1	/*! \file */
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	#define KMP_WEIGHTED_ITERATIONS_SUPPORTED \
31	(KMP_AFFINITY_SUPPORTED && KMP_STATIC_STEAL_ENABLED && \
32	(KMP_ARCH_X86 || KMP_ARCH_X86_64))
33
34	#define TASK_CURRENT_NOT_QUEUED 0
35	#define TASK_CURRENT_QUEUED 1
36
37	#ifdef BUILD_TIED_TASK_STACK
38	#define TASK_STACK_EMPTY 0 // entries when the stack is empty
39	#define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
40	// Number of entries in each task stack array
41	#define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
42	// Mask for determining index into stack block
43	#define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
44	#endif // BUILD_TIED_TASK_STACK
45
46	#define TASK_NOT_PUSHED 1
47	#define TASK_SUCCESSFULLY_PUSHED 0
48	#define TASK_TIED 1
49	#define TASK_UNTIED 0
50	#define TASK_EXPLICIT 1
51	#define TASK_IMPLICIT 0
52	#define TASK_PROXY 1
53	#define TASK_FULL 0
54	#define TASK_DETACHABLE 1
55	#define TASK_UNDETACHABLE 0
56
57	#define KMP_CANCEL_THREADS
58	#define KMP_THREAD_ATTR
59
60	// Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
61	// built on Android
62	#if defined(__ANDROID__)
63	#undef KMP_CANCEL_THREADS
64	#endif
65
66	// Some WASI targets (e.g., wasm32-wasi-threads) do not support thread
67	// cancellation.
68	#if KMP_OS_WASI
69	#undef KMP_CANCEL_THREADS
70	#endif
71
72	#if !KMP_OS_WASI
73	#include <signal.h>
74	#endif
75	#include <stdarg.h>
76	#include <stddef.h>
77	#include <stdio.h>
78	#include <stdlib.h>
79	#include <string.h>
80	#include <limits>
81	#include <type_traits>
82	/* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
83	Microsoft library. Some macros provided below to replace these functions */
84	#ifndef __ABSOFT_WIN
85	#include <sys/types.h>
86	#endif
87	#include <limits.h>
88	#include <time.h>
89
90	#include <errno.h>
91
92	#include "kmp_os.h"
93
94	#include "kmp_safe_c_api.h"
95
96	#if KMP_STATS_ENABLED
97	class kmp_stats_list;
98	#endif
99
100	#if KMP_USE_HIER_SCHED
101	// Only include hierarchical scheduling if affinity is supported
102	#undef KMP_USE_HIER_SCHED
103	#define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
104	#endif
105
106	// OMPD_SKIP_HWLOC used in libompd/omp-icv.cpp to avoid OMPD depending on hwloc
107	#if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED && !defined(OMPD_SKIP_HWLOC)
108	#include "hwloc.h"
109	#ifndef HWLOC_OBJ_NUMANODE
110	#define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
111	#endif
112	#ifndef HWLOC_OBJ_PACKAGE
113	#define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
114	#endif
115	#endif
116
117	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
118	#include <xmmintrin.h>
119	#endif
120
121	// The below has to be defined before including "kmp_barrier.h".
122	#define KMP_INTERNAL_MALLOC(sz) malloc(sz)
123	#define KMP_INTERNAL_FREE(p) free(p)
124	#define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
125	#define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
126
127	#include "kmp_debug.h"
128	#include "kmp_lock.h"
129	#include "kmp_version.h"
130	#include "kmp_barrier.h"
131	#if USE_DEBUGGER
132	#include "kmp_debugger.h"
133	#endif
134	#include "kmp_i18n.h"
135
136	#define KMP_HANDLE_SIGNALS ((KMP_OS_UNIX && !KMP_OS_WASI) || KMP_OS_WINDOWS)
137
138	#include "kmp_wrapper_malloc.h"
139	#if KMP_OS_UNIX
140	#include <unistd.h>
141	#if !defined NSIG && defined _NSIG
142	#define NSIG _NSIG
143	#endif
144	#endif
145
146	#if KMP_OS_LINUX
147	#pragma weak clock_gettime
148	#endif
149
150	#if OMPT_SUPPORT
151	#include "ompt-internal.h"
152	#endif
153
154	#if OMPD_SUPPORT
155	#include "ompd-specific.h"
156	#endif
157
158	#ifndef UNLIKELY
159	#define UNLIKELY(x) (x)
160	#endif
161
162	// Affinity format function
163	#include "kmp_str.h"
164
165	// 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
166	// 3 - fast allocation using sync, non-sync free lists of any size, non-self
167	// free lists of limited size.
168	#ifndef USE_FAST_MEMORY
169	#define USE_FAST_MEMORY 3
170	#endif
171
172	#ifndef KMP_NESTED_HOT_TEAMS
173	#define KMP_NESTED_HOT_TEAMS 0
174	#define USE_NESTED_HOT_ARG(x)
175	#else
176	#if KMP_NESTED_HOT_TEAMS
177	#define USE_NESTED_HOT_ARG(x) , x
178	#else
179	#define USE_NESTED_HOT_ARG(x)
180	#endif
181	#endif
182
183	// Assume using BGET compare_exchange instruction instead of lock by default.
184	#ifndef USE_CMP_XCHG_FOR_BGET
185	#define USE_CMP_XCHG_FOR_BGET 1
186	#endif
187
188	// Test to see if queuing lock is better than bootstrap lock for bget
189	// #ifndef USE_QUEUING_LOCK_FOR_BGET
190	// #define USE_QUEUING_LOCK_FOR_BGET
191	// #endif
192
193	#define KMP_NSEC_PER_SEC 1000000000L
194	#define KMP_USEC_PER_SEC 1000000L
195	#define KMP_NSEC_PER_USEC 1000L
196
197	/*!
198	@ingroup BASIC_TYPES
199	@{
200	*/
201
202	/*!
203	Values for bit flags used in the ident_t to describe the fields.
204	*/
205	enum {
206	/*! Use trampoline for internal microtasks */
207	KMP_IDENT_IMB = 0x01,
208	/*! Use c-style ident structure */
209	KMP_IDENT_KMPC = 0x02,
210	/* 0x04 is no longer used */
211	/*! Entry point generated by auto-parallelization */
212	KMP_IDENT_AUTOPAR = 0x08,
213	/*! Compiler generates atomic reduction option for kmpc_reduce* */
214	KMP_IDENT_ATOMIC_REDUCE = 0x10,
215	/*! To mark a 'barrier' directive in user code */
216	KMP_IDENT_BARRIER_EXPL = 0x20,
217	/*! To Mark implicit barriers. */
218	KMP_IDENT_BARRIER_IMPL = 0x0040,
219	KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
220	KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
221	KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
222
223	KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
224	KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
225
226	/*! To mark a static loop in OMPT callbacks */
227	KMP_IDENT_WORK_LOOP = 0x200,
228	/*! To mark a sections directive in OMPT callbacks */
229	KMP_IDENT_WORK_SECTIONS = 0x400,
230	/*! To mark a distribute construct in OMPT callbacks */
231	KMP_IDENT_WORK_DISTRIBUTE = 0x800,
232	/*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
233	not currently used. If one day we need more bits, then we can use
234	an invalid combination of hints to mean that another, larger field
235	should be used in a different flag. */
236	KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
237	KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
238	KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
239	KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
240	KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
241	KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
242	};
243
244	/*!
245	* The ident structure that describes a source location.
246	*/
247	typedef struct ident {
248	kmp_int32 reserved_1; /**< might be used in Fortran; see above */
249	kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
250	identifies this union member */
251	kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */
252	#if USE_ITT_BUILD
253	/* but currently used for storing region-specific ITT */
254	/* contextual information. */
255	#endif /* USE_ITT_BUILD */
256	kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */
257	char const *psource; /**< String describing the source location.
258	The string is composed of semi-colon separated fields
259	which describe the source file, the function and a pair
260	of line numbers that delimit the construct. */
261	// Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
262	kmp_int32 get_openmp_version() {
263	return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
264	}
265	} ident_t;
266	/*!
267	@}
268	*/
269
270	// Some forward declarations.
271	typedef union kmp_team kmp_team_t;
272	typedef struct kmp_taskdata kmp_taskdata_t;
273	typedef union kmp_task_team kmp_task_team_t;
274	typedef union kmp_team kmp_team_p;
275	typedef union kmp_info kmp_info_p;
276	typedef union kmp_root kmp_root_p;
277
278	template <bool C = false, bool S = true> class kmp_flag_32;
279	template <bool C = false, bool S = true> class kmp_flag_64;
280	template <bool C = false, bool S = true> class kmp_atomic_flag_64;
281	class kmp_flag_oncore;
282
283	#ifdef __cplusplus
284	extern "C" {
285	#endif
286
287	/* ------------------------------------------------------------------------ */
288
289	/* Pack two 32-bit signed integers into a 64-bit signed integer */
290	/* ToDo: Fix word ordering for big-endian machines. */
291	#define KMP_PACK_64(HIGH_32, LOW_32) \
292	((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
293
294	// Generic string manipulation macros. Assume that _x is of type char *
295	#define SKIP_WS(_x) \
296	{ \
297	while (*(_x) == ' ' || *(_x) == '\t') \
298	(_x)++; \
299	}
300	#define SKIP_DIGITS(_x) \
301	{ \
302	while (*(_x) >= '0' && *(_x) <= '9') \
303	(_x)++; \
304	}
305	#define SKIP_TOKEN(_x) \
306	{ \
307	while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
308	(*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
309	(_x)++; \
310	}
311	#define SKIP_TO(_x, _c) \
312	{ \
313	while (*(_x) != '\0' && *(_x) != (_c)) \
314	(_x)++; \
315	}
316
317	/* ------------------------------------------------------------------------ */
318
319	#define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
320	#define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
321
322	/* ------------------------------------------------------------------------ */
323	/* Enumeration types */
324
325	enum kmp_state_timer {
326	ts_stop,
327	ts_start,
328	ts_pause,
329
330	ts_last_state
331	};
332
333	enum dynamic_mode {
334	dynamic_default,
335	#ifdef USE_LOAD_BALANCE
336	dynamic_load_balance,
337	#endif /* USE_LOAD_BALANCE */
338	dynamic_random,
339	dynamic_thread_limit,
340	dynamic_max
341	};
342
343	/* external schedule constants, duplicate enum omp_sched in omp.h in order to
344	* not include it here */
345	#ifndef KMP_SCHED_TYPE_DEFINED
346	#define KMP_SCHED_TYPE_DEFINED
347	typedef enum kmp_sched {
348	kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
349	// Note: need to adjust __kmp_sch_map global array in case enum is changed
350	kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
351	kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
352	kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
353	kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
354	kmp_sched_upper_std = 5, // upper bound for standard schedules
355	kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
356	kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
357	#if KMP_STATIC_STEAL_ENABLED
358	kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
359	#endif
360	kmp_sched_upper,
361	kmp_sched_default = kmp_sched_static, // default scheduling
362	kmp_sched_monotonic = 0x80000000
363	} kmp_sched_t;
364	#endif
365
366	/*!
367	@ingroup WORK_SHARING
368	* Describes the loop schedule to be used for a parallel for loop.
369	*/
370	enum sched_type : kmp_int32 {
371	kmp_sch_lower = 32, /**< lower bound for unordered values */
372	kmp_sch_static_chunked = 33,
373	kmp_sch_static = 34, /**< static unspecialized */
374	kmp_sch_dynamic_chunked = 35,
375	kmp_sch_guided_chunked = 36, /**< guided unspecialized */
376	kmp_sch_runtime = 37,
377	kmp_sch_auto = 38, /**< auto */
378	kmp_sch_trapezoidal = 39,
379
380	/* accessible only through KMP_SCHEDULE environment variable */
381	kmp_sch_static_greedy = 40,
382	kmp_sch_static_balanced = 41,
383	/* accessible only through KMP_SCHEDULE environment variable */
384	kmp_sch_guided_iterative_chunked = 42,
385	kmp_sch_guided_analytical_chunked = 43,
386	/* accessible only through KMP_SCHEDULE environment variable */
387	kmp_sch_static_steal = 44,
388
389	/* static with chunk adjustment (e.g., simd) */
390	kmp_sch_static_balanced_chunked = 45,
391	kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
392	kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
393
394	/* accessible only through KMP_SCHEDULE environment variable */
395	kmp_sch_upper, /**< upper bound for unordered values */
396
397	kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
398	kmp_ord_static_chunked = 65,
399	kmp_ord_static = 66, /**< ordered static unspecialized */
400	kmp_ord_dynamic_chunked = 67,
401	kmp_ord_guided_chunked = 68,
402	kmp_ord_runtime = 69,
403	kmp_ord_auto = 70, /**< ordered auto */
404	kmp_ord_trapezoidal = 71,
405	kmp_ord_upper, /**< upper bound for ordered values */
406
407	/* Schedules for Distribute construct */
408	kmp_distribute_static_chunked = 91, /**< distribute static chunked */
409	kmp_distribute_static = 92, /**< distribute static unspecialized */
410
411	/* For the "nomerge" versions, kmp_dispatch_next*() will always return a
412	single iteration/chunk, even if the loop is serialized. For the schedule
413	types listed above, the entire iteration vector is returned if the loop is
414	serialized. This doesn't work for gcc/gcomp sections. */
415	kmp_nm_lower = 160, /**< lower bound for nomerge values */
416
417	kmp_nm_static_chunked =
418	(kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
419	kmp_nm_static = 162, /**< static unspecialized */
420	kmp_nm_dynamic_chunked = 163,
421	kmp_nm_guided_chunked = 164, /**< guided unspecialized */
422	kmp_nm_runtime = 165,
423	kmp_nm_auto = 166, /**< auto */
424	kmp_nm_trapezoidal = 167,
425
426	/* accessible only through KMP_SCHEDULE environment variable */
427	kmp_nm_static_greedy = 168,
428	kmp_nm_static_balanced = 169,
429	/* accessible only through KMP_SCHEDULE environment variable */
430	kmp_nm_guided_iterative_chunked = 170,
431	kmp_nm_guided_analytical_chunked = 171,
432	kmp_nm_static_steal =
433	172, /* accessible only through OMP_SCHEDULE environment variable */
434
435	kmp_nm_ord_static_chunked = 193,
436	kmp_nm_ord_static = 194, /**< ordered static unspecialized */
437	kmp_nm_ord_dynamic_chunked = 195,
438	kmp_nm_ord_guided_chunked = 196,
439	kmp_nm_ord_runtime = 197,
440	kmp_nm_ord_auto = 198, /**< auto */
441	kmp_nm_ord_trapezoidal = 199,
442	kmp_nm_upper, /**< upper bound for nomerge values */
443
444	/* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
445	we need to distinguish the three possible cases (no modifier, monotonic
446	modifier, nonmonotonic modifier), we need separate bits for each modifier.
447	The absence of monotonic does not imply nonmonotonic, especially since 4.5
448	says that the behaviour of the "no modifier" case is implementation defined
449	in 4.5, but will become "nonmonotonic" in 5.0.
450
451	Since we're passing a full 32 bit value, we can use a couple of high bits
452	for these flags; out of paranoia we avoid the sign bit.
