmp_task_counter;
3096 
3097 #if USE_DEBUGGER
3098 #define _KMP_GEN_ID(counter) \
3099  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3100 #else
3101 #define _KMP_GEN_ID(counter) (~0)
3102 #endif /* USE_DEBUGGER */
3103 
3104 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3105 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3106 
3107 /* ------------------------------------------------------------------------ */
3108 
3109 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3110  size_t size, char const *format, ...);
3111 
3112 extern void __kmp_serial_initialize(void);
3113 extern void __kmp_middle_initialize(void);
3114 extern void __kmp_parallel_initialize(void);
3115 
3116 extern void __kmp_internal_begin(void);
3117 extern void __kmp_internal_end_library(int gtid);
3118 extern void __kmp_internal_end_thread(int gtid);
3119 extern void __kmp_internal_end_atexit(void);
3120 extern void __kmp_internal_end_dtor(void);
3121 extern void __kmp_internal_end_dest(void *);
3122 
3123 extern int __kmp_register_root(int initial_thread);
3124 extern void __kmp_unregister_root(int gtid);
3125 
3126 extern int __kmp_ignore_mppbeg(void);
3127 extern int __kmp_ignore_mppend(void);
3128 
3129 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3130 extern void __kmp_exit_single(int gtid);
3131 
3132 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3133 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3134 
3135 #ifdef USE_LOAD_BALANCE
3136 extern int __kmp_get_load_balance(int);
3137 #endif
3138 
3139 extern int __kmp_get_global_thread_id(void);
3140 extern int __kmp_get_global_thread_id_reg(void);
3141 extern void __kmp_exit_thread(int exit_status);
3142 extern void __kmp_abort(char const *format, ...);
3143 extern void __kmp_abort_thread(void);
3144 KMP_NORETURN extern void __kmp_abort_process(void);
3145 extern void __kmp_warn(char const *format, ...);
3146 
3147 extern void __kmp_set_num_threads(int new_nth, int gtid);
3148 
3149 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3150 // registered.
3151 static inline kmp_info_t *__kmp_entry_thread() {
3152  int gtid = __kmp_entry_gtid();
3153 
3154  return __kmp_threads[gtid];
3155 }
3156 
3157 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3158 extern int __kmp_get_max_active_levels(int gtid);
3159 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3160 extern int __kmp_get_team_size(int gtid, int level);
3161 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3162 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3163 
3164 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3165 extern void __kmp_init_random(kmp_info_t *thread);
3166 
3167 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3168 extern void __kmp_adjust_num_threads(int new_nproc);
3169 extern void __kmp_check_stksize(size_t *val);
3170 
3171 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3172 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3173 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3174 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3175 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3176 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3177 
3178 #if USE_FAST_MEMORY
3179 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3180  size_t size KMP_SRC_LOC_DECL);
3181 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3182 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3183 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3184 #define __kmp_fast_allocate(this_thr, size) \
3185  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3186 #define __kmp_fast_free(this_thr, ptr) \
3187  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3188 #endif
3189 
3190 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3191 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3192  size_t elsize KMP_SRC_LOC_DECL);
3193 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3194  size_t size KMP_SRC_LOC_DECL);
3195 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3196 #define __kmp_thread_malloc(th, size) \
3197  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3198 #define __kmp_thread_calloc(th, nelem, elsize) \
3199  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3200 #define __kmp_thread_realloc(th, ptr, size) \
3201  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3202 #define __kmp_thread_free(th, ptr) \
3203  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3204 
3205 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3206 #define KMP_INTERNAL_FREE(p) free(p)
3207 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3208 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3209 
3210 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3211 
3212 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3213  kmp_proc_bind_t proc_bind);
3214 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3215  int num_threads);
3216 
3217 extern void __kmp_yield();
3218 
3219 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3220  enum sched_type schedule, kmp_int32 lb,
3221  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3222 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3223  enum sched_type schedule, kmp_uint32 lb,
3224  kmp_uint32 ub, kmp_int32 st,
3225  kmp_int32 chunk);
3226 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3227  enum sched_type schedule, kmp_int64 lb,
3228  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3229 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3230  enum sched_type schedule, kmp_uint64 lb,