453
454	These modifiers can be or-ed into non-static schedules by the compiler to
455	pass the additional information. They will be stripped early in the
456	processing in __kmp_dispatch_init when setting up schedules, so most of the
457	code won't ever see schedules with these bits set. */
458	kmp_sch_modifier_monotonic =
459	(1 << 29), /**< Set if the monotonic schedule modifier was present */
460	kmp_sch_modifier_nonmonotonic =
461	(1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
462
463	#define SCHEDULE_WITHOUT_MODIFIERS(s) \
464	(enum sched_type)( \
465	(s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
466	#define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
467	#define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
468	#define SCHEDULE_HAS_NO_MODIFIERS(s) \
469	(((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
470	#define SCHEDULE_GET_MODIFIERS(s) \
471	((enum sched_type)( \
472	(s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
473	#define SCHEDULE_SET_MODIFIERS(s, m) \
474	(s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
475	#define SCHEDULE_NONMONOTONIC 0
476	#define SCHEDULE_MONOTONIC 1
477
478	kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
479	};
480
481	// Apply modifiers on internal kind to standard kind
482	static inline void
483	__kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
484	enum sched_type internal_kind) {
485	if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
486	*kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
487	}
488	}
489
490	// Apply modifiers on standard kind to internal kind
491	static inline void
492	__kmp_sched_apply_mods_intkind(kmp_sched_t kind,
493	enum sched_type *internal_kind) {
494	if ((int)kind & (int)kmp_sched_monotonic) {
495	*internal_kind = (enum sched_type)((int)*internal_kind |
496	(int)kmp_sch_modifier_monotonic);
497	}
498	}
499
500	// Get standard schedule without modifiers
501	static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
502	return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
503	}
504
505	/* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
506	typedef union kmp_r_sched {
507	struct {
508	enum sched_type r_sched_type;
509	int chunk;
510	};
511	kmp_int64 sched;
512	} kmp_r_sched_t;
513
514	extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
515	// internal schedule types
516
517	enum library_type {
518	library_none,
519	library_serial,
520	library_turnaround,
521	library_throughput
522	};
523
524	#if KMP_MIC_SUPPORTED
525	enum mic_type { non_mic, mic1, mic2, mic3, dummy };
526	#endif
527
528	// OpenMP 3.1 - Nested num threads array
529	typedef struct kmp_nested_nthreads_t {
530	int *nth;
531	int size;
532	int used;
533	} kmp_nested_nthreads_t;
534
535	extern kmp_nested_nthreads_t __kmp_nested_nth;
536
537	/* -- fast reduction stuff ------------------------------------------------ */
538
539	#undef KMP_FAST_REDUCTION_BARRIER
540	#define KMP_FAST_REDUCTION_BARRIER 1
541
542	#undef KMP_FAST_REDUCTION_CORE_DUO
543	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
544	#define KMP_FAST_REDUCTION_CORE_DUO 1
545	#endif
546
547	enum _reduction_method {
548	reduction_method_not_defined = 0,
549	critical_reduce_block = (1 << 8),
550	atomic_reduce_block = (2 << 8),
551	tree_reduce_block = (3 << 8),
552	empty_reduce_block = (4 << 8)
553	};
554
555	// Description of the packed_reduction_method variable:
556	// The packed_reduction_method variable consists of two enum types variables
557	// that are packed together into 0-th byte and 1-st byte:
558	// 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
559	// barrier that will be used in fast reduction: bs_plain_barrier or
560	// bs_reduction_barrier
561	// 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
562	// be used in fast reduction;
563	// Reduction method is of 'enum _reduction_method' type and it's defined the way
564	// so that the bits of 0-th byte are empty, so no need to execute a shift
565	// instruction while packing/unpacking
566
567	#if KMP_FAST_REDUCTION_BARRIER
568	#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
569	((reduction_method) | (barrier_type))
570
571	#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
572	((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
573
574	#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
575	((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
576	#else
577	#define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
578	(reduction_method)
579
580	#define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
581	(packed_reduction_method)
582
583	#define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
584	#endif
585
586	#define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
587	((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
588	(which_reduction_block))
589
590	#if KMP_FAST_REDUCTION_BARRIER
591	#define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
592	(PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
593
594	#define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
595	(PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
596	#endif
597
598	typedef int PACKED_REDUCTION_METHOD_T;
599
600	/* -- end of fast reduction stuff ----------------------------------------- */
601
602	#if KMP_OS_WINDOWS
603	#define USE_CBLKDATA
604	#if KMP_MSVC_COMPAT
605	#pragma warning(push)
606	#pragma warning(disable : 271 310)
607	#endif
608	#include <windows.h>
609	#if KMP_MSVC_COMPAT
610	#pragma warning(pop)
611	#endif
612	#endif
613
614	#if KMP_OS_UNIX
615	#if !KMP_OS_WASI
616	#include <dlfcn.h>
617	#endif
618	#include <pthread.h>
619	#endif
620
621	enum kmp_hw_t : int {
622	KMP_HW_UNKNOWN = -1,
623	KMP_HW_SOCKET = 0,
624	KMP_HW_PROC_GROUP,
625	KMP_HW_NUMA,
626	KMP_HW_DIE,
627	KMP_HW_LLC,
628	KMP_HW_L3,
629	KMP_HW_TILE,
630	KMP_HW_MODULE,
631	KMP_HW_L2,
632	KMP_HW_L1,
633	KMP_HW_CORE,
634	KMP_HW_THREAD,
635	KMP_HW_LAST
636	};
637
638	typedef enum kmp_hw_core_type_t {
639	KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
640	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
641	KMP_HW_CORE_TYPE_ATOM = 0x20,
642	KMP_HW_CORE_TYPE_CORE = 0x40,
643	KMP_HW_MAX_NUM_CORE_TYPES = 3,
644	#else
645	KMP_HW_MAX_NUM_CORE_TYPES = 1,
646	#endif
647	} kmp_hw_core_type_t;
648
649	#define KMP_HW_MAX_NUM_CORE_EFFS 8
650
651	#define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
652	KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
653	#define KMP_ASSERT_VALID_HW_TYPE(type) \
654	KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
655
656	#define KMP_FOREACH_HW_TYPE(type) \
657	for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
658	type = (kmp_hw_t)((int)type + 1))
659
660	const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
661	const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
662	const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
663
664	/* Only Linux* OS and Windows* OS support thread affinity. */
665	#if KMP_AFFINITY_SUPPORTED
666
667	// GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
668	#if KMP_OS_WINDOWS
669	#if _MSC_VER < 1600 && KMP_MSVC_COMPAT
670	typedef struct GROUP_AFFINITY {
671	KAFFINITY Mask;
672	WORD Group;
673	WORD Reserved[3];
674	} GROUP_AFFINITY;
675	#endif /* _MSC_VER < 1600 */
676	#if KMP_GROUP_AFFINITY
677	extern int __kmp_num_proc_groups;
678	#else
679	static const int __kmp_num_proc_groups = 1;
680	#endif /* KMP_GROUP_AFFINITY */
681	typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
682	extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
683
684	typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
685	extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
686
687	typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
688	extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
689
690	typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
691	GROUP_AFFINITY *);
692	extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
693	#endif /* KMP_OS_WINDOWS */
694
695	#if KMP_USE_HWLOC && !defined(OMPD_SKIP_HWLOC)
696	extern hwloc_topology_t __kmp_hwloc_topology;
697	extern int __kmp_hwloc_error;
698	#endif
699
700	extern size_t __kmp_affin_mask_size;
701	#define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
702	#define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
703	#define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
704	#define KMP_CPU_SET_ITERATE(i, mask) \
705	for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
706	#define KMP_CPU_SET(i, mask) (mask)->set(i)
707	#define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
708	#define KMP_CPU_CLR(i, mask) (mask)->clear(i)
709	#define KMP_CPU_ZERO(mask) (mask)->zero()
710	#define KMP_CPU_ISEMPTY(mask) (mask)->empty()
711	#define KMP_CPU_COPY(dest, src) (dest)->copy(src)
712	#define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
713	#define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
714	#define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
715	#define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
716	#define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
717	#define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
718	#define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
719	#define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
720	#define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
721	#define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
722	#define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
723	#define KMP_CPU_ALLOC_ARRAY(arr, n) \
724	(arr = __kmp_affinity_dispatch->allocate_mask_array(n))
725	#define KMP_CPU_FREE_ARRAY(arr, n) \
726	__kmp_affinity_dispatch->deallocate_mask_array(arr)
727	#define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
728	#define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
729	#define __kmp_get_system_affinity(mask, abort_bool) \
730	(mask)->get_system_affinity(abort_bool)
731	#define __kmp_set_system_affinity(mask, abort_bool) \
732	(mask)->set_system_affinity(abort_bool)
733	#define __kmp_get_proc_group(mask) (mask)->get_proc_group()
734
735	class KMPAffinity {
736	public:
737	class Mask {
738	public:
739	void *operator new(size_t n);
740	void operator delete(void *p);
741	void *operator new[](size_t n);
742	void operator delete[](void *p);
743	virtual ~Mask() {}
744	// Set bit i to 1
745	virtual void set(int i) {}
746	// Return bit i
747	virtual bool is_set(int i) const { return false; }
748	// Set bit i to 0
749	virtual void clear(int i) {}
750	// Zero out entire mask
751	virtual void zero() {}
752	// Check whether mask is empty
753	virtual bool empty() const { return true; }
754	// Copy src into this mask
755	virtual void copy(const Mask *src) {}
756	// this &= rhs
757	virtual void bitwise_and(const Mask *rhs) {}
758	// this |= rhs
759	virtual void bitwise_or(const Mask *rhs) {}
760	// this = ~this
761	virtual void bitwise_not() {}
762	// this == rhs
763	virtual bool is_equal(const Mask *rhs) const { return false; }
764	// API for iterating over an affinity mask
765	// for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
766	virtual int begin() const { return 0; }
767	virtual int end() const { return 0; }
768	virtual int next(int previous) const { return 0; }
769	#if KMP_OS_WINDOWS
770	virtual int set_process_affinity(bool abort_on_error) const { return -1; }
771	#endif
772	// Set the system's affinity to this affinity mask's value
773	virtual int set_system_affinity(bool abort_on_error) const { return -1; }
774	// Set this affinity mask to the current system affinity
775	virtual int get_system_affinity(bool abort_on_error) { return -1; }
776	// Only 1 DWORD in the mask should have any procs set.
777	// Return the appropriate index, or -1 for an invalid mask.
778	virtual int get_proc_group() const { return -1; }
779	int get_max_cpu() const {
780	int cpu;
781	int max_cpu = -1;
782	KMP_CPU_SET_ITERATE(cpu, this) {
783	if (cpu > max_cpu)
784	max_cpu = cpu;
785	}
786	return max_cpu;
787	}
788	};
789	void *operator new(size_t n);
790	void operator delete(void *p);
791	// Need virtual destructor
792	virtual ~KMPAffinity() = default;
793	// Determine if affinity is capable
794	virtual void determine_capable(const char *env_var) {}
795	// Bind the current thread to os proc
796	virtual void bind_thread(int proc) {}
797	// Factory functions to allocate/deallocate a mask
798	virtual Mask *allocate_mask() { return nullptr; }
799	virtual void deallocate_mask(Mask *m) {}
800	virtual Mask *allocate_mask_array(int num) { return nullptr; }
801	virtual void deallocate_mask_array(Mask *m) {}
802	virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
803	static void pick_api();
804	static void destroy_api();
805	enum api_type {
806	NATIVE_OS
807	#if KMP_USE_HWLOC
808	,
809	HWLOC
810	#endif
811	};
812	virtual api_type get_api_type() const {
813	KMP_ASSERT(0);
814	return NATIVE_OS;
815	}
816
817	private:
818	static bool picked_api;
819	};
820
821	typedef KMPAffinity::Mask kmp_affin_mask_t;
822	extern KMPAffinity *__kmp_affinity_dispatch;
823
824	#if !KMP_OS_AIX
825	class kmp_affinity_raii_t {
826	kmp_affin_mask_t *mask;
827	bool restored;
828
829	public:
830	kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
831	: mask(nullptr), restored(false) {
832	if (KMP_AFFINITY_CAPABLE()) {
833	KMP_CPU_ALLOC(mask);
834	KMP_ASSERT(mask != NULL);
835	__kmp_get_system_affinity(mask, /*abort_on_error=*/true);
836	if (new_mask)
837	__kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
838	}
839	}
840	void restore() {
841	if (mask && KMP_AFFINITY_CAPABLE() && !restored) {
842	__kmp_set_system_affinity(mask, /*abort_on_error=*/true);
843	KMP_CPU_FREE(mask);
844	}
845	restored = true;
846	}
847	~kmp_affinity_raii_t() { restore(); }
848	};
849	#endif // !KMP_OS_AIX
850
851	// Declare local char buffers with this size for printing debug and info
852	// messages, using __kmp_affinity_print_mask().
853	#define KMP_AFFIN_MASK_PRINT_LEN 1024
854
855	enum affinity_type {
856	affinity_none = 0,
857	affinity_physical,
858	affinity_logical,
859	affinity_compact,
860	affinity_scatter,
861	affinity_explicit,
862	affinity_balanced,
863	affinity_disabled, // not used outsize the env var parser
864	affinity_default
865	};
866
867	enum affinity_top_method {
868	affinity_top_method_all = 0, // try all (supported) methods, in order
869	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
870	affinity_top_method_apicid,
871	affinity_top_method_x2apicid,
872	affinity_top_method_x2apicid_1f,
873	#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
874	affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
875	#if KMP_GROUP_AFFINITY
876	affinity_top_method_group,
877	#endif /* KMP_GROUP_AFFINITY */
878	affinity_top_method_flat,
879	#if KMP_USE_HWLOC
880	affinity_top_method_hwloc,
881	#endif
882	affinity_top_method_default
883	};
884
885	#define affinity_respect_mask_default (2)
886
887	typedef struct kmp_affinity_flags_t {
888	unsigned dups : 1;
889	unsigned verbose : 1;
890	unsigned warnings : 1;
891	unsigned respect : 2;
892	unsigned reset : 1;
893	unsigned initialized : 1;
894	unsigned core_types_gran : 1;
895	unsigned core_effs_gran : 1;
896	unsigned omp_places : 1;
897	unsigned reserved : 22;
898	} kmp_affinity_flags_t;
899	KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
900
901	typedef struct kmp_affinity_ids_t {
902	int os_id;
903	int ids[KMP_HW_LAST];
904	} kmp_affinity_ids_t;
905
906	typedef struct kmp_affinity_attrs_t {
907	int core_type : 8;
908	int core_eff : 8;
909	unsigned valid : 1;
910	unsigned reserved : 15;
911	} kmp_affinity_attrs_t;
912	#define KMP_AFFINITY_ATTRS_UNKNOWN \
913	{ KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
914
915	typedef struct kmp_affinity_t {
916	char *proclist;
917	enum affinity_type type;
918	kmp_hw_t gran;
919	int gran_levels;
920	kmp_affinity_attrs_t core_attr_gran;
921	int compact;
922	int offset;
923	kmp_affinity_flags_t flags;
924	unsigned num_masks;
925	kmp_affin_mask_t *masks;
926	kmp_affinity_ids_t *ids;
927	kmp_affinity_attrs_t *attrs;
928	unsigned num_os_id_masks;
929	kmp_affin_mask_t *os_id_masks;
930	const char *env_var;
931	} kmp_affinity_t;
932
933	#define KMP_AFFINITY_INIT(env) \
934	{ \
935	nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
936	0, 0, \
937	{TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE, \
938	FALSE, FALSE, FALSE}, \
939	0, nullptr, nullptr, nullptr, 0, nullptr, env \
940	}
941
942	extern enum affinity_top_method __kmp_affinity_top_method;
943	extern kmp_affinity_t __kmp_affinity;
944	extern kmp_affinity_t __kmp_hh_affinity;
945	extern kmp_affinity_t *__kmp_affinities[2];
946
947	extern void __kmp_affinity_bind_thread(int which);
948
949	extern kmp_affin_mask_t *__kmp_affin_fullMask;
950	extern kmp_affin_mask_t *__kmp_affin_origMask;
951	extern char *__kmp_cpuinfo_file;
952
953	#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
954	extern int __kmp_first_osid_with_ecore;
955	#endif
956
957	#endif /* KMP_AFFINITY_SUPPORTED */
958
959	// This needs to be kept in sync with the values in omp.h !!!
960	typedef enum kmp_proc_bind_t {
961	proc_bind_false = 0,
962	proc_bind_true,
963	proc_bind_primary,
964	proc_bind_close,
965	proc_bind_spread,
966	proc_bind_intel, // use KMP_AFFINITY interface
967	proc_bind_default
968	} kmp_proc_bind_t;
969
970	typedef struct kmp_nested_proc_bind_t {
971	kmp_proc_bind_t *bind_types;
972	int size;
973	int used;
974	} kmp_nested_proc_bind_t;
975
976	extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
977	extern kmp_proc_bind_t __kmp_teams_proc_bind;
978
979	extern int __kmp_display_affinity;
980	extern char *__kmp_affinity_format;
981	static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
982	#if OMPT_SUPPORT
983	extern int __kmp_tool;
984	extern char *__kmp_tool_libraries;
985	#endif // OMPT_SUPPORT
986
987	#if KMP_AFFINITY_SUPPORTED
988	#define KMP_PLACE_ALL (-1)
989	#define KMP_PLACE_UNDEFINED (-2)
990	// Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
991	#define KMP_AFFINITY_NON_PROC_BIND \
992	((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
993	__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
994	(__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
995	#endif /* KMP_AFFINITY_SUPPORTED */
996
997	extern int __kmp_affinity_num_places;
998
999	typedef enum kmp_cancel_kind_t {
1000	cancel_noreq = 0,
1001	cancel_parallel = 1,
1002	cancel_loop = 2,
1003	cancel_sections = 3,
1004	cancel_taskgroup = 4
1005	} kmp_cancel_kind_t;
1006
1007	// KMP_HW_SUBSET support:
1008	typedef struct kmp_hws_item {
1009	int num;
1010	int offset;
1011	} kmp_hws_item_t;
1012
1013	extern kmp_hws_item_t __kmp_hws_socket;
1014	extern kmp_hws_item_t __kmp_hws_die;
1015	extern kmp_hws_item_t __kmp_hws_node;
1016	extern kmp_hws_item_t __kmp_hws_tile;
1017	extern kmp_hws_item_t __kmp_hws_core;
1018	extern kmp_hws_item_t __kmp_hws_proc;
1019	extern int __kmp_hws_requested;
1020	extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1021
1022	/* ------------------------------------------------------------------------ */
1023
1024	#define KMP_PAD(type, sz) \
1025	(sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1026
1027	// We need to avoid using -1 as a GTID as +1 is added to the gtid
1028	// when storing it in a lock, and the value 0 is reserved.
1029	#define KMP_GTID_DNE (-2) /* Does not exist */
1030	#define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1031	#define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1032	#define KMP_GTID_UNKNOWN (-5) /* Is not known */
1033	#define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1034
1035	/* OpenMP 5.0 Memory Management support */
1036
1037	#ifndef __OMP_H
1038	// Duplicate type definitions from omp.h
1039	typedef uintptr_t omp_uintptr_t;
1040
1041	typedef enum {
1042	omp_atk_sync_hint = 1,
1043	omp_atk_alignment = 2,
1044	omp_atk_access = 3,
1045	omp_atk_pool_size = 4,
1046	omp_atk_fallback = 5,
1047	omp_atk_fb_data = 6,
1048	omp_atk_pinned = 7,
1049	omp_atk_partition = 8,
1050	omp_atk_pin_device = 9,
1051	omp_atk_preferred_device = 10,
1052	omp_atk_device_access = 11,
1053	omp_atk_target_access = 12,
1054	omp_atk_atomic_scope = 13,
1055	omp_atk_part_size = 14
1056	} omp_alloctrait_key_t;
1057
1058	typedef enum {
1059	omp_atv_false = 0,
1060	omp_atv_true = 1,
1061	omp_atv_contended = 3,
1062	omp_atv_uncontended = 4,
1063	omp_atv_serialized = 5,
1064	omp_atv_sequential = omp_atv_serialized, // (deprecated)
1065	omp_atv_private = 6,
1066	omp_atv_device = 7,
1067	omp_atv_thread = 8,
1068	omp_atv_pteam = 9,
1069	omp_atv_cgroup = 10,
1070	omp_atv_default_mem_fb = 11,
1071	omp_atv_null_fb = 12,
1072	omp_atv_abort_fb = 13,
1073	omp_atv_allocator_fb = 14,
1074	omp_atv_environment = 15,
1075	omp_atv_nearest = 16,
1076	omp_atv_blocked = 17,
1077	omp_atv_interleaved = 18,
1078	omp_atv_all = 19,
1079	omp_atv_single = 20,
1080	omp_atv_multiple = 21,
1081	omp_atv_memspace = 22
1082	} omp_alloctrait_value_t;
1083	#define omp_atv_default ((omp_uintptr_t)-1)
1084
1085	typedef void *omp_memspace_handle_t;
1086	extern omp_memspace_handle_t const omp_null_mem_space;
1087	extern omp_memspace_handle_t const omp_default_mem_space;
1088	extern omp_memspace_handle_t const omp_large_cap_mem_space;
1089	extern omp_memspace_handle_t const omp_const_mem_space;
1090	extern omp_memspace_handle_t const omp_high_bw_mem_space;
1091	extern omp_memspace_handle_t const omp_low_lat_mem_space;
1092	extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1093	extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1094	extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1095	extern omp_memspace_handle_t const kmp_max_mem_space;
1096
1097	typedef struct {
1098	omp_alloctrait_key_t key;
1099	omp_uintptr_t value;
1100	} omp_alloctrait_t;
1101
1102	typedef void *omp_allocator_handle_t;
1103	extern omp_allocator_handle_t const omp_null_allocator;
1104	extern omp_allocator_handle_t const omp_default_mem_alloc;
1105	extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1106	extern omp_allocator_handle_t const omp_const_mem_alloc;
1107	extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1108	extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1109	extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1110	extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1111	extern omp_allocator_handle_t const omp_thread_mem_alloc;
1112	extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1113	extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1114	extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1115	extern omp_allocator_handle_t const kmp_max_mem_alloc;
1116	extern omp_allocator_handle_t __kmp_def_allocator;
1117
1118	// end of duplicate type definitions from omp.h
1119	#endif
1120
1121	extern int __kmp_memkind_available;
1122	extern bool __kmp_hwloc_available;
1123
1124	/// Memory space informaition is shared with offload runtime.
1125	typedef struct kmp_memspace_t {
1126	omp_memspace_handle_t memspace; // predefined input memory space
1127	int num_resources = 0; // number of available resources
1128	int *resources = nullptr; // available resources
1129	kmp_memspace_t *next = nullptr; // next memory space handle
1130	} kmp_memspace_t;
1131
1132	/// Memory allocator information is shared with offload runtime.
1133	typedef struct kmp_allocator_t {
1134	omp_memspace_handle_t memspace;
1135	void **memkind; // pointer to memkind
1136	size_t alignment;
1137	omp_alloctrait_value_t fb;
1138	kmp_allocator_t *fb_data;
1139	kmp_uint64 pool_size;
1140	kmp_uint64 pool_used;
1141	bool pinned;
1142	omp_alloctrait_value_t partition;
1143	int pin_device;
1144	int preferred_device;
1145	omp_alloctrait_value_t target_access;
1146	omp_alloctrait_value_t atomic_scope;
1147	size_t part_size;
1148	#if KMP_USE_HWLOC
1149	omp_alloctrait_value_t membind;
1150	#endif
1151	} kmp_allocator_t;
1152
1153	extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1154	omp_memspace_handle_t,
1155	int ntraits,
1156	omp_alloctrait_t traits[]);
1157	extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1158	extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1159	extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1160	// external interfaces, may be used by compiler
1161	extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1162	extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1163	omp_allocator_handle_t al);
1164	extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1165	omp_allocator_handle_t al);
1166	extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1167	omp_allocator_handle_t al,
1168	omp_allocator_handle_t free_al);
1169	extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1170	// internal interfaces, contain real implementation
1171	extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1172	omp_allocator_handle_t al);
1173	extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1174	omp_allocator_handle_t al);
1175	extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1176	omp_allocator_handle_t al,
1177	omp_allocator_handle_t free_al);
1178	extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1179
1180	extern void __kmp_init_memkind();
1181	extern void __kmp_fini_memkind();
1182	extern void __kmp_init_target_mem();
1183	extern void __kmp_fini_target_mem();
1184
1185	// OpenMP 6.0 (TR11) Memory Management support
1186	extern omp_memspace_handle_t __kmp_get_devices_memspace(int ndevs,
1187	const int *devs,
1188	omp_memspace_handle_t,
1189	int host);
1190	extern omp_allocator_handle_t __kmp_get_devices_allocator(int ndevs,
1191	const int *devs,
1192	omp_memspace_handle_t,
1193	int host);
1194	extern int __kmp_get_memspace_num_resources(omp_memspace_handle_t memspace);
1195	extern omp_memspace_handle_t
1196	__kmp_get_submemspace(omp_memspace_handle_t memspace, int num_resources,
1197	int *resources);
1198
1199	/* ------------------------------------------------------------------------ */
1200
1201	#if ENABLE_LIBOMPTARGET
1202	extern void __kmp_init_target_task();
1203	#endif
1204
1205	/* ------------------------------------------------------------------------ */
1206
1207	#define KMP_UINT64_MAX \
1208	(~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1209
1210	#define KMP_MIN_NTH 1
1211
1212	#ifndef KMP_MAX_NTH
1213	#if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1214	#define KMP_MAX_NTH PTHREAD_THREADS_MAX
1215	#else
1216	#ifdef __ve__
1217	// VE's pthread supports only up to 64 threads per a VE process.
1218	// Please check p. 14 of following documentation for more details.
1219	// https://sxauroratsubasa.sakura.ne.jp/documents/veos/en/VEOS_high_level_design.pdf
1220	#define KMP_MAX_NTH 64
1221	#else
1222	#define KMP_MAX_NTH INT_MAX
1223	#endif
1224	#endif
1225	#endif /* KMP_MAX_NTH */
1226
1227	#ifdef PTHREAD_STACK_MIN
1228	#define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1229	#else
1230	#define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1231	#endif
1232
1233	#if KMP_OS_AIX && KMP_ARCH_PPC
1234	#define KMP_MAX_STKSIZE 0x10000000 /* 256Mb max size on 32-bit AIX */
1235	#else
1236	#define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1237	#endif
1238
1239	#if KMP_ARCH_X86
1240	#define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1241	#elif KMP_ARCH_X86_64
1242	#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1243	#define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1244	#elif KMP_ARCH_VE
1245	// Minimum stack size for pthread for VE is 4MB.
1246	// https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1247	#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1248	#elif KMP_OS_AIX
1249	// The default stack size for worker threads on AIX is 4MB.
1250	#define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1251	#else
1252	#define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1253	#endif
1254
1255	#define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1256	#define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1257	#define KMP_MAX_MALLOC_POOL_INCR \
1258	(~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1259
1260	#define KMP_MIN_STKOFFSET (0)
1261	#define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1262	#if KMP_OS_DARWIN
1263	#define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1264	#else
1265	#define KMP_DEFAULT_STKOFFSET CACHE_LINE
1266	#endif
1267
1268	#define KMP_MIN_STKPADDING (0)
1269	#define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1270
1271	#define KMP_BLOCKTIME_MULTIPLIER \
1272	(1000000) /* number of blocktime units per second */
1273	#define KMP_MIN_BLOCKTIME (0)
1274	#define KMP_MAX_BLOCKTIME \
1275	(INT_MAX) /* Must be this for "infinite" setting the work */
1276
1277	/* __kmp_blocktime is in microseconds */
1278	#define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1279
1280	#if KMP_USE_MONITOR
1281	#define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1282	#define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1283	#define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1284
1285	/* Calculate new number of monitor wakeups for a specific block time based on
1286	previous monitor_wakeups. Only allow increasing number of wakeups */
1287	#define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1288	(((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1289	: ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1290	: ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1291	? (monitor_wakeups) \
1292	: (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1293
1294	/* Calculate number of intervals for a specific block time based on
1295	monitor_wakeups */
1296	#define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1297	(((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1298	(KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1299	#else
1300	#define KMP_BLOCKTIME(team, tid) \
1301	(get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1302	#if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1303	// HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1304	extern kmp_uint64 __kmp_ticks_per_msec;
1305	extern kmp_uint64 __kmp_ticks_per_usec;
1306	#if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1307	#define KMP_NOW() ((kmp_uint64)_rdtsc())
1308	#else
1309	#define KMP_NOW() __kmp_hardware_timestamp()
1310	#endif
1311	#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1312	((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1313	#define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1314	#else
1315	// System time is retrieved sporadically while blocking.
1316	extern kmp_uint64 __kmp_now_nsec();
1317	#define KMP_NOW() __kmp_now_nsec()
1318	#define KMP_BLOCKTIME_INTERVAL(team, tid) \
1319	((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1320	#define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1321	#endif
1322	#endif // KMP_USE_MONITOR
1323
1324	#define KMP_MIN_STATSCOLS 40
1325	#define KMP_MAX_STATSCOLS 4096
1326	#define KMP_DEFAULT_STATSCOLS 80
1327
1328	#define KMP_MIN_INTERVAL 0
1329	#define KMP_MAX_INTERVAL (INT_MAX - 1)
1330	#define KMP_DEFAULT_INTERVAL 0
1331
1332	#define KMP_MIN_CHUNK 1
1333	#define KMP_MAX_CHUNK (INT_MAX - 1)
1334	#define KMP_DEFAULT_CHUNK 1
1335
1336	#define KMP_MIN_DISP_NUM_BUFF 1
1337	#define KMP_DFLT_DISP_NUM_BUFF 7
1338	#define KMP_MAX_DISP_NUM_BUFF 4096
1339
1340	#define KMP_MAX_ORDERED 8
1341
1342	#define KMP_MAX_FIELDS 32
1343
1344	#define KMP_MAX_BRANCH_BITS 31
1345
1346	#define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1347
1348	#define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1349
1350	#define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1351
1352	/* Minimum number of threads before switch to TLS gtid (experimentally
1353	determined) */
1354	/* josh TODO: what about OS X* tuning? */
1355	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1356	#define KMP_TLS_GTID_MIN 5
1357	#else
1358	#define KMP_TLS_GTID_MIN INT_MAX
1359	#endif
1360
1361	#define KMP_MASTER_TID(tid) (0 == (tid))
1362	#define KMP_WORKER_TID(tid) (0 != (tid))
1363
1364	#define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1365	#define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1366	#define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1367
1368	#ifndef TRUE
1369	#define FALSE 0
1370	#define TRUE (!FALSE)
1371	#endif
1372
1373	/* NOTE: all of the following constants must be even */
1374
1375	#if KMP_OS_WINDOWS
1376	#define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1377	#define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1378	#elif KMP_OS_LINUX
1379	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1380	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1381	#elif KMP_OS_DARWIN
1382	/* TODO: tune for KMP_OS_DARWIN */
1383	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1384	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1385	#elif KMP_OS_DRAGONFLY
1386	/* TODO: tune for KMP_OS_DRAGONFLY */
1387	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1388	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1389	#elif KMP_OS_FREEBSD
1390	/* TODO: tune for KMP_OS_FREEBSD */
1391	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1392	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1393	#elif KMP_OS_NETBSD
1394	/* TODO: tune for KMP_OS_NETBSD */
1395	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1396	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1397	#elif KMP_OS_OPENBSD
1398	/* TODO: tune for KMP_OS_OPENBSD */
1399	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1400	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1401	#elif KMP_OS_HAIKU
1402	/* TODO: tune for KMP_OS_HAIKU */
1403	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1404	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1405	#elif KMP_OS_HURD
1406	/* TODO: tune for KMP_OS_HURD */
1407	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1408	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1409	#elif KMP_OS_SOLARIS
1410	/* TODO: tune for KMP_OS_SOLARIS */
1411	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1412	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1413	#elif KMP_OS_WASI
1414	/* TODO: tune for KMP_OS_WASI */
1415	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1416	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1417	#elif KMP_OS_AIX
1418	/* TODO: tune for KMP_OS_AIX */
1419	#define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1420	#define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1421	#endif
1422
1423	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
1424	typedef struct kmp_cpuid {
1425	kmp_uint32 eax;
1426	kmp_uint32 ebx;
1427	kmp_uint32 ecx;
1428	kmp_uint32 edx;
1429	} kmp_cpuid_t;
1430
1431	typedef struct kmp_cpuinfo_flags_t {
1432	unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1433	unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1434	unsigned hybrid : 1;
1435	unsigned reserved : 29; // Ensure size of 32 bits
1436	} kmp_cpuinfo_flags_t;
1437
1438	typedef struct kmp_cpuinfo {
1439	int initialized; // If 0, other fields are not initialized.