3231  kmp_uint64 ub, kmp_int64 st,
3232  kmp_int64 chunk);
3233 
3234 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3235  kmp_int32 *p_last, kmp_int32 *p_lb,
3236  kmp_int32 *p_ub, kmp_int32 *p_st);
3237 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3238  kmp_int32 *p_last, kmp_uint32 *p_lb,
3239  kmp_uint32 *p_ub, kmp_int32 *p_st);
3240 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3241  kmp_int32 *p_last, kmp_int64 *p_lb,
3242  kmp_int64 *p_ub, kmp_int64 *p_st);
3243 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3244  kmp_int32 *p_last, kmp_uint64 *p_lb,
3245  kmp_uint64 *p_ub, kmp_int64 *p_st);
3246 
3247 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3248 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3249 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3250 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3251 
3252 #ifdef KMP_GOMP_COMPAT
3253 
3254 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3255  enum sched_type schedule, kmp_int32 lb,
3256  kmp_int32 ub, kmp_int32 st,
3257  kmp_int32 chunk, int push_ws);
3258 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3259  enum sched_type schedule, kmp_uint32 lb,
3260  kmp_uint32 ub, kmp_int32 st,
3261  kmp_int32 chunk, int push_ws);
3262 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3263  enum sched_type schedule, kmp_int64 lb,
3264  kmp_int64 ub, kmp_int64 st,
3265  kmp_int64 chunk, int push_ws);
3266 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3267  enum sched_type schedule, kmp_uint64 lb,
3268  kmp_uint64 ub, kmp_int64 st,
3269  kmp_int64 chunk, int push_ws);
3270 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3271 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3272 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3273 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3274 
3275 #endif /* KMP_GOMP_COMPAT */
3276 
3277 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3278 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3279 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3280 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3281 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3282 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3283  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3284  void *obj);
3285 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3286  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3287 
3288 class kmp_flag_32;
3289 class kmp_flag_64;
3290 class kmp_flag_oncore;
3291 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3292  int final_spin
3293 #if USE_ITT_BUILD
3294  ,
3295  void *itt_sync_obj
3296 #endif
3297  );
3298 extern void __kmp_release_64(kmp_flag_64 *flag);
3299 
3300 extern void __kmp_infinite_loop(void);
3301 
3302 extern void __kmp_cleanup(void);
3303 
3304 #if KMP_HANDLE_SIGNALS
3305 extern int __kmp_handle_signals;
3306 extern void __kmp_install_signals(int parallel_init);
3307 extern void __kmp_remove_signals(void);
3308 #endif
3309 
3310 extern void __kmp_clear_system_time(void);
3311 extern void __kmp_read_system_time(double *delta);
3312 
3313 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3314 
3315 extern void __kmp_expand_host_name(char *buffer, size_t size);
3316 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3317 
3318 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3319 extern void
3320 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3321 #endif
3322 
3323 extern void
3324 __kmp_runtime_initialize(void); /* machine specific initialization */
3325 extern void __kmp_runtime_destroy(void);
3326 
3327 #if KMP_AFFINITY_SUPPORTED
3328 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3329  kmp_affin_mask_t *mask);
3330 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3331  kmp_affin_mask_t *mask);
3332 extern void __kmp_affinity_initialize(void);
3333 extern void __kmp_affinity_uninitialize(void);
3334 extern void __kmp_affinity_set_init_mask(
3335  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3336 extern void __kmp_affinity_set_place(int gtid);
3337 extern void __kmp_affinity_determine_capable(const char *env_var);
3338 extern int __kmp_aux_set_affinity(void **mask);
3339 extern int __kmp_aux_get_affinity(void **mask);
3340 extern int __kmp_aux_get_affinity_max_proc();
3341 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3342 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3343 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3344 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3345 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3346 extern int kmp_set_thread_affinity_mask_initial(void);
3347 #endif
3348 #endif /* KMP_AFFINITY_SUPPORTED */
3349 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3350 // format string is for affinity, so platforms that do not support
3351 // affinity can still use the other fields, e.g., %n for num_threads
3352 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3353  kmp_str_buf_t *buffer);
3354 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3355 
3356 extern void __kmp_cleanup_hierarchy();
3357 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3358 
3359 #if KMP_USE_FUTEX
3360 
3361 extern int __kmp_futex_determine_capable(void);
3362 
3363 #endif // KMP_USE_FUTEX
3364 
3365 extern void __kmp_gtid_set_specific(int gtid);
3366 extern int __kmp_gtid_get_specific(void);
3367 
3368 extern double __kmp_read_cpu_time(void);
3369 
3370 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3371 
3372 #if KMP_USE_MONITOR
3373 extern void __kmp_create_monitor(kmp_info_t *th);