1440	int signature; // CPUID(1).EAX
1441	int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1442	int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1443	// Model << 4 ) + Model)
1444	int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1445	kmp_cpuinfo_flags_t flags;
1446	int apic_id;
1447	kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1448	char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1449	} kmp_cpuinfo_t;
1450
1451	extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1452
1453	#if KMP_OS_UNIX
1454	// subleaf is only needed for cache and topology discovery and can be set to
1455	// zero in most cases
1456	static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1457	__asm__ __volatile__("cpuid"
1458	: "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1459	: "a"(leaf), "c"(subleaf));
1460	}
1461	// Load p into FPU control word
1462	static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1463	__asm__ __volatile__("fldcw %0" : : "m"(*p));
1464	}
1465	// Store FPU control word into p
1466	static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1467	__asm__ __volatile__("fstcw %0" : "=m"(*p));
1468	}
1469	static inline void __kmp_clear_x87_fpu_status_word() {
1470	#if KMP_MIC
1471	// 32-bit protected mode x87 FPU state
1472	struct x87_fpu_state {
1473	unsigned cw;
1474	unsigned sw;
1475	unsigned tw;
1476	unsigned fip;
1477	unsigned fips;
1478	unsigned fdp;
1479	unsigned fds;
1480	};
1481	struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1482	__asm__ __volatile__("fstenv %0\n\t" // store FP env
1483	"andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1484	"fldenv %0\n\t" // load FP env back
1485	: "+m"(fpu_state), "+m"(fpu_state.sw));
1486	#else
1487	__asm__ __volatile__("fnclex");
1488	#endif // KMP_MIC
1489	}
1490	#if __SSE__
1491	static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(i: *p); }
1492	static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1493	#else
1494	static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1495	static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1496	#endif
1497	#else
1498	// Windows still has these as external functions in assembly file
1499	extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1500	extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1501	extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1502	extern void __kmp_clear_x87_fpu_status_word();
1503	static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1504	static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1505	#endif // KMP_OS_UNIX
1506
1507	#define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1508
1509	// User-level Monitor/Mwait
1510	#if KMP_HAVE_UMWAIT
1511	// We always try for UMWAIT first
1512	#if KMP_HAVE_WAITPKG_INTRINSICS
1513	#if KMP_HAVE_IMMINTRIN_H
1514	#include <immintrin.h>
1515	#elif KMP_HAVE_INTRIN_H
1516	#include <intrin.h>
1517	#endif
1518	#endif // KMP_HAVE_WAITPKG_INTRINSICS
1519
1520	KMP_ATTRIBUTE_TARGET_WAITPKG
1521	static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1522	#if !KMP_HAVE_WAITPKG_INTRINSICS
1523	uint32_t timeHi = uint32_t(counter >> 32);
1524	uint32_t timeLo = uint32_t(counter & 0xffffffff);
1525	char flag;
1526	__asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1527	"setb %0"
1528	// The "=q" restraint means any register accessible as rl
1529	// in 32-bit mode: a, b, c, and d;
1530	// in 64-bit mode: any integer register
1531	: "=q"(flag)
1532	: "a"(timeLo), "d"(timeHi), "c"(hint)
1533	:);
1534	return flag;
1535	#else
1536	return _tpause(control: hint, counter: counter);
1537	#endif
1538	}
1539	KMP_ATTRIBUTE_TARGET_WAITPKG
1540	static inline void __kmp_umonitor(void *cacheline) {
1541	#if !KMP_HAVE_WAITPKG_INTRINSICS
1542	__asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1543	:
1544	: "a"(cacheline)
1545	:);
1546	#else
1547	_umonitor(address: cacheline);
1548	#endif
1549	}
1550	KMP_ATTRIBUTE_TARGET_WAITPKG
1551	static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1552	#if !KMP_HAVE_WAITPKG_INTRINSICS
1553	uint32_t timeHi = uint32_t(counter >> 32);
1554	uint32_t timeLo = uint32_t(counter & 0xffffffff);
1555	char flag;
1556	__asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1557	"setb %0"
1558	// The "=q" restraint means any register accessible as rl
1559	// in 32-bit mode: a, b, c, and d;
1560	// in 64-bit mode: any integer register
1561	: "=q"(flag)
1562	: "a"(timeLo), "d"(timeHi), "c"(hint)
1563	:);
1564	return flag;
1565	#else
1566	return _umwait(control: hint, counter: counter);
1567	#endif
1568	}
1569	#elif KMP_HAVE_MWAIT
1570	#if KMP_OS_UNIX
1571	#include <pmmintrin.h>
1572	#else
1573	#include <intrin.h>
1574	#endif
1575	#if KMP_OS_UNIX
1576	__attribute__((target("sse3")))
1577	#endif
1578	static inline void
1579	__kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1580	_mm_monitor(cacheline, extensions, hints);
1581	}
1582	#if KMP_OS_UNIX
1583	__attribute__((target("sse3")))
1584	#endif
1585	static inline void
1586	__kmp_mm_mwait(unsigned extensions, unsigned hints) {
1587	_mm_mwait(extensions, hints);
1588	}
1589	#endif // KMP_HAVE_UMWAIT
1590
1591	#if KMP_ARCH_X86
1592	extern void __kmp_x86_pause(void);
1593	#elif KMP_MIC
1594	// Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1595	// regression after removal of extra PAUSE from spin loops. Changing
1596	// the delay from 100 to 300 showed even better performance than double PAUSE
1597	// on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1598	static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1599	#else
1600	static inline void __kmp_x86_pause(void) { _mm_pause(); }
1601	#endif
1602	#define KMP_CPU_PAUSE() __kmp_x86_pause()
1603	#elif KMP_ARCH_PPC64
1604	#define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1605	#define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1606	#define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1607	#define KMP_CPU_PAUSE() \
1608	do { \
1609	KMP_PPC64_PRI_LOW(); \
1610	KMP_PPC64_PRI_MED(); \
1611	KMP_PPC64_PRI_LOC_MB(); \
1612	} while (0)
1613	#else
1614	#define KMP_CPU_PAUSE() /* nothing to do */
1615	#endif
1616
1617	#define KMP_INIT_YIELD(count) \
1618	{ (count) = __kmp_yield_init; }
1619
1620	#define KMP_INIT_BACKOFF(time) \
1621	{ (time) = __kmp_pause_init; }
1622
1623	#define KMP_OVERSUBSCRIBED \
1624	(TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1625
1626	#define KMP_TRY_YIELD \
1627	((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1628
1629	#define KMP_TRY_YIELD_OVERSUB \
1630	((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1631
1632	#define KMP_YIELD(cond) \
1633	{ \
1634	KMP_CPU_PAUSE(); \
1635	if ((cond) && (KMP_TRY_YIELD)) \
1636	__kmp_yield(); \
1637	}
1638
1639	#define KMP_YIELD_OVERSUB() \
1640	{ \
1641	KMP_CPU_PAUSE(); \
1642	if ((KMP_TRY_YIELD_OVERSUB)) \
1643	__kmp_yield(); \
1644	}
1645
1646	// Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1647	// there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1648	#define KMP_YIELD_SPIN(count) \
1649	{ \
1650	KMP_CPU_PAUSE(); \
1651	if (KMP_TRY_YIELD) { \
1652	(count) -= 2; \
1653	if (!(count)) { \
1654	__kmp_yield(); \
1655	(count) = __kmp_yield_next; \
1656	} \
1657	} \
1658	}
1659
1660	// If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1661	// (C0.2) state, which improves performance of other SMT threads on the same
1662	// core, otherwise, use the fast (C0.1) default state, or whatever the user has
1663	// requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1664	// available, fall back to the regular CPU pause and yield combination.
1665	#if KMP_HAVE_UMWAIT
1666	#define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1667	#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1668	{ \
1669	if (__kmp_tpause_enabled) { \
1670	if (KMP_OVERSUBSCRIBED) { \
1671	__kmp_tpause(0, (time)); \
1672	} else { \
1673	__kmp_tpause(__kmp_tpause_hint, (time)); \
1674	} \
1675	(time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1676	} else { \
1677	KMP_CPU_PAUSE(); \
1678	if ((KMP_TRY_YIELD_OVERSUB)) { \
1679	__kmp_yield(); \
1680	} else if (__kmp_use_yield == 1) { \
1681	(count) -= 2; \
1682	if (!(count)) { \
1683	__kmp_yield(); \
1684	(count) = __kmp_yield_next; \
1685	} \
1686	} \
1687	} \
1688	}
1689	#else
1690	#define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1691	{ \
1692	KMP_CPU_PAUSE(); \
1693	if ((KMP_TRY_YIELD_OVERSUB)) \
1694	__kmp_yield(); \
1695	else if (__kmp_use_yield == 1) { \
1696	(count) -= 2; \
1697	if (!(count)) { \
1698	__kmp_yield(); \
1699	(count) = __kmp_yield_next; \
1700	} \
1701	} \
1702	}
1703	#endif // KMP_HAVE_UMWAIT
1704
1705	/* ------------------------------------------------------------------------ */
1706	/* Support datatypes for the orphaned construct nesting checks. */
1707	/* ------------------------------------------------------------------------ */
1708
1709	/* When adding to this enum, add its corresponding string in cons_text_c[]
1710	* array in kmp_error.cpp */
1711	enum cons_type {
1712	ct_none,
1713	ct_parallel,
1714	ct_pdo,
1715	ct_pdo_ordered,
1716	ct_psections,
1717	ct_psingle,
1718	ct_critical,
1719	ct_ordered_in_parallel,
1720	ct_ordered_in_pdo,
1721	ct_master,
1722	ct_reduce,
1723	ct_barrier,
1724	ct_masked
1725	};
1726
1727	#define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1728
1729	struct cons_data {
1730	ident_t const *ident;
1731	enum cons_type type;
1732	int prev;
1733	kmp_user_lock_p
1734	name; /* address exclusively for critical section name comparison */
1735	};
1736
1737	struct cons_header {
1738	int p_top, w_top, s_top;
1739	int stack_size, stack_top;
1740	struct cons_data *stack_data;
1741	};
1742
1743	struct kmp_region_info {
1744	char *text;
1745	int offset[KMP_MAX_FIELDS];
1746	int length[KMP_MAX_FIELDS];
1747	};
1748
1749	/* ---------------------------------------------------------------------- */
1750	/* ---------------------------------------------------------------------- */
1751
1752	#if KMP_OS_WINDOWS
1753	typedef HANDLE kmp_thread_t;
1754	typedef DWORD kmp_key_t;
1755	#endif /* KMP_OS_WINDOWS */
1756
1757	#if KMP_OS_UNIX
1758	typedef pthread_t kmp_thread_t;
1759	typedef pthread_key_t kmp_key_t;
1760	#endif
1761
1762	extern kmp_key_t __kmp_gtid_threadprivate_key;
1763
1764	typedef struct kmp_sys_info {
1765	long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1766	long minflt; /* the number of page faults serviced without any I/O */
1767	long majflt; /* the number of page faults serviced that required I/O */
1768	long nswap; /* the number of times a process was "swapped" out of memory */
1769	long inblock; /* the number of times the file system had to perform input */
1770	long oublock; /* the number of times the file system had to perform output */
1771	long nvcsw; /* the number of times a context switch was voluntarily */
1772	long nivcsw; /* the number of times a context switch was forced */
1773	} kmp_sys_info_t;
1774
1775	#if USE_ITT_BUILD
1776	// We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1777	// required type here. Later we will check the type meets requirements.
1778	typedef int kmp_itt_mark_t;
1779	#define KMP_ITT_DEBUG 0
1780	#endif /* USE_ITT_BUILD */
1781
1782	typedef kmp_int32 kmp_critical_name[8];
1783
1784	/*!
1785	@ingroup PARALLEL
1786	The type for a microtask which gets passed to @ref __kmpc_fork_call().
1787	The arguments to the outlined function are
1788	@param global_tid the global thread identity of the thread executing the
1789	function.
1790	@param bound_tid the local identity of the thread executing the function
1791	@param ... pointers to shared variables accessed by the function.
1792	*/
1793	typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1794	typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1795	...);
1796
1797	/*!
1798	@ingroup THREADPRIVATE
1799	@{
1800	*/
1801	/* ---------------------------------------------------------------------------
1802	*/
1803	/* Threadprivate initialization/finalization function declarations */
1804
1805	/* for non-array objects: __kmpc_threadprivate_register() */
1806
1807	/*!
1808	Pointer to the constructor function.
1809	The first argument is the <tt>this</tt> pointer
1810	*/
1811	typedef void *(*kmpc_ctor)(void *);
1812
1813	/*!
1814	Pointer to the destructor function.
1815	The first argument is the <tt>this</tt> pointer
1816	*/
1817	typedef void (*kmpc_dtor)(
1818	void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1819	compiler */
1820	/*!
1821	Pointer to an alternate constructor.
1822	The first argument is the <tt>this</tt> pointer.
1823	*/
1824	typedef void *(*kmpc_cctor)(void *, void *);
1825
1826	/* for array objects: __kmpc_threadprivate_register_vec() */
1827	/* First arg: "this" pointer */
1828	/* Last arg: number of array elements */
1829	/*!
1830	Array constructor.
1831	First argument is the <tt>this</tt> pointer
1832	Second argument the number of array elements.
1833	*/
1834	typedef void *(*kmpc_ctor_vec)(void *, size_t);
1835	/*!
1836	Pointer to the array destructor function.
1837	The first argument is the <tt>this</tt> pointer
1838	Second argument the number of array elements.
1839	*/
1840	typedef void (*kmpc_dtor_vec)(void *, size_t);
1841	/*!
1842	Array constructor.
1843	First argument is the <tt>this</tt> pointer
1844	Third argument the number of array elements.
1845	*/
1846	typedef void *(*kmpc_cctor_vec)(void *, void *,
1847	size_t); /* function unused by compiler */
1848
1849	/*!
1850	@}
1851	*/
1852
1853	/* keeps tracked of threadprivate cache allocations for cleanup later */
1854	typedef struct kmp_cached_addr {
1855	void **addr; /* address of allocated cache */
1856	void ***compiler_cache; /* pointer to compiler's cache */
1857	void *data; /* pointer to global data */
1858	struct kmp_cached_addr *next; /* pointer to next cached address */
1859	} kmp_cached_addr_t;
1860
1861	struct private_data {
1862	struct private_data *next; /* The next descriptor in the list */
1863	void *data; /* The data buffer for this descriptor */
1864	int more; /* The repeat count for this descriptor */
1865	size_t size; /* The data size for this descriptor */
1866	};
1867
1868	struct private_common {
1869	struct private_common *next;
1870	struct private_common *link;
1871	void *gbl_addr;
1872	void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1873	size_t cmn_size;
1874	};
1875
1876	struct shared_common {
1877	struct shared_common *next;
1878	struct private_data *pod_init;
1879	void *obj_init;
1880	void *gbl_addr;
1881	union {
1882	kmpc_ctor ctor;
1883	kmpc_ctor_vec ctorv;
1884	} ct;
1885	union {
1886	kmpc_cctor cctor;
1887	kmpc_cctor_vec cctorv;
1888	} cct;
1889	union {
1890	kmpc_dtor dtor;
1891	kmpc_dtor_vec dtorv;
1892	} dt;
1893	size_t vec_len;
1894	int is_vec;
1895	size_t cmn_size;
1896	};
1897
1898	#define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1899	#define KMP_HASH_TABLE_SIZE \
1900	(1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1901	#define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1902	#define KMP_HASH(x) \
1903	((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1904
1905	struct common_table {
1906	struct private_common *data[KMP_HASH_TABLE_SIZE];
1907	};
1908
1909	struct shared_table {
1910	struct shared_common *data[KMP_HASH_TABLE_SIZE];
1911	};
1912
1913	/* ------------------------------------------------------------------------ */
1914
1915	#if KMP_USE_HIER_SCHED
1916	// Shared barrier data that exists inside a single unit of the scheduling
1917	// hierarchy
1918	typedef struct kmp_hier_private_bdata_t {
1919	kmp_int32 num_active;
1920	kmp_uint64 index;
1921	kmp_uint64 wait_val[2];
1922	} kmp_hier_private_bdata_t;
1923	#endif
1924
1925	typedef struct kmp_sched_flags {
1926	unsigned ordered : 1;
1927	unsigned nomerge : 1;
1928	unsigned contains_last : 1;
1929	unsigned use_hier : 1; // Used in KMP_USE_HIER_SCHED code
1930	unsigned use_hybrid : 1; // Used in KMP_WEIGHTED_ITERATIONS_SUPPORTED code
1931	unsigned unused : 27;
1932	} kmp_sched_flags_t;
1933
1934	KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1935
1936	#if KMP_STATIC_STEAL_ENABLED
1937	typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1938	kmp_int32 count;
1939	kmp_int32 ub;
1940	/* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1941	kmp_int32 lb;
1942	kmp_int32 st;
1943	kmp_int32 tc;
1944	kmp_lock_t *steal_lock; // lock used for chunk stealing
1945
1946	kmp_uint32 ordered_lower;
1947	kmp_uint32 ordered_upper;
1948
1949	// KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1950	// a) parm3 is properly aligned and
1951	// b) all parm1-4 are on the same cache line.
1952	// Because of parm1-4 are used together, performance seems to be better
1953	// if they are on the same cache line (not measured though).
1954
1955	struct KMP_ALIGN(32) {
1956	kmp_int32 parm1;
1957	kmp_int32 parm2;
1958	kmp_int32 parm3;
1959	kmp_int32 parm4;
1960	};
1961
1962	#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1963	kmp_uint32 pchunks;
1964	kmp_uint32 num_procs_with_pcore;
1965	kmp_int32 first_thread_with_ecore;
1966	#endif
1967	#if KMP_OS_WINDOWS
1968	kmp_int32 last_upper;
1969	#endif /* KMP_OS_WINDOWS */
1970	} dispatch_private_info32_t;
1971
1972	#if CACHE_LINE <= 128
1973	KMP_BUILD_ASSERT(sizeof(dispatch_private_info32_t) <= 128);
1974	#endif
1975
1976	typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1977	kmp_int64 count; // current chunk number for static & static-steal scheduling
1978	kmp_int64 ub; /* upper-bound */
1979	/* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1980	kmp_int64 lb; /* lower-bound */
1981	kmp_int64 st; /* stride */
1982	kmp_int64 tc; /* trip count (number of iterations) */
1983	kmp_lock_t *steal_lock; // lock used for chunk stealing
1984
1985	kmp_uint64 ordered_lower;
1986	kmp_uint64 ordered_upper;
1987	/* parm[1-4] are used in different ways by different scheduling algorithms */
1988
1989	// KMP_ALIGN(32) ensures ( if the KMP_ALIGN macro is turned on )
1990	// a) parm3 is properly aligned and
1991	// b) all parm1-4 are in the same cache line.
1992	// Because of parm1-4 are used together, performance seems to be better
1993	// if they are in the same line (not measured though).
1994	struct KMP_ALIGN(32) {
1995	kmp_int64 parm1;
1996	kmp_int64 parm2;
1997	kmp_int64 parm3;
1998	kmp_int64 parm4;
1999	};
2000
2001	#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
2002	kmp_uint64 pchunks;
2003	kmp_uint64 num_procs_with_pcore;
2004	kmp_int64 first_thread_with_ecore;
2005	#endif
2006
2007	#if KMP_OS_WINDOWS
2008	kmp_int64 last_upper;
2009	#endif /* KMP_OS_WINDOWS */
2010	} dispatch_private_info64_t;
2011
2012	#if CACHE_LINE <= 128
2013	KMP_BUILD_ASSERT(sizeof(dispatch_private_info64_t) <= 128);
2014	#endif
2015
2016	#else /* KMP_STATIC_STEAL_ENABLED */
2017	typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
2018	kmp_int32 lb;
2019	kmp_int32 ub;
2020	kmp_int32 st;
2021	kmp_int32 tc;
2022
2023	kmp_int32 parm1;
2024	kmp_int32 parm2;
2025	kmp_int32 parm3;
2026	kmp_int32 parm4;
2027
2028	kmp_int32 count;
2029
2030	kmp_uint32 ordered_lower;
2031	kmp_uint32 ordered_upper;
2032	#if KMP_OS_WINDOWS
2033	kmp_int32 last_upper;
2034	#endif /* KMP_OS_WINDOWS */
2035	} dispatch_private_info32_t;
2036
2037	typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
2038	kmp_int64 lb; /* lower-bound */
2039	kmp_int64 ub; /* upper-bound */
2040	kmp_int64 st; /* stride */
2041	kmp_int64 tc; /* trip count (number of iterations) */
2042
2043	/* parm[1-4] are used in different ways by different scheduling algorithms */
2044	kmp_int64 parm1;
2045	kmp_int64 parm2;
2046	kmp_int64 parm3;
2047	kmp_int64 parm4;
2048
2049	kmp_int64 count; /* current chunk number for static scheduling */
2050
2051	kmp_uint64 ordered_lower;
2052	kmp_uint64 ordered_upper;
2053	#if KMP_OS_WINDOWS
2054	kmp_int64 last_upper;
2055	#endif /* KMP_OS_WINDOWS */
2056	} dispatch_private_info64_t;
2057	#endif /* KMP_STATIC_STEAL_ENABLED */
2058
2059	typedef struct KMP_ALIGN_CACHE dispatch_private_info {
2060	union private_info {
2061	dispatch_private_info32_t p32;
2062	dispatch_private_info64_t p64;
2063	} u;
2064	enum sched_type schedule; /* scheduling algorithm */
2065	kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
2066	std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
2067	kmp_int32 ordered_bumped;
2068	// Stack of buffers for nest of serial regions
2069	struct dispatch_private_info *next;
2070	kmp_int32 type_size; /* the size of types in private_info */
2071	#if KMP_USE_HIER_SCHED
2072	kmp_int32 hier_id;
2073	void *parent; /* hierarchical scheduling parent pointer */
2074	#endif
2075	enum cons_type pushed_ws;
2076	} dispatch_private_info_t;
2077
2078	typedef struct dispatch_shared_info32 {
2079	/* chunk index under dynamic, number of idle threads under static-steal;
2080	iteration index otherwise */
2081	volatile kmp_uint32 iteration;
2082	volatile kmp_int32 num_done;
2083	volatile kmp_uint32 ordered_iteration;
2084	// Dummy to retain the structure size after making ordered_iteration scalar
2085	kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
2086	} dispatch_shared_info32_t;
2087
2088	typedef struct dispatch_shared_info64 {
2089	/* chunk index under dynamic, number of idle threads under static-steal;
2090	iteration index otherwise */
2091	volatile kmp_uint64 iteration;
2092	volatile kmp_int64 num_done;
2093	volatile kmp_uint64 ordered_iteration;
2094	// Dummy to retain the structure size after making ordered_iteration scalar
2095	kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
2096	} dispatch_shared_info64_t;
2097
2098	typedef struct dispatch_shared_info {
2099	union shared_info {
2100	dispatch_shared_info32_t s32;
2101	dispatch_shared_info64_t s64;
2102	} u;
2103	volatile kmp_uint32 buffer_index;
2104	volatile kmp_int32 doacross_buf_idx; // teamwise index
2105	volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
2106	kmp_int32 doacross_num_done; // count finished threads
2107	#if KMP_USE_HIER_SCHED
2108	void *hier;
2109	#endif
2110	#if KMP_USE_HWLOC
2111	// When linking with libhwloc, the ORDERED EPCC test slows down on big
2112	// machines (> 48 cores). Performance analysis showed that a cache thrash
2113	// was occurring and this padding helps alleviate the problem.
2114	char padding[64];
2115	#endif
2116	} dispatch_shared_info_t;
2117
2118	typedef struct kmp_disp {
2119	/* Vector for ORDERED SECTION */
2120	void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2121	/* Vector for END ORDERED SECTION */
2122	void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2123
2124	dispatch_shared_info_t *th_dispatch_sh_current;
2125	dispatch_private_info_t *th_dispatch_pr_current;
2126
2127	dispatch_private_info_t *th_disp_buffer;
2128	kmp_uint32 th_disp_index;
2129	kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2130	volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2131	kmp_int64 *th_doacross_info; // info on loop bounds
2132	#if KMP_USE_INTERNODE_ALIGNMENT
2133	char more_padding[INTERNODE_CACHE_LINE];
2134	#endif
2135	} kmp_disp_t;
2136
2137	/* ------------------------------------------------------------------------ */
2138	/* Barrier stuff */
2139
2140	/* constants for barrier state update */
2141	#define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2142	#define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2143	#define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2144	#define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2145
2146	#define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2147	#define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2148	#define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2149
2150	#if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2151	#error "Barrier sleep bit must be smaller than barrier bump bit"
2152	#endif
2153	#if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2154	#error "Barrier unused bit must be smaller than barrier bump bit"
2155	#endif
2156
2157	// Constants for release barrier wait state: currently, hierarchical only
2158	#define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2159	#define KMP_BARRIER_OWN_FLAG \
2160	1 // Normal state; worker waiting on own b_go flag in release
2161	#define KMP_BARRIER_PARENT_FLAG \
2162	2 // Special state; worker waiting on parent's b_go flag in release
2163	#define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2164	3 // Special state; tells worker to shift from parent to own b_go
2165	#define KMP_BARRIER_SWITCHING \
2166	4 // Special state; worker resets appropriate flag on wake-up
2167
2168	#define KMP_NOT_SAFE_TO_REAP \
2169	0 // Thread th_reap_state: not safe to reap (tasking)
2170	#define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2171
2172	// The flag_type describes the storage used for the flag.
2173	enum flag_type {
2174	flag32, /**< atomic 32 bit flags */
2175	flag64, /**< 64 bit flags */
2176	atomic_flag64, /**< atomic 64 bit flags */
2177	flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */
2178	flag_unset
2179	};
2180
2181	enum barrier_type {
2182	bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2183	barriers if enabled) */
2184	bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2185	#if KMP_FAST_REDUCTION_BARRIER
2186	bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2187	#endif // KMP_FAST_REDUCTION_BARRIER
2188	bs_last_barrier /* Just a placeholder to mark the end */
2189	};
2190
2191	// to work with reduction barriers just like with plain barriers
2192	#if !KMP_FAST_REDUCTION_BARRIER
2193	#define bs_reduction_barrier bs_plain_barrier
2194	#endif // KMP_FAST_REDUCTION_BARRIER
2195
2196	typedef enum kmp_bar_pat { /* Barrier communication patterns */
2197	bp_linear_bar =
2198	0, /* Single level (degenerate) tree */
2199	bp_tree_bar =
2200	1, /* Balanced tree with branching factor 2^n */
2201	bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2202	branching factor 2^n */
2203	bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2204	bp_dist_bar = 4, /* Distributed barrier */
2205	bp_last_bar /* Placeholder to mark the end */
2206	} kmp_bar_pat_e;
2207
2208	#define KMP_BARRIER_ICV_PUSH 1
2209
2210	/* Record for holding the values of the internal controls stack records */
2211	typedef struct kmp_internal_control {
2212	int serial_nesting_level; /* corresponds to the value of the
2213	th_team_serialized field */
2214	kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2215	thread) */
2216	kmp_int8
2217	bt_set; /* internal control for whether blocktime is explicitly set */
2218	int blocktime; /* internal control for blocktime */
2219	#if KMP_USE_MONITOR
2220	int bt_intervals; /* internal control for blocktime intervals */
2221	#endif
2222	int nproc; /* internal control for #threads for next parallel region (per
2223	thread) */
2224	int thread_limit; /* internal control for thread-limit-var */
2225	int task_thread_limit; /* internal control for thread-limit-var of a task*/
2226	int max_active_levels; /* internal control for max_active_levels */
2227	kmp_r_sched_t
2228	sched; /* internal control for runtime schedule {sched,chunk} pair */
2229	kmp_proc_bind_t proc_bind; /* internal control for affinity */
2230	kmp_int32 default_device; /* internal control for default device */
2231	struct kmp_internal_control *next;
2232	} kmp_internal_control_t;
2233
2234	static inline void copy_icvs(kmp_internal_control_t *dst,
2235	kmp_internal_control_t *src) {
2236	*dst = *src;
2237	}
2238
2239	/* Thread barrier needs volatile barrier fields */
2240	typedef struct KMP_ALIGN_CACHE kmp_bstate {
2241	// th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2242	// uses of it). It is not explicitly aligned below, because we *don't* want
2243	// it to be padded -- instead, we fit b_go into the same cache line with
2244	// th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2245	kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2246	// Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2247	// same NGO store
2248	volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2249	KMP_ALIGN_CACHE volatile kmp_uint64
2250	b_arrived; // STATE => task reached synch point.
2251	kmp_uint32 *skip_per_level;
2252	kmp_uint32 my_level;
2253	kmp_int32 parent_tid;
2254	kmp_int32 old_tid;
2255	kmp_uint32 depth;
2256	struct kmp_bstate *parent_bar;
2257	kmp_team_t *team;
2258	kmp_uint64 leaf_state;
2259	kmp_uint32 nproc;
2260	kmp_uint8 base_leaf_kids;
2261	kmp_uint8 leaf_kids;
2262	kmp_uint8 offset;
2263	kmp_uint8 wait_flag;
2264	kmp_uint8 use_oncore_barrier;
2265	#if USE_DEBUGGER
2266	// The following field is intended for the debugger solely. Only the worker
2267	// thread itself accesses this field: the worker increases it by 1 when it
2268	// arrives to a barrier.
2269	KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2270	#endif /* USE_DEBUGGER */
2271	} kmp_bstate_t;
2272
2273	union KMP_ALIGN_CACHE kmp_barrier_union {
2274	double b_align; /* use worst case alignment */
2275	char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2276	kmp_bstate_t bb;
2277	};
2278
2279	typedef union kmp_barrier_union kmp_balign_t;
2280
2281	/* Team barrier needs only non-volatile arrived counter */
2282	union KMP_ALIGN_CACHE kmp_barrier_team_union {
2283	double b_align; /* use worst case alignment */
2284	char b_pad[CACHE_LINE];
2285	struct {
2286	kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2287	#if USE_DEBUGGER
2288	// The following two fields are indended for the debugger solely. Only
2289	// primary thread of the team accesses these fields: the first one is
2290	// increased by 1 when the primary thread arrives to a barrier, the second
2291	// one is increased by one when all the threads arrived.
2292	kmp_uint b_master_arrived;
2293	kmp_uint b_team_arrived;
2294	#endif
2295	};
2296	};
2297
2298	typedef union kmp_barrier_team_union kmp_balign_team_t;
2299
2300	/* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2301	threads when a condition changes. This is to workaround an NPTL bug where
2302	padding was added to pthread_cond_t which caused the initialization routine
2303	to write outside of the structure if compiled on pre-NPTL threads. */
2304	#if KMP_OS_WINDOWS
2305	typedef struct kmp_win32_mutex {
2306	/* The Lock */
2307	CRITICAL_SECTION cs;
2308	} kmp_win32_mutex_t;
2309
2310	typedef struct kmp_win32_cond {
2311	/* Count of the number of waiters. */
2312	int waiters_count_;
2313
2314	/* Serialize access to <waiters_count_> */
2315	kmp_win32_mutex_t waiters_count_lock_;
2316
2317	/* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2318	int release_count_;
2319
2320	/* Keeps track of the current "generation" so that we don't allow */
2321	/* one thread to steal all the "releases" from the broadcast. */
2322	int wait_generation_count_;
2323
2324	/* A manual-reset event that's used to block and release waiting threads. */
2325	HANDLE event_;
2326	} kmp_win32_cond_t;
2327	#endif
2328
2329	#if KMP_OS_UNIX
2330
2331	union KMP_ALIGN_CACHE kmp_cond_union {
2332	double c_align;
2333	char c_pad[CACHE_LINE];
2334	pthread_cond_t c_cond;
2335	};
2336
2337	typedef union kmp_cond_union kmp_cond_align_t;
2338
2339	union KMP_ALIGN_CACHE kmp_mutex_union {
2340	double m_align;
2341	char m_pad[CACHE_LINE];
2342	pthread_mutex_t m_mutex;
2343	};
2344
2345	typedef union kmp_mutex_union kmp_mutex_align_t;
2346
2347	#endif /* KMP_OS_UNIX */
2348
2349	typedef struct kmp_desc_base {
2350	void *ds_stackbase;
2351	size_t ds_stacksize;
2352	int ds_stackgrow;
2353	kmp_thread_t ds_thread;
2354	volatile int ds_tid;
2355	int ds_gtid;
2356	#if KMP_OS_WINDOWS
2357	volatile int ds_alive;
2358	DWORD ds_thread_id;
2359	/* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2360	However, debugger support (libomp_db) cannot work with handles, because they
2361	uncomparable. For example, debugger requests info about thread with handle h.
2362	h is valid within debugger process, and meaningless within debugee process.
2363	Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2364	within debugee process, but it is a *new* handle which does *not* equal to
2365	any other handle in debugee... The only way to compare handles is convert
2366	them to system-wide ids. GetThreadId() function is available only in
2367	Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2368	on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2369	thread id by call to GetCurrentThreadId() from within the thread and save it
2370	to let libomp_db identify threads. */
2371	#endif /* KMP_OS_WINDOWS */
2372	} kmp_desc_base_t;
2373
2374	typedef union KMP_ALIGN_CACHE kmp_desc {
2375	double ds_align; /* use worst case alignment */
2376	char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2377	kmp_desc_base_t ds;
2378	} kmp_desc_t;
2379
2380	typedef struct kmp_local {
2381	volatile int this_construct; /* count of single's encountered by thread */
2382	void *reduce_data;
2383	#if KMP_USE_BGET
2384	void *bget_data;
2385	void *bget_list;
2386	#if !USE_CMP_XCHG_FOR_BGET
2387	#ifdef USE_QUEUING_LOCK_FOR_BGET
2388	kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2389	#else
2390	kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2391	// bootstrap lock so we can use it at library
2392	// shutdown.
2393	#endif /* USE_LOCK_FOR_BGET */
2394	#endif /* ! USE_CMP_XCHG_FOR_BGET */
2395	#endif /* KMP_USE_BGET */
2396
2397	PACKED_REDUCTION_METHOD_T
2398	packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2399	__kmpc_end_reduce*() */
2400
2401	} kmp_local_t;
2402
2403	#define KMP_CHECK_UPDATE(a, b) \
2404	if ((a) != (b)) \
2405	(a) = (b)
2406	#define KMP_CHECK_UPDATE_SYNC(a, b) \
2407	if ((a) != (b)) \
2408	TCW_SYNC_PTR((a), (b))
2409
2410	#define get__blocktime(xteam, xtid) \
2411	((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2412	#define get__bt_set(xteam, xtid) \
2413	((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2414	#if KMP_USE_MONITOR
2415	#define get__bt_intervals(xteam, xtid) \
2416	((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2417	#endif
2418
2419	#define get__dynamic_2(xteam, xtid) \
2420	((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2421	#define get__nproc_2(xteam, xtid) \
2422	((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2423	#define get__sched_2(xteam, xtid) \
2424	((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2425
2426	#define set__blocktime_team(xteam, xtid, xval) \
2427	(((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2428	(xval))
2429
2430	#if KMP_USE_MONITOR
2431	#define set__bt_intervals_team(xteam, xtid, xval) \
2432	(((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2433	(xval))
2434	#endif
2435
2436	#define set__bt_set_team(xteam, xtid, xval) \
2437	(((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2438
2439	#define set__dynamic(xthread, xval) \
2440	(((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2441	#define get__dynamic(xthread) \
2442	(((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2443
2444	#define set__nproc(xthread, xval) \
2445	(((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2446
2447	#define set__thread_limit(xthread, xval) \
2448	(((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2449
2450	#define set__max_active_levels(xthread, xval) \
2451	(((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2452
2453	#define get__max_active_levels(xthread) \
2454	((xthread)->th.th_current_task->td_icvs.max_active_levels)
2455
2456	#define set__sched(xthread, xval) \
2457	(((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2458
2459	#define set__proc_bind(xthread, xval) \
2460	(((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2461	#define get__proc_bind(xthread) \
2462	((xthread)->th.th_current_task->td_icvs.proc_bind)
2463
2464	// OpenMP tasking data structures
2465
2466	typedef enum kmp_tasking_mode {
2467	tskm_immediate_exec = 0,
2468	tskm_extra_barrier = 1,
2469	tskm_task_teams = 2,
2470	tskm_max = 2
2471	} kmp_tasking_mode_t;
2472
2473	extern kmp_tasking_mode_t
2474	__kmp_tasking_mode; /* determines how/when to execute tasks */
2475	extern int __kmp_task_stealing_constraint;
2476	extern int __kmp_enable_task_throttling;
2477	extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2478	// specified, defaults to 0 otherwise
2479	// Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2480	extern kmp_int32 __kmp_max_task_priority;
2481	// Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2482	extern kmp_uint64 __kmp_taskloop_min_tasks;
2483
2484	/* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2485	taskdata first */
2486	#define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2487	#define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2488
2489	// The tt_found_tasks flag is a signal to all threads in the team that tasks
2490	// were spawned and queued since the previous barrier release.