3374 #endif
3375 
3376 extern void *__kmp_launch_thread(kmp_info_t *thr);
3377 
3378 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3379 
3380 #if KMP_OS_WINDOWS
3381 extern int __kmp_still_running(kmp_info_t *th);
3382 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3383 extern void __kmp_free_handle(kmp_thread_t tHandle);
3384 #endif
3385 
3386 #if KMP_USE_MONITOR
3387 extern void __kmp_reap_monitor(kmp_info_t *th);
3388 #endif
3389 extern void __kmp_reap_worker(kmp_info_t *th);
3390 extern void __kmp_terminate_thread(int gtid);
3391 
3392 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3393 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3394 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3395 
3396 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag);
3397 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag);
3398 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
3399 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag);
3400 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag);
3401 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
3402 
3403 extern void __kmp_elapsed(double *);
3404 extern void __kmp_elapsed_tick(double *);
3405 
3406 extern void __kmp_enable(int old_state);
3407 extern void __kmp_disable(int *old_state);
3408 
3409 extern void __kmp_thread_sleep(int millis);
3410 
3411 extern void __kmp_common_initialize(void);
3412 extern void __kmp_common_destroy(void);
3413 extern void __kmp_common_destroy_gtid(int gtid);
3414 
3415 #if KMP_OS_UNIX
3416 extern void __kmp_register_atfork(void);
3417 #endif
3418 extern void __kmp_suspend_initialize(void);
3419 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3420 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3421 
3422 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3423  int tid);
3424 extern kmp_team_t *
3425 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3426 #if OMPT_SUPPORT
3427  ompt_data_t ompt_parallel_data,
3428 #endif
3429  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3430  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3431 extern void __kmp_free_thread(kmp_info_t *);
3432 extern void __kmp_free_team(kmp_root_t *,
3433  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3434 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3435 
3436 /* ------------------------------------------------------------------------ */
3437 
3438 extern void __kmp_initialize_bget(kmp_info_t *th);
3439 extern void __kmp_finalize_bget(kmp_info_t *th);
3440 
3441 KMP_EXPORT void *kmpc_malloc(size_t size);
3442 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3443 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3444 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3445 KMP_EXPORT void kmpc_free(void *ptr);
3446 
3447 /* declarations for internal use */
3448 
3449 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3450  size_t reduce_size, void *reduce_data,
3451  void (*reduce)(void *, void *));
3452 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3453 extern int __kmp_barrier_gomp_cancel(int gtid);
3454 
3459 enum fork_context_e {
3460  fork_context_gnu,
3462  fork_context_intel,
3463  fork_context_last
3464 };
3465 extern int __kmp_fork_call(ident_t *loc, int gtid,
3466  enum fork_context_e fork_context, kmp_int32 argc,
3467  microtask_t microtask, launch_t invoker,
3468  kmp_va_list ap);
3469 
3470 extern void __kmp_join_call(ident_t *loc, int gtid
3471 #if OMPT_SUPPORT
3472  ,
3473  enum fork_context_e fork_context
3474 #endif
3475  ,
3476  int exit_teams = 0);
3477 
3478 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3479 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3480 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3481 extern int __kmp_invoke_task_func(int gtid);
3482 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3483  kmp_info_t *this_thr,
3484  kmp_team_t *team);
3485 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3486  kmp_info_t *this_thr,
3487  kmp_team_t *team);
3488 
3489 // should never have been exported
3490 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3491 extern int __kmp_invoke_teams_master(int gtid);
3492 extern void __kmp_teams_master(int gtid);
3493 extern int __kmp_aux_get_team_num();
3494 extern int __kmp_aux_get_num_teams();
3495 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3496 extern void __kmp_user_set_library(enum library_type arg);
3497 extern void __kmp_aux_set_library(enum library_type arg);
3498 extern void __kmp_aux_set_stacksize(size_t arg);
3499 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3500 extern void __kmp_aux_set_defaults(char const *str, int len);
3501 
3502 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3503 void kmpc_set_blocktime(int arg);
3504 void ompc_set_nested(int flag);
3505 void ompc_set_dynamic(int flag);
3506 void ompc_set_num_threads(int arg);
3507 
3508 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3509  kmp_team_t *team, int tid);
3510 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3511 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3512  kmp_tasking_flags_t *flags,
3513  size_t sizeof_kmp_task_t,
3514  size_t sizeof_shareds,
3515  kmp_routine_entry_t task_entry);
3516 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3517  kmp_team_t *team, int tid,
3518  int set_curr_task);
3519 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3520 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3521 
3522 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3523  int gtid,