2491	#define KMP_TASKING_ENABLED(task_team) \
2492	(TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2493	/*!
2494	@ingroup BASIC_TYPES
2495	@{
2496	*/
2497
2498	/*!
2499	*/
2500	typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2501
2502	typedef union kmp_cmplrdata {
2503	kmp_int32 priority; /**< priority specified by user for the task */
2504	kmp_routine_entry_t
2505	destructors; /* pointer to function to invoke deconstructors of
2506	firstprivate C++ objects */
2507	/* future data */
2508	} kmp_cmplrdata_t;
2509
2510	/* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2511	/*!
2512	*/
2513	typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2514	void *shareds; /**< pointer to block of pointers to shared vars */
2515	kmp_routine_entry_t
2516	routine; /**< pointer to routine to call for executing task */
2517	kmp_int32 part_id; /**< part id for the task */
2518	kmp_cmplrdata_t
2519	data1; /* Two known optional additions: destructors and priority */
2520	kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2521	/* future data */
2522	/* private vars */
2523	} kmp_task_t;
2524
2525	/*!
2526	@}
2527	*/
2528
2529	typedef struct kmp_taskgroup {
2530	std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2531	std::atomic<kmp_int32>
2532	cancel_request; // request for cancellation of this taskgroup
2533	struct kmp_taskgroup *parent; // parent taskgroup
2534	// Block of data to perform task reduction
2535	void *reduce_data; // reduction related info
2536	kmp_int32 reduce_num_data; // number of data items to reduce
2537	uintptr_t *gomp_data; // gomp reduction data
2538	} kmp_taskgroup_t;
2539
2540	// forward declarations
2541	typedef union kmp_depnode kmp_depnode_t;
2542	typedef struct kmp_depnode_list kmp_depnode_list_t;
2543	typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2544
2545	// macros for checking dep flag as an integer
2546	#define KMP_DEP_IN 0x1
2547	#define KMP_DEP_OUT 0x2
2548	#define KMP_DEP_INOUT 0x3
2549	#define KMP_DEP_MTX 0x4
2550	#define KMP_DEP_SET 0x8
2551	#define KMP_DEP_ALL 0x80
2552	// Compiler sends us this info. Note: some test cases contain an explicit copy
2553	// of this struct and should be in sync with any changes here.
2554	typedef struct kmp_depend_info {
2555	kmp_intptr_t base_addr;
2556	size_t len;
2557	union {
2558	kmp_uint8 flag; // flag as an unsigned char
2559	struct { // flag as a set of 8 bits
2560	#if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2561	/* Same fields as in the #else branch, but in reverse order */
2562	unsigned all : 1;
2563	unsigned unused : 3;
2564	unsigned set : 1;
2565	unsigned mtx : 1;
2566	unsigned out : 1;
2567	unsigned in : 1;
2568	#else
2569	unsigned in : 1;
2570	unsigned out : 1;
2571	unsigned mtx : 1;
2572	unsigned set : 1;
2573	unsigned unused : 3;
2574	unsigned all : 1;
2575	#endif
2576	} flags;
2577	};
2578	} kmp_depend_info_t;
2579
2580	// Internal structures to work with task dependencies:
2581	struct kmp_depnode_list {
2582	kmp_depnode_t *node;
2583	kmp_depnode_list_t *next;
2584	};
2585
2586	// Max number of mutexinoutset dependencies per node
2587	#define MAX_MTX_DEPS 4
2588
2589	typedef struct kmp_base_depnode {
2590	kmp_depnode_list_t *successors; /* used under lock */
2591	kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2592	kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2593	kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2594	kmp_lock_t lock; /* guards shared fields: task, successors */
2595	#if KMP_SUPPORT_GRAPH_OUTPUT
2596	kmp_uint32 id;
2597	#endif
2598	std::atomic<kmp_int32> npredecessors;
2599	std::atomic<kmp_int32> nrefs;
2600	} kmp_base_depnode_t;
2601
2602	union KMP_ALIGN_CACHE kmp_depnode {
2603	double dn_align; /* use worst case alignment */
2604	char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2605	kmp_base_depnode_t dn;
2606	};
2607
2608	struct kmp_dephash_entry {
2609	kmp_intptr_t addr;
2610	kmp_depnode_t *last_out;
2611	kmp_depnode_list_t *last_set;
2612	kmp_depnode_list_t *prev_set;
2613	kmp_uint8 last_flag;
2614	kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2615	kmp_dephash_entry_t *next_in_bucket;
2616	};
2617
2618	typedef struct kmp_dephash {
2619	kmp_dephash_entry_t **buckets;
2620	size_t size;
2621	kmp_depnode_t *last_all;
2622	size_t generation;
2623	kmp_uint32 nelements;
2624	kmp_uint32 nconflicts;
2625	} kmp_dephash_t;
2626
2627	typedef struct kmp_task_affinity_info {
2628	kmp_intptr_t base_addr;
2629	size_t len;
2630	struct {
2631	bool flag1 : 1;
2632	bool flag2 : 1;
2633	kmp_int32 reserved : 30;
2634	} flags;
2635	} kmp_task_affinity_info_t;
2636
2637	typedef enum kmp_event_type_t {
2638	KMP_EVENT_UNINITIALIZED = 0,
2639	KMP_EVENT_ALLOW_COMPLETION = 1
2640	} kmp_event_type_t;
2641
2642	typedef struct {
2643	kmp_event_type_t type;
2644	kmp_tas_lock_t lock;
2645	union {
2646	kmp_task_t *task;
2647	} ed;
2648	} kmp_event_t;
2649
2650	#if OMPX_TASKGRAPH
2651	// Initial number of allocated nodes while recording
2652	#define INIT_MAPSIZE 50
2653
2654	typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2655	unsigned nowait : 1;
2656	unsigned re_record : 1;
2657	unsigned reserved : 30;
2658	} kmp_taskgraph_flags_t;
2659
2660	/// Represents a TDG node
2661	typedef struct kmp_node_info {
2662	kmp_task_t *task; // Pointer to the actual task
2663	kmp_int32 *successors; // Array of the succesors ids
2664	kmp_int32 nsuccessors; // Number of succesors of the node
2665	std::atomic<kmp_int32>
2666	npredecessors_counter; // Number of predessors on the fly
2667	kmp_int32 npredecessors; // Total number of predecessors
2668	kmp_int32 successors_size; // Number of allocated succesors ids
2669	kmp_taskdata_t *parent_task; // Parent implicit task
2670	} kmp_node_info_t;
2671
2672	/// Represent a TDG's current status
2673	typedef enum kmp_tdg_status {
2674	KMP_TDG_NONE = 0,
2675	KMP_TDG_RECORDING = 1,
2676	KMP_TDG_READY = 2
2677	} kmp_tdg_status_t;
2678
2679	/// Structure that contains a TDG
2680	typedef struct kmp_tdg_info {
2681	kmp_int32 tdg_id; // Unique idenfifier of the TDG
2682	kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2683	kmp_int32 map_size; // Number of allocated TDG nodes
2684	kmp_int32 num_roots; // Number of roots tasks int the TDG
2685	kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2686	kmp_node_info_t *record_map; // Array of TDG nodes
2687	kmp_tdg_status_t tdg_status =
2688	KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2689	std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2690	kmp_bootstrap_lock_t
2691	graph_lock; // Protect graph attributes when updated via taskloop_recur
2692	// Taskloop reduction related
2693	void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2694	// __kmpc_taskred_init
2695	kmp_int32 rec_num_taskred;
2696	} kmp_tdg_info_t;
2697
2698	extern int __kmp_tdg_dot;
2699	extern kmp_int32 __kmp_max_tdgs;
2700	extern kmp_tdg_info_t **__kmp_global_tdgs;
2701	extern kmp_int32 __kmp_curr_tdg_idx;
2702	extern kmp_int32 __kmp_successors_size;
2703	extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2704	extern kmp_int32 __kmp_num_tdg;
2705	#endif
2706
2707	#ifdef BUILD_TIED_TASK_STACK
2708
2709	/* Tied Task stack definitions */
2710	typedef struct kmp_stack_block {
2711	kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2712	struct kmp_stack_block *sb_next;
2713	struct kmp_stack_block *sb_prev;
2714	} kmp_stack_block_t;
2715
2716	typedef struct kmp_task_stack {
2717	kmp_stack_block_t ts_first_block; // first block of stack entries
2718	kmp_taskdata_t **ts_top; // pointer to the top of stack
2719	kmp_int32 ts_entries; // number of entries on the stack
2720	} kmp_task_stack_t;
2721
2722	#endif // BUILD_TIED_TASK_STACK
2723
2724	typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2725	#if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2726	/* Same fields as in the #else branch, but in reverse order */
2727	#if OMPX_TASKGRAPH
2728	unsigned reserved31 : 5;
2729	unsigned onced : 1;
2730	#else
2731	unsigned reserved31 : 6;
2732	#endif
2733	unsigned target : 1;
2734	unsigned native : 1;
2735	unsigned freed : 1;
2736	unsigned complete : 1;
2737	unsigned executing : 1;
2738	unsigned started : 1;
2739	unsigned team_serial : 1;
2740	unsigned tasking_ser : 1;
2741	unsigned task_serial : 1;
2742	unsigned tasktype : 1;
2743	unsigned reserved : 8;
2744	unsigned hidden_helper : 1;
2745	unsigned detachable : 1;
2746	unsigned priority_specified : 1;
2747	unsigned proxy : 1;
2748	unsigned destructors_thunk : 1;
2749	unsigned merged_if0 : 1;
2750	unsigned final : 1;
2751	unsigned tiedness : 1;
2752	#else
2753	/* Compiler flags */ /* Total compiler flags must be 16 bits */
2754	unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2755	unsigned final : 1; /* task is final(1) so execute immediately */
2756	unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2757	code path */
2758	unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2759	invoke destructors from the runtime */
2760	unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2761	context of the RTL) */
2762	unsigned priority_specified : 1; /* set if the compiler provides priority
2763	setting for the task */
2764	unsigned detachable : 1; /* 1 == can detach */
2765	unsigned hidden_helper : 1; /* 1 == hidden helper task */
2766	unsigned reserved : 8; /* reserved for compiler use */
2767
2768	/* Library flags */ /* Total library flags must be 16 bits */
2769	unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2770	unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2771	unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2772	// (1) or may be deferred (0)
2773	unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2774	// (0) [>= 2 threads]
2775	/* If either team_serial or tasking_ser is set, task team may be NULL */
2776	/* Task State Flags: */
2777	unsigned started : 1; /* 1==started, 0==not started */
2778	unsigned executing : 1; /* 1==executing, 0==not executing */
2779	unsigned complete : 1; /* 1==complete, 0==not complete */
2780	unsigned freed : 1; /* 1==freed, 0==allocated */
2781	unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2782	unsigned target : 1;
2783	#if OMPX_TASKGRAPH
2784	unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2785	unsigned reserved31 : 5; /* reserved for library use */
2786	#else
2787	unsigned reserved31 : 6; /* reserved for library use */
2788	#endif
2789	#endif
2790	} kmp_tasking_flags_t;
2791
2792	typedef struct kmp_target_data {
2793	void *async_handle; // libomptarget async handle for task completion query
2794	} kmp_target_data_t;
2795
2796	struct kmp_taskdata { /* aligned during dynamic allocation */
2797	kmp_int32 td_task_id; /* id, assigned by debugger */
2798	kmp_tasking_flags_t td_flags; /* task flags */
2799	kmp_team_t *td_team; /* team for this task */
2800	kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2801	/* Currently not used except for perhaps IDB */
2802	kmp_taskdata_t *td_parent; /* parent task */
2803	kmp_int32 td_level; /* task nesting level */
2804	std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2805	ident_t *td_ident; /* task identifier */
2806	// Taskwait data.
2807	ident_t *td_taskwait_ident;
2808	kmp_uint32 td_taskwait_counter;
2809	kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2810	KMP_ALIGN_CACHE kmp_internal_control_t
2811	td_icvs; /* Internal control variables for the task */
2812	KMP_ALIGN_CACHE std::atomic<kmp_int32>
2813	td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2814	deallocated */
2815	std::atomic<kmp_int32>
2816	td_incomplete_child_tasks; /* Child tasks not yet complete */
2817	kmp_taskgroup_t
2818	*td_taskgroup; // Each task keeps pointer to its current taskgroup
2819	kmp_dephash_t
2820	*td_dephash; // Dependencies for children tasks are tracked from here
2821	kmp_depnode_t
2822	*td_depnode; // Pointer to graph node if this task has dependencies
2823	kmp_task_team_t *td_task_team;
2824	size_t td_size_alloc; // Size of task structure, including shareds etc.
2825	#if defined(KMP_GOMP_COMPAT)
2826	// 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2827	kmp_int32 td_size_loop_bounds;
2828	#endif
2829	kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2830	#if defined(KMP_GOMP_COMPAT)
2831	// GOMP sends in a copy function for copy constructors
2832	void (*td_copy_func)(void *, void *);
2833	#endif
2834	kmp_event_t td_allow_completion_event;
2835	#if OMPT_SUPPORT
2836	ompt_task_info_t ompt_task_info;
2837	#endif
2838	#if OMPX_TASKGRAPH
2839	bool is_taskgraph = 0; // whether the task is within a TDG
2840	kmp_tdg_info_t *tdg; // used to associate task with a TDG
2841	kmp_int32 td_tdg_task_id; // local task id in its TDG
2842	#endif
2843	kmp_target_data_t td_target_data;
2844	}; // struct kmp_taskdata
2845
2846	// Make sure padding above worked
2847	KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2848
2849	// Data for task team but per thread
2850	typedef struct kmp_base_thread_data {
2851	kmp_info_p *td_thr; // Pointer back to thread info
2852	// Used only in __kmp_execute_tasks_template, maybe not avail until task is
2853	// queued?
2854	kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2855	kmp_taskdata_t *
2856	*td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2857	kmp_int32 td_deque_size; // Size of deck
2858	kmp_uint32 td_deque_head; // Head of deque (will wrap)
2859	kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2860	kmp_int32 td_deque_ntasks; // Number of tasks in deque
2861	// GEH: shouldn't this be volatile since used in while-spin?
2862	kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2863	#ifdef BUILD_TIED_TASK_STACK
2864	kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2865	// scheduling constraint
2866	#endif // BUILD_TIED_TASK_STACK
2867	} kmp_base_thread_data_t;
2868
2869	#define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2870	#define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2871
2872	#define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2873	#define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2874
2875	typedef union KMP_ALIGN_CACHE kmp_thread_data {
2876	kmp_base_thread_data_t td;
2877	double td_align; /* use worst case alignment */
2878	char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2879	} kmp_thread_data_t;
2880
2881	typedef struct kmp_task_pri {
2882	kmp_thread_data_t td;
2883	kmp_int32 priority;
2884	kmp_task_pri *next;
2885	} kmp_task_pri_t;
2886
2887	// Data for task teams which are used when tasking is enabled for the team
2888	typedef struct kmp_base_task_team {
2889	kmp_bootstrap_lock_t
2890	tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2891	/* must be bootstrap lock since used at library shutdown*/
2892
2893	// TODO: check performance vs kmp_tas_lock_t
2894	kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2895	kmp_task_pri_t *tt_task_pri_list;
2896
2897	kmp_task_team_t *tt_next; /* For linking the task team free list */
2898	kmp_thread_data_t
2899	*tt_threads_data; /* Array of per-thread structures for task team */
2900	/* Data survives task team deallocation */
2901	kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2902	executing this team? */
2903	/* TRUE means tt_threads_data is set up and initialized */
2904	kmp_int32 tt_nproc; /* #threads in team */
2905	kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2906	kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2907	kmp_int32 tt_untied_task_encountered;
2908	std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2909	// There is hidden helper thread encountered in this task team so that we must
2910	// wait when waiting on task team
2911	kmp_int32 tt_hidden_helper_task_encountered;
2912
2913	KMP_ALIGN_CACHE
2914	std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2915
2916	KMP_ALIGN_CACHE
2917	volatile kmp_uint32
2918	tt_active; /* is the team still actively executing tasks */
2919	} kmp_base_task_team_t;
2920
2921	union KMP_ALIGN_CACHE kmp_task_team {
2922	kmp_base_task_team_t tt;
2923	double tt_align; /* use worst case alignment */
2924	char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2925	};
2926
2927	typedef struct kmp_task_team_list_t {
2928	kmp_task_team_t *task_team;
2929	kmp_task_team_list_t *next;
2930	} kmp_task_team_list_t;
2931
2932	#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2933	// Free lists keep same-size free memory slots for fast memory allocation
2934	// routines
2935	typedef struct kmp_free_list {
2936	void *th_free_list_self; // Self-allocated tasks free list
2937	void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2938	// threads
2939	void *th_free_list_other; // Non-self free list (to be returned to owner's
2940	// sync list)
2941	} kmp_free_list_t;
2942	#endif
2943	#if KMP_NESTED_HOT_TEAMS
2944	// Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2945	// are not put in teams pool, and they don't put threads in threads pool.
2946	typedef struct kmp_hot_team_ptr {
2947	kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2948	kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2949	} kmp_hot_team_ptr_t;
2950	#endif
2951	typedef struct kmp_teams_size {
2952	kmp_int32 nteams; // number of teams in a league
2953	kmp_int32 nth; // number of threads in each team of the league
2954	} kmp_teams_size_t;
2955
2956	// This struct stores a thread that acts as a "root" for a contention
2957	// group. Contention groups are rooted at kmp_root threads, but also at
2958	// each primary thread of each team created in the teams construct.
2959	// This struct therefore also stores a thread_limit associated with
2960	// that contention group, and a counter to track the number of threads
2961	// active in that contention group. Each thread has a list of these: CG
2962	// root threads have an entry in their list in which cg_root refers to
2963	// the thread itself, whereas other workers in the CG will have a
2964	// single entry where cg_root is same as the entry containing their CG
2965	// root. When a thread encounters a teams construct, it will add a new
2966	// entry to the front of its list, because it now roots a new CG.
2967	typedef struct kmp_cg_root {
2968	kmp_info_p *cg_root; // "root" thread for a contention group
2969	// The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2970	// thread_limit clause for teams primary threads
2971	kmp_int32 cg_thread_limit;
2972	kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2973	struct kmp_cg_root *up; // pointer to higher level CG root in list
2974	} kmp_cg_root_t;
2975
2976	// OpenMP thread data structures
2977
2978	typedef struct KMP_ALIGN_CACHE kmp_base_info {
2979	/* Start with the readonly data which is cache aligned and padded. This is
2980	written before the thread starts working by the primary thread. Uber
2981	masters may update themselves later. Usage does not consider serialized
2982	regions. */
2983	kmp_desc_t th_info;
2984	kmp_team_p *th_team; /* team we belong to */
2985	kmp_root_p *th_root; /* pointer to root of task hierarchy */
2986	kmp_info_p *th_next_pool; /* next available thread in the pool */
2987	kmp_disp_t *th_dispatch; /* thread's dispatch data */
2988	int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2989
2990	/* The following are cached from the team info structure */
2991	/* TODO use these in more places as determined to be needed via profiling */
2992	int th_team_nproc; /* number of threads in a team */
2993	kmp_info_p *th_team_master; /* the team's primary thread */
2994	int th_team_serialized; /* team is serialized */
2995	microtask_t th_teams_microtask; /* save entry address for teams construct */
2996	int th_teams_level; /* save initial level of teams construct */
2997	/* it is 0 on device but may be any on host */
2998
2999	/* The blocktime info is copied from the team struct to the thread struct */
3000	/* at the start of a barrier, and the values stored in the team are used */
3001	/* at points in the code where the team struct is no longer guaranteed */
3002	/* to exist (from the POV of worker threads). */
3003	#if KMP_USE_MONITOR
3004	int th_team_bt_intervals;
3005	int th_team_bt_set;
3006	#else
3007	kmp_uint64 th_team_bt_intervals;
3008	#endif
3009
3010	#if KMP_AFFINITY_SUPPORTED
3011	kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
3012	kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
3013	kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
3014	#endif
3015	omp_allocator_handle_t th_def_allocator; /* default allocator */
3016	/* The data set by the primary thread at reinit, then R/W by the worker */
3017	KMP_ALIGN_CACHE int
3018	th_set_nproc; /* if > 0, then only use this request for the next fork */
3019	int *th_set_nested_nth;
3020	bool th_nt_strict; // num_threads clause has strict modifier
3021	ident_t *th_nt_loc; // loc for strict modifier
3022	int th_nt_sev; // error severity for strict modifier
3023	const char *th_nt_msg; // error message for strict modifier
3024	int th_set_nested_nth_sz;
3025	#if KMP_NESTED_HOT_TEAMS
3026	kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
3027	#endif
3028	kmp_proc_bind_t
3029	th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
3030	kmp_teams_size_t
3031	th_teams_size; /* number of teams/threads in teams construct */
3032	#if KMP_AFFINITY_SUPPORTED
3033	int th_current_place; /* place currently bound to */
3034	int th_new_place; /* place to bind to in par reg */
3035	int th_first_place; /* first place in partition */
3036	int th_last_place; /* last place in partition */
3037	#endif
3038	int th_prev_level; /* previous level for affinity format */
3039	int th_prev_num_threads; /* previous num_threads for affinity format */
3040	#if USE_ITT_BUILD
3041	kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
3042	kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
3043	kmp_uint64 th_frame_time; /* frame timestamp */
3044	#endif /* USE_ITT_BUILD */
3045	kmp_local_t th_local;
3046	struct private_common *th_pri_head;
3047
3048	/* Now the data only used by the worker (after initial allocation) */
3049	/* TODO the first serial team should actually be stored in the info_t
3050	structure. this will help reduce initial allocation overhead */
3051	KMP_ALIGN_CACHE kmp_team_p
3052	*th_serial_team; /*serialized team held in reserve*/
3053
3054	#if OMPT_SUPPORT
3055	ompt_thread_info_t ompt_thread_info;
3056	#endif
3057
3058	/* The following are also read by the primary thread during reinit */
3059	struct common_table *th_pri_common;
3060
3061	volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
3062	/* while awaiting queuing lock acquire */
3063
3064	volatile void *th_sleep_loc; // this points at a kmp_flag<T>
3065	flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
3066
3067	ident_t *th_ident;
3068	unsigned th_x; // Random number generator data
3069	unsigned th_a; // Random number generator data
3070
3071	/* Tasking-related data for the thread */
3072	kmp_task_team_t *th_task_team; // Task team struct
3073	kmp_taskdata_t *th_current_task; // Innermost Task being executed
3074	kmp_uint8 th_task_state; // alternating 0/1 for task team identification
3075	kmp_uint32 th_reap_state; // Non-zero indicates thread is not
3076	// tasking, thus safe to reap
3077
3078	/* More stuff for keeping track of active/sleeping threads (this part is
3079	written by the worker thread) */
3080	kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
3081	int th_active; // ! sleeping; 32 bits for TCR/TCW
3082	std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
3083	// 0 = not used in team; 1 = used in team;
3084	// 2 = transitioning to not used in team; 3 = transitioning to used in team
3085	struct cons_header *th_cons; // used for consistency check
3086	#if KMP_USE_HIER_SCHED
3087	// used for hierarchical scheduling
3088	kmp_hier_private_bdata_t *th_hier_bar_data;
3089	#endif
3090
3091	/* Add the syncronizing data which is cache aligned and padded. */
3092	KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
3093
3094	KMP_ALIGN_CACHE volatile kmp_int32
3095	th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
3096
3097	#if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
3098	#define NUM_LISTS 4
3099	kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
3100	// allocation routines
3101	#endif
3102
3103	#if KMP_OS_WINDOWS
3104	kmp_win32_cond_t th_suspend_cv;
3105	kmp_win32_mutex_t th_suspend_mx;
3106	std::atomic<int> th_suspend_init;
3107	#endif
3108	#if KMP_OS_UNIX
3109	kmp_cond_align_t th_suspend_cv;
3110	kmp_mutex_align_t th_suspend_mx;
3111	std::atomic<int> th_suspend_init_count;
3112	#endif
3113
3114	#if USE_ITT_BUILD
3115	kmp_itt_mark_t th_itt_mark_single;
3116	// alignment ???
3117	#endif /* USE_ITT_BUILD */
3118	#if KMP_STATS_ENABLED
3119	kmp_stats_list *th_stats;
3120	#endif
3121	#if KMP_OS_UNIX
3122	std::atomic<bool> th_blocking;
3123	#endif
3124	kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
3125	} kmp_base_info_t;
3126
3127	typedef union KMP_ALIGN_CACHE kmp_info {
3128	double th_align; /* use worst case alignment */
3129	char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
3130	kmp_base_info_t th;
3131	} kmp_info_t;
3132
3133	// OpenMP thread team data structures
3134
3135	typedef struct kmp_base_data {
3136	volatile kmp_uint32 t_value;
3137	} kmp_base_data_t;
3138
3139	typedef union KMP_ALIGN_CACHE kmp_sleep_team {
3140	double dt_align; /* use worst case alignment */
3141	char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3142	kmp_base_data_t dt;
3143	} kmp_sleep_team_t;
3144
3145	typedef union KMP_ALIGN_CACHE kmp_ordered_team {
3146	double dt_align; /* use worst case alignment */
3147	char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3148	kmp_base_data_t dt;
3149	} kmp_ordered_team_t;
3150
3151	typedef int (*launch_t)(int gtid);
3152
3153	/* Minimum number of ARGV entries to malloc if necessary */
3154	#define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3155
3156	// Set up how many argv pointers will fit in cache lines containing
3157	// t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3158	// larger value for more space between the primary write/worker read section and
3159	// read/write by all section seems to buy more performance on EPCC PARALLEL.
3160	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3161	#define KMP_INLINE_ARGV_BYTES \
3162	(4 * CACHE_LINE - \
3163	((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
3164	sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
3165	CACHE_LINE))
3166	#else
3167	#define KMP_INLINE_ARGV_BYTES \
3168	(2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3169	#endif
3170	#define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3171
3172	typedef struct KMP_ALIGN_CACHE kmp_base_team {
3173	// Synchronization Data
3174	// ---------------------------------------------------------------------------
3175	KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3176	kmp_balign_team_t t_bar[bs_last_barrier];
3177	std::atomic<int> t_construct; // count of single directive encountered by team
3178	char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3179
3180	// [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3181	std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3182	std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3183
3184	// Primary thread only
3185	// ---------------------------------------------------------------------------
3186	KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3187	int t_master_this_cons; // "this_construct" single counter of primary thread
3188	// in parent team
3189	ident_t *t_ident; // if volatile, have to change too much other crud to
3190	// volatile too
3191	kmp_team_p *t_parent; // parent team
3192	kmp_team_p *t_next_pool; // next free team in the team pool
3193	kmp_disp_t *t_dispatch; // thread's dispatch data
3194	kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3195	kmp_proc_bind_t t_proc_bind; // bind type for par region
3196	int t_primary_task_state; // primary thread's task state saved
3197	#if USE_ITT_BUILD
3198	kmp_uint64 t_region_time; // region begin timestamp
3199	#endif /* USE_ITT_BUILD */
3200
3201	// Primary thread write, workers read
3202	// --------------------------------------------------------------------------
3203	KMP_ALIGN_CACHE void **t_argv;
3204	int t_argc;
3205	int t_nproc; // number of threads in team
3206	microtask_t t_pkfn;
3207	launch_t t_invoke; // procedure to launch the microtask
3208
3209	#if OMPT_SUPPORT
3210	ompt_team_info_t ompt_team_info;
3211	ompt_lw_taskteam_t *ompt_serialized_team_info;
3212	#endif
3213
3214	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3215	kmp_int8 t_fp_control_saved;
3216	kmp_int8 t_pad2b;
3217	kmp_int16 t_x87_fpu_control_word; // FP control regs
3218	kmp_uint32 t_mxcsr;
3219	#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3220
3221	void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3222
3223	KMP_ALIGN_CACHE kmp_info_t **t_threads;
3224	kmp_taskdata_t
3225	*t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3226	int t_level; // nested parallel level
3227
3228	KMP_ALIGN_CACHE int t_max_argc;
3229	int t_max_nproc; // max threads this team can handle (dynamically expandable)
3230	int t_serialized; // levels deep of serialized teams
3231	dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3232	int t_id; // team's id, assigned by debugger.
3233	int t_active_level; // nested active parallel level
3234	kmp_r_sched_t t_sched; // run-time schedule for the team
3235	#if KMP_AFFINITY_SUPPORTED
3236	int t_first_place; // first & last place in parent thread's partition.
3237	int t_last_place; // Restore these values to primary thread after par region.
3238	#endif // KMP_AFFINITY_SUPPORTED
3239	int t_display_affinity;
3240	int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3241	// omp_set_num_threads() call
3242	omp_allocator_handle_t t_def_allocator; /* default allocator */
3243
3244	// Read/write by workers as well
3245	#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3246	// Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3247	// regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3248	// padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3249	// CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3250	char dummy_padding[1024];
3251	#endif
3252	// Internal control stack for additional nested teams.
3253	KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3254	// for SERIALIZED teams nested 2 or more levels deep
3255	// typed flag to store request state of cancellation
3256	std::atomic<kmp_int32> t_cancel_request;
3257	int t_master_active; // save on fork, restore on join
3258	void *t_copypriv_data; // team specific pointer to copyprivate data array
3259	#if KMP_OS_WINDOWS
3260	std::atomic<kmp_uint32> t_copyin_counter;
3261	#endif
3262	#if USE_ITT_BUILD
3263	void *t_stack_id; // team specific stack stitching id (for ittnotify)
3264	#endif /* USE_ITT_BUILD */
3265	distributedBarrier *b; // Distributed barrier data associated with team
3266	kmp_nested_nthreads_t *t_nested_nth;
3267	} kmp_base_team_t;
3268
3269	// Assert that the list structure fits and aligns within
3270	// the double task team pointer
3271	KMP_BUILD_ASSERT(sizeof(kmp_task_team_t *[2]) == sizeof(kmp_task_team_list_t));
3272	KMP_BUILD_ASSERT(alignof(kmp_task_team_t *[2]) ==
3273	alignof(kmp_task_team_list_t));
3274
3275	union KMP_ALIGN_CACHE kmp_team {
3276	kmp_base_team_t t;
3277	double t_align; /* use worst case alignment */
3278	char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3279	};
3280
3281	typedef union KMP_ALIGN_CACHE kmp_time_global {
3282	double dt_align; /* use worst case alignment */
3283	char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3284	kmp_base_data_t dt;
3285	} kmp_time_global_t;
3286
3287	typedef struct kmp_base_global {
3288	/* cache-aligned */
3289	kmp_time_global_t g_time;
3290
3291	/* non cache-aligned */
3292	volatile int g_abort;
3293	volatile int g_done;
3294
3295	int g_dynamic;
3296	enum dynamic_mode g_dynamic_mode;
3297	} kmp_base_global_t;
3298
3299	typedef union KMP_ALIGN_CACHE kmp_global {
3300	kmp_base_global_t g;
3301	double g_align; /* use worst case alignment */
3302	char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3303	} kmp_global_t;
3304
3305	typedef struct kmp_base_root {
3306	// TODO: GEH - combine r_active with r_in_parallel then r_active ==
3307	// (r_in_parallel>= 0)
3308	// TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3309	// the synch overhead or keeping r_active
3310	volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3311	// keeps a count of active parallel regions per root
3312	std::atomic<int> r_in_parallel;
3313	// GEH: This is misnamed, should be r_active_levels
3314	kmp_team_t *r_root_team;
3315	kmp_team_t *r_hot_team;
3316	kmp_info_t *r_uber_thread;
3317	kmp_lock_t r_begin_lock;
3318	volatile int r_begin;
3319	int r_blocktime; /* blocktime for this root and descendants */
3320	#if KMP_AFFINITY_SUPPORTED
3321	int r_affinity_assigned;
3322	#endif // KMP_AFFINITY_SUPPORTED
3323	} kmp_base_root_t;
3324
3325	typedef union KMP_ALIGN_CACHE kmp_root {
3326	kmp_base_root_t r;
3327	double r_align; /* use worst case alignment */
3328	char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3329	} kmp_root_t;
3330
3331	struct fortran_inx_info {
3332	kmp_int32 data;
3333	};
3334
3335	// This list type exists to hold old __kmp_threads arrays so that
3336	// old references to them may complete while reallocation takes place when
3337	// expanding the array. The items in this list are kept alive until library
3338	// shutdown.