3524  kmp_task_t *task);
3525 extern void __kmp_fulfill_event(kmp_event_t *event);
3526 
3527 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3528  kmp_flag_32 *flag, int final_spin,
3529  int *thread_finished,
3530 #if USE_ITT_BUILD
3531  void *itt_sync_obj,
3532 #endif /* USE_ITT_BUILD */
3533  kmp_int32 is_constrained);
3534 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3535  kmp_flag_64 *flag, int final_spin,
3536  int *thread_finished,
3537 #if USE_ITT_BUILD
3538  void *itt_sync_obj,
3539 #endif /* USE_ITT_BUILD */
3540  kmp_int32 is_constrained);
3541 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3542  kmp_flag_oncore *flag, int final_spin,
3543  int *thread_finished,
3544 #if USE_ITT_BUILD
3545  void *itt_sync_obj,
3546 #endif /* USE_ITT_BUILD */
3547  kmp_int32 is_constrained);
3548 
3549 extern void __kmp_free_task_team(kmp_info_t *thread,
3550  kmp_task_team_t *task_team);
3551 extern void __kmp_reap_task_teams(void);
3552 extern void __kmp_wait_to_unref_task_teams(void);
3553 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3554  int always);
3555 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3556 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3557 #if USE_ITT_BUILD
3558  ,
3559  void *itt_sync_obj
3560 #endif /* USE_ITT_BUILD */
3561  ,
3562  int wait = 1);
3563 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3564  int gtid);
3565 
3566 extern int __kmp_is_address_mapped(void *addr);
3567 extern kmp_uint64 __kmp_hardware_timestamp(void);
3568 
3569 #if KMP_OS_UNIX
3570 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3571 #endif
3572 
3573 /* ------------------------------------------------------------------------ */
3574 //
3575 // Assembly routines that have no compiler intrinsic replacement
3576 //
3577 
3578 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3579  void *argv[]
3580 #if OMPT_SUPPORT
3581  ,
3582  void **exit_frame_ptr
3583 #endif
3584  );
3585 
3586 /* ------------------------------------------------------------------------ */
3587 
3588 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3589 KMP_EXPORT void __kmpc_end(ident_t *);
3590 
3591 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3592  kmpc_ctor_vec ctor,
3593  kmpc_cctor_vec cctor,
3594  kmpc_dtor_vec dtor,
3595  size_t vector_length);
3596 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3597  kmpc_ctor ctor, kmpc_cctor cctor,
3598  kmpc_dtor dtor);
3599 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3600  void *data, size_t size);
3601 
3602 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3603 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3604 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3605 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3606 
3607 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3608 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3609  kmpc_micro microtask, ...);
3610 
3611 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3612 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3613 
3614 KMP_EXPORT void __kmpc_flush(ident_t *);
3615 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3616 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3617 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3618 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3619 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3620 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3621  kmp_critical_name *);
3622 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3623  kmp_critical_name *);
3624 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3625  kmp_critical_name *, uint32_t hint);
3626 
3627 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3628 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3629 
3630 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3631  kmp_int32 global_tid);
3632 
3633 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3634 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3635 
3636 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3637  kmp_int32 schedtype, kmp_int32 *plastiter,
3638  kmp_int *plower, kmp_int *pupper,
3639  kmp_int *pstride, kmp_int incr,
3640  kmp_int chunk);
3641 
3642 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3643 
3644 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3645  size_t cpy_size, void *cpy_data,
3646  void (*cpy_func)(void *, void *),
3647  kmp_int32 didit);
3648 
3649 extern void KMPC_SET_NUM_THREADS(int arg);
3650 extern void KMPC_SET_DYNAMIC(int flag);
3651 extern void KMPC_SET_NESTED(int flag);
3652 
3653 /* OMP 3.0 tasking interface routines */
3654 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3655  kmp_task_t *new_task);
3656 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3657  kmp_int32 flags,
3658  size_t sizeof_kmp_task_t,
3659  size_t sizeof_shareds,
3660  kmp_routine_entry_t task_entry);
3661 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3662  kmp_int32 flags,
3663  size_t sizeof_kmp_task_t,
3664  size_t sizeof_shareds,
3665  kmp_routine_entry_t task_entry,
3666  kmp_int64 device_id);
3667 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3668  kmp_task_t *task);
3669 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3670  kmp_task_t *task);
3671 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3672  kmp_task_t *new_task);
3673 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3674 
3675 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3676  int end_part);
3677 
3678 #if TASK_UNUSED
3679 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3680 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3681  kmp_task_t *task);
3682 #endif // TASK_UNUSED
3683 
3684 /* ------------------------------------------------------------------------ */
3685 
3686 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3687 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3688 
3689 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3690  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3691  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3692  kmp_depend_info_t *noalias_dep_list);
3693 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3694  kmp_int32 ndeps,
3695  kmp_depend_info_t *dep_list,
3696  kmp_int32 ndeps_noalias,
3697  kmp_depend_info_t *noalias_dep_list);
3698 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3699  bool serialize_immediate);
3700 
3701 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3702  kmp_int32 cncl_kind);
3703 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3704  kmp_int32 cncl_kind);
3705 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3706 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3707 
3708 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3709 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3710 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3711  kmp_int32 if_val, kmp_uint64 *lb,
3712  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3713  kmp_int32 sched, kmp_uint64 grainsize,
3714  void *task_dup);
3715 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3716 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3717 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3718 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3719  int is_ws, int num,
3720  void *data);
3721 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3722  int num, void *data);
3723 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3724  int is_ws);
3725 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3726  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3727  kmp_task_affinity_info_t *affin_list);
3728 
3729 /* Lock interface routines (fast versions with gtid passed in) */
3730 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3731  void **user_lock);
3732 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3733  void **user_lock);
3734 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3735  void **user_lock);
3736 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3737  void **user_lock);
3738 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3739 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3740  void **user_lock);
3741 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3742  void **user_lock);
3743 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3744  void **user_lock);
3745 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3746 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3747  void **user_lock);
3748 
3749 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3750  void **user_lock, uintptr_t hint);
3751 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3752  void **user_lock,
3753  uintptr_t hint);
3754 
3755 /* Interface to fast scalable reduce methods routines */
3756 
3757 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3758  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3759  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3760  kmp_critical_name *lck);
3761 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3762  kmp_critical_name *lck);
3763 KMP_EXPORT kmp_int32 __kmpc_reduce(
3764  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3765  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3766  kmp_critical_name *lck);
3767 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3768  kmp_critical_name *lck);
3769 
3770 /* Internal fast reduction routines */
3771 
3772 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3773  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3774  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3775  kmp_critical_name *lck);
3776 
3777 // this function is for testing set/get/determine reduce method
3778 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3779 
3780 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3781 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3782 
3783 // C++ port
3784 // missing 'extern "C"' declarations
3785 
3786 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3787 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3788 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3789  kmp_int32 num_threads);
3790 
3791 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3792  int proc_bind);
3793 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3794  kmp_int32 num_teams,
3795  kmp_int32 num_threads);
3796 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3797  kmpc_micro microtask, ...);
3798 struct kmp_dim { // loop bounds info casted to kmp_int64
3799  kmp_int64 lo; // lower
3800  kmp_int64 up; // upper
3801  kmp_int64 st; // stride
3802 };
3803 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3804  kmp_int32 num_dims,
3805  const struct kmp_dim *dims);
3806 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3807  const kmp_int64 *vec);
3808 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3809  const kmp_int64 *vec);
3810 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3811 
3812 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3813  void *data, size_t size,
3814  void ***cache);
3815 
3816 // Symbols for MS mutual detection.
3817 extern int _You_must_link_with_exactly_one_OpenMP_library;
3818 extern int _You_must_link_with_Intel_OpenMP_library;
3819 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3820 extern int _You_must_link_with_Microsoft_OpenMP_library;
3821 #endif
3822 
3823 // The routines below are not exported.
3824 // Consider making them 'static' in corresponding source files.
3825 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3826  void *data_addr, size_t pc_size);
3827 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3828  void *data_addr,
3829  size_t pc_size);
3830 void __kmp_threadprivate_resize_cache(int newCapacity);
3831 void __kmp_cleanup_threadprivate_caches();
3832 
3833 // ompc_, kmpc_ entries moved from omp.h.
3834 #if KMP_OS_WINDOWS
3835 #define KMPC_CONVENTION __cdecl
3836 #else
3837 #define KMPC_CONVENTION
3838 #endif
3839 
3840 #ifndef __OMP_H
3841 typedef enum omp_sched_t {
3842  omp_sched_static = 1,
3843  omp_sched_dynamic = 2,
3844  omp_sched_guided = 3,
3845  omp_sched_auto = 4
3846 } omp_sched_t;
3847 typedef void *kmp_affinity_mask_t;
3848 #endif
3849 
3850 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3851 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3852 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3853 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3854 KMP_EXPORT int KMPC_CONVENTION
3855 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3856 KMP_EXPORT int KMPC_CONVENTION
3857 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3858 KMP_EXPORT int KMPC_CONVENTION
3859 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3860 
3861 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3862 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3863 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3864 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
3865 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
3866 
3867 enum kmp_target_offload_kind {
3868  tgt_disabled = 0,
3869  tgt_default = 1,
3870  tgt_mandatory = 2
3871 };
3872 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3873 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
3874 extern kmp_target_offload_kind_t __kmp_target_offload;
3875 extern int __kmpc_get_target_offload();
3876 
3877 // Constants used in libomptarget
3878 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
3879 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
3880 
3881 // OMP Pause Resource
3882 
3883 // The following enum is used both to set the status in __kmp_pause_status, and
3884 // as the internal equivalent of the externally-visible omp_pause_resource_t.
3885 typedef enum kmp_pause_status_t {
3886  kmp_not_paused = 0, // status is not paused, or, requesting resume
3887  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
3888  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
3889 } kmp_pause_status_t;
3890 
3891 // This stores the pause state of the runtime
3892 extern kmp_pause_status_t __kmp_pause_status;
3893 extern int __kmpc_pause_resource(kmp_pause_status_t level);
3894 extern int __kmp_pause_resource(kmp_pause_status_t level);
3895 // Soft resume sets __kmp_pause_status, and wakes up all threads.
3896 extern void __kmp_resume_if_soft_paused();
3897 // Hard resume simply resets the status to not paused. Library will appear to
3898 // be uninitialized after hard pause. Let OMP constructs trigger required
3899 // initializations.
3900 static inline void __kmp_resume_if_hard_paused() {
3901  if (__kmp_pause_status == kmp_hard_paused) {
3902  __kmp_pause_status = kmp_not_paused;
3903  }
3904 }
3905 
3906 extern void __kmp_omp_display_env(int verbose);
3907 
3908 #ifdef __cplusplus
3909 }
3910 #endif
3911 
3912 #endif /* KMP_H */
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
kmp_int32 reserved_2
Definition: kmp.h:230
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
void(* kmpc_dtor)(void *)
Definition: kmp.h:1420
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)
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 kmp_int32 __kmpc_global_thread_num(ident_t *)
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)
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1443
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
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_int32 reserved_1
Definition: kmp.h:227
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1437
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_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, void *data, size_t size, void ***cache)
kmp_int32 reserved_3
Definition: kmp.h:235
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition: kmp.h:1449
KMP_EXPORT void __kmpc_flush(ident_t *)
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)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
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)
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)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1427
void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims, const struct kmp_dim *dims)
KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor)
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 void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
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)
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)
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)
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
sched_type
Definition: kmp.h:340
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
Definition: kmp.h:226
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_barrier(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 void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1414
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
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)
KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, kmp_critical_name *, uint32_t hint)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1396
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(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)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
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)
char const * psource
Definition: kmp.h:236
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_int32 flags
Definition: kmp.h:228
struct ident ident_t