3339	typedef struct kmp_old_threads_list_t {
3340	kmp_info_t **threads;
3341	struct kmp_old_threads_list_t *next;
3342	} kmp_old_threads_list_t;
3343
3344	/* ------------------------------------------------------------------------ */
3345
3346	extern int __kmp_settings;
3347	extern int __kmp_duplicate_library_ok;
3348	#if USE_ITT_BUILD
3349	extern int __kmp_forkjoin_frames;
3350	extern int __kmp_forkjoin_frames_mode;
3351	#endif
3352	extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3353	extern int __kmp_determ_red;
3354
3355	#ifdef KMP_DEBUG
3356	extern int kmp_a_debug;
3357	extern int kmp_b_debug;
3358	extern int kmp_c_debug;
3359	extern int kmp_d_debug;
3360	extern int kmp_e_debug;
3361	extern int kmp_f_debug;
3362	#endif /* KMP_DEBUG */
3363
3364	/* For debug information logging using rotating buffer */
3365	#define KMP_DEBUG_BUF_LINES_INIT 512
3366	#define KMP_DEBUG_BUF_LINES_MIN 1
3367
3368	#define KMP_DEBUG_BUF_CHARS_INIT 128
3369	#define KMP_DEBUG_BUF_CHARS_MIN 2
3370
3371	extern int
3372	__kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3373	extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3374	extern int
3375	__kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3376	extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3377	entry pointer */
3378
3379	extern char *__kmp_debug_buffer; /* Debug buffer itself */
3380	extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3381	printed in buffer so far */
3382	extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3383	recommended in warnings */
3384	/* end rotating debug buffer */
3385
3386	#ifdef KMP_DEBUG
3387	extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3388
3389	#define KMP_PAR_RANGE_ROUTINE_LEN 1024
3390	extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3391	#define KMP_PAR_RANGE_FILENAME_LEN 1024
3392	extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3393	extern int __kmp_par_range_lb;
3394	extern int __kmp_par_range_ub;
3395	#endif
3396
3397	/* For printing out dynamic storage map for threads and teams */
3398	extern int
3399	__kmp_storage_map; /* True means print storage map for threads and teams */
3400	extern int __kmp_storage_map_verbose; /* True means storage map includes
3401	placement info */
3402	extern int __kmp_storage_map_verbose_specified;
3403
3404	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3405	extern kmp_cpuinfo_t __kmp_cpuinfo;
3406	static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3407	#elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3408	static inline bool __kmp_is_hybrid_cpu() { return true; }
3409	#else
3410	static inline bool __kmp_is_hybrid_cpu() { return false; }
3411	#endif
3412
3413	extern volatile int __kmp_init_serial;
3414	extern volatile int __kmp_init_gtid;
3415	extern volatile int __kmp_init_common;
3416	extern volatile int __kmp_need_register_serial;
3417	extern volatile int __kmp_init_middle;
3418	extern volatile int __kmp_init_parallel;
3419	#if KMP_USE_MONITOR
3420	extern volatile int __kmp_init_monitor;
3421	#endif
3422	extern volatile int __kmp_init_user_locks;
3423	extern volatile int __kmp_init_hidden_helper_threads;
3424	extern int __kmp_init_counter;
3425	extern int __kmp_root_counter;
3426	extern int __kmp_version;
3427
3428	/* list of address of allocated caches for commons */
3429	extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3430
3431	/* Barrier algorithm types and options */
3432	extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3433	extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3434	extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3435	extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3436	extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3437	extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3438	extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3439	extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3440	extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3441	extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3442	extern char const *__kmp_barrier_type_name[bs_last_barrier];
3443	extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3444
3445	/* Global Locks */
3446	extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3447	extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3448	extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3449	extern kmp_bootstrap_lock_t
3450	__kmp_exit_lock; /* exit() is not always thread-safe */
3451	#if KMP_USE_MONITOR
3452	extern kmp_bootstrap_lock_t
3453	__kmp_monitor_lock; /* control monitor thread creation */
3454	#endif
3455	extern kmp_bootstrap_lock_t
3456	__kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3457	__kmp_threads expansion to co-exist */
3458
3459	extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3460
3461	extern enum library_type __kmp_library;
3462
3463	extern enum sched_type __kmp_sched; /* default runtime scheduling */
3464	extern enum sched_type __kmp_static; /* default static scheduling method */
3465	extern enum sched_type __kmp_guided; /* default guided scheduling method */
3466	extern enum sched_type __kmp_auto; /* default auto scheduling method */
3467	extern int __kmp_chunk; /* default runtime chunk size */
3468	extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3469
3470	extern size_t __kmp_stksize; /* stack size per thread */
3471	#if KMP_USE_MONITOR
3472	extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3473	#endif
3474	extern size_t __kmp_stkoffset; /* stack offset per thread */
3475	extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3476
3477	extern size_t
3478	__kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3479	extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3480	extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3481	extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3482	extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3483	extern int __kmp_generate_warnings; /* should we issue warnings? */
3484	extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3485
3486	#ifdef DEBUG_SUSPEND
3487	extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3488	#endif
3489
3490	extern kmp_int32 __kmp_use_yield;
3491	extern kmp_int32 __kmp_use_yield_exp_set;
3492	extern kmp_uint32 __kmp_yield_init;
3493	extern kmp_uint32 __kmp_yield_next;
3494	extern kmp_uint64 __kmp_pause_init;
3495
3496	/* ------------------------------------------------------------------------- */
3497	extern int __kmp_allThreadsSpecified;
3498
3499	extern size_t __kmp_align_alloc;
3500	/* following data protected by initialization routines */
3501	extern int __kmp_xproc; /* number of processors in the system */
3502	extern int __kmp_avail_proc; /* number of processors available to the process */
3503	extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3504	extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3505	// maximum total number of concurrently-existing threads on device
3506	extern int __kmp_max_nth;
3507	// maximum total number of concurrently-existing threads in a contention group
3508	extern int __kmp_cg_max_nth;
3509	extern int __kmp_task_max_nth; // max threads used in a task
3510	extern int __kmp_teams_max_nth; // max threads used in a teams construct
3511	extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3512	__kmp_root */
3513	extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3514	region a la OMP_NUM_THREADS */
3515	extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3516	initialization */
3517	extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3518	used (fixed) */
3519	extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3520	(__kmpc_threadprivate_cached()) */
3521	extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3522	blocking (env setting) */
3523	extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3524	extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3525
3526	// Convert raw blocktime from ms to us if needed.
3527	static inline void __kmp_aux_convert_blocktime(int *bt) {
3528	if (__kmp_blocktime_units == 'm') {
3529	if (*bt > INT_MAX / 1000) {
3530	*bt = INT_MAX / 1000;
3531	KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3532	}
3533	*bt = *bt * 1000;
3534	}
3535	}
3536
3537	#if KMP_USE_MONITOR
3538	extern int
3539	__kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3540	extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3541	blocking */
3542	#endif
3543	#ifdef KMP_ADJUST_BLOCKTIME
3544	extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3545	#endif /* KMP_ADJUST_BLOCKTIME */
3546	#ifdef KMP_DFLT_NTH_CORES
3547	extern int __kmp_ncores; /* Total number of cores for threads placement */
3548	#endif
3549	/* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3550	extern int __kmp_abort_delay;
3551
3552	extern int __kmp_need_register_atfork_specified;
3553	extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3554	to install fork handler */
3555	extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3556	0 - not set, will be set at runtime
3557	1 - using stack search
3558	2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3559	X*) or TlsGetValue(Windows* OS))
3560	3 - static TLS (__declspec(thread) __kmp_gtid),
3561	Linux* OS .so only. */
3562	extern int
3563	__kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3564	#ifdef KMP_TDATA_GTID
3565	extern KMP_THREAD_LOCAL int __kmp_gtid;
3566	#endif
3567	extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3568	extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3569	#if KMP_ARCH_X86 || KMP_ARCH_X86_64
3570	extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3571	extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3572	extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3573	#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3574
3575	// max_active_levels for nested parallelism enabled by default via
3576	// OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3577	extern int __kmp_dflt_max_active_levels;
3578	// Indicates whether value of __kmp_dflt_max_active_levels was already
3579	// explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3580	extern bool __kmp_dflt_max_active_levels_set;
3581	extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3582	concurrent execution per team */
3583	#if KMP_NESTED_HOT_TEAMS
3584	extern int __kmp_hot_teams_mode;
3585	extern int __kmp_hot_teams_max_level;
3586	#endif
3587
3588	#if KMP_MIC_SUPPORTED
3589	extern enum mic_type __kmp_mic_type;
3590	#endif
3591
3592	#ifdef USE_LOAD_BALANCE
3593	extern double __kmp_load_balance_interval; // load balance algorithm interval
3594	#endif /* USE_LOAD_BALANCE */
3595
3596	#if KMP_USE_ADAPTIVE_LOCKS
3597
3598	// Parameters for the speculative lock backoff system.
3599	struct kmp_adaptive_backoff_params_t {
3600	// Number of soft retries before it counts as a hard retry.
3601	kmp_uint32 max_soft_retries;
3602	// Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3603	// the right
3604	kmp_uint32 max_badness;
3605	};
3606
3607	extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3608
3609	#if KMP_DEBUG_ADAPTIVE_LOCKS
3610	extern const char *__kmp_speculative_statsfile;
3611	#endif
3612
3613	#endif // KMP_USE_ADAPTIVE_LOCKS
3614
3615	extern int __kmp_display_env; /* TRUE or FALSE */
3616	extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3617	extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3618	extern int __kmp_nteams;
3619	extern int __kmp_teams_thread_limit;
3620
3621	/* ------------------------------------------------------------------------- */
3622
3623	/* the following are protected by the fork/join lock */
3624	/* write: lock read: anytime */
3625	extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3626	/* Holds old arrays of __kmp_threads until library shutdown */
3627	extern kmp_old_threads_list_t *__kmp_old_threads_list;
3628	/* read/write: lock */
3629	extern volatile kmp_team_t *__kmp_team_pool;
3630	extern volatile kmp_info_t *__kmp_thread_pool;
3631	extern kmp_info_t *__kmp_thread_pool_insert_pt;
3632
3633	// total num threads reachable from some root thread including all root threads
3634	extern volatile int __kmp_nth;
3635	/* total number of threads reachable from some root thread including all root
3636	threads, and those in the thread pool */
3637	extern volatile int __kmp_all_nth;
3638	extern std::atomic<int> __kmp_thread_pool_active_nth;
3639
3640	extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3641	/* end data protected by fork/join lock */
3642	/* ------------------------------------------------------------------------- */
3643
3644	#define __kmp_get_gtid() __kmp_get_global_thread_id()
3645	#define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3646	#define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3647	#define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3648	#define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3649
3650	// AT: Which way is correct?
3651	// AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3652	// AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3653	#define __kmp_get_team_num_threads(gtid) \
3654	(__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3655
3656	static inline bool KMP_UBER_GTID(int gtid) {
3657	KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3658	KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3659	return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3660	__kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3661	}
3662
3663	static inline int __kmp_tid_from_gtid(int gtid) {
3664	KMP_DEBUG_ASSERT(gtid >= 0);
3665	return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3666	}
3667
3668	static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3669	KMP_DEBUG_ASSERT(tid >= 0 && team);
3670	return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3671	}
3672
3673	static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3674	KMP_DEBUG_ASSERT(thr);
3675	return thr->th.th_info.ds.ds_gtid;
3676	}
3677
3678	static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3679	KMP_DEBUG_ASSERT(gtid >= 0);
3680	return __kmp_threads[gtid];
3681	}
3682
3683	static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3684	KMP_DEBUG_ASSERT(gtid >= 0);
3685	return __kmp_threads[gtid]->th.th_team;
3686	}
3687
3688	static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3689	if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3690	KMP_FATAL(ThreadIdentInvalid);
3691	}
3692
3693	#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3694	extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3695	extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3696	extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3697	extern int __kmp_mwait_hints; // Hints to pass in to mwait
3698	#endif
3699
3700	#if KMP_HAVE_UMWAIT
3701	extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3702	extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3703	extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3704	extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3705	#endif
3706
3707	/* ------------------------------------------------------------------------- */
3708
3709	extern kmp_global_t __kmp_global; /* global status */
3710
3711	extern kmp_info_t __kmp_monitor;
3712	// For Debugging Support Library
3713	extern std::atomic<kmp_int32> __kmp_team_counter;
3714	// For Debugging Support Library
3715	extern std::atomic<kmp_int32> __kmp_task_counter;
3716
3717	#if USE_DEBUGGER
3718	#define _KMP_GEN_ID(counter) \
3719	(__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3720	#else
3721	#define _KMP_GEN_ID(counter) (~0)
3722	#endif /* USE_DEBUGGER */
3723
3724	#define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3725	#define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3726
3727	/* ------------------------------------------------------------------------ */
3728
3729	extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3730	size_t size, char const *format, ...);
3731
3732	extern void __kmp_serial_initialize(void);
3733	extern void __kmp_middle_initialize(void);
3734	extern void __kmp_parallel_initialize(void);
3735
3736	extern void __kmp_internal_begin(void);
3737	extern void __kmp_internal_end_library(int gtid);
3738	extern void __kmp_internal_end_thread(int gtid);
3739	extern void __kmp_internal_end_atexit(void);
3740	extern void __kmp_internal_end_dtor(void);
3741	extern void __kmp_internal_end_dest(void *);
3742
3743	extern int __kmp_register_root(int initial_thread);
3744	extern void __kmp_unregister_root(int gtid);
3745	extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3746
3747	extern int __kmp_ignore_mppbeg(void);
3748	extern int __kmp_ignore_mppend(void);
3749
3750	extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3751	extern void __kmp_exit_single(int gtid);
3752
3753	extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3754	extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3755
3756	#ifdef USE_LOAD_BALANCE
3757	extern int __kmp_get_load_balance(int);
3758	#endif
3759
3760	extern int __kmp_get_global_thread_id(void);
3761	extern int __kmp_get_global_thread_id_reg(void);
3762	extern void __kmp_exit_thread(int exit_status);
3763	extern void __kmp_abort(char const *format, ...);
3764	extern void __kmp_abort_thread(void);
3765	KMP_NORETURN extern void __kmp_abort_process(void);
3766	extern void __kmp_warn(char const *format, ...);
3767
3768	extern void __kmp_set_num_threads(int new_nth, int gtid);
3769
3770	extern bool __kmp_detect_shm();
3771	extern bool __kmp_detect_tmp();
3772
3773	// Returns current thread (pointer to kmp_info_t). Current thread *must* be
3774	// registered.
3775	static inline kmp_info_t *__kmp_entry_thread() {
3776	int gtid = __kmp_entry_gtid();
3777
3778	return __kmp_threads[gtid];
3779	}
3780
3781	extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3782	extern int __kmp_get_max_active_levels(int gtid);
3783	extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3784	extern int __kmp_get_team_size(int gtid, int level);
3785	extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3786	extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3787
3788	extern unsigned short __kmp_get_random(kmp_info_t *thread);
3789	extern void __kmp_init_random(kmp_info_t *thread);
3790
3791	extern kmp_r_sched_t __kmp_get_schedule_global(void);
3792	extern void __kmp_adjust_num_threads(int new_nproc);
3793	extern void __kmp_check_stksize(size_t *val);
3794
3795	extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3796	extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3797	extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3798	#define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3799	#define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3800	#define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3801
3802	#if USE_FAST_MEMORY
3803	extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3804	size_t size KMP_SRC_LOC_DECL);
3805	extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3806	extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3807	extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3808	#define __kmp_fast_allocate(this_thr, size) \
3809	___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3810	#define __kmp_fast_free(this_thr, ptr) \
3811	___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3812	#endif
3813
3814	extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3815	extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3816	size_t elsize KMP_SRC_LOC_DECL);
3817	extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3818	size_t size KMP_SRC_LOC_DECL);
3819	extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3820	#define __kmp_thread_malloc(th, size) \
3821	___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3822	#define __kmp_thread_calloc(th, nelem, elsize) \
3823	___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3824	#define __kmp_thread_realloc(th, ptr, size) \
3825	___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3826	#define __kmp_thread_free(th, ptr) \
3827	___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3828
3829	extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3830	extern void __kmp_push_num_threads_list(ident_t *loc, int gtid,
3831	kmp_uint32 list_length,
3832	int *num_threads_list);
3833	extern void __kmp_set_strict_num_threads(ident_t *loc, int gtid, int sev,
3834	const char *msg);
3835
3836	extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3837	kmp_proc_bind_t proc_bind);
3838	extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3839	int num_threads);
3840	extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3841	int num_teams_ub, int num_threads);
3842
3843	extern void __kmp_yield();
3844
3845	extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3846	enum sched_type schedule, kmp_int32 lb,
3847	kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3848	extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3849	enum sched_type schedule, kmp_uint32 lb,
3850	kmp_uint32 ub, kmp_int32 st,
3851	kmp_int32 chunk);
3852	extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3853	enum sched_type schedule, kmp_int64 lb,
3854	kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3855	extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3856	enum sched_type schedule, kmp_uint64 lb,
3857	kmp_uint64 ub, kmp_int64 st,
3858	kmp_int64 chunk);
3859
3860	extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3861	kmp_int32 *p_last, kmp_int32 *p_lb,
3862	kmp_int32 *p_ub, kmp_int32 *p_st);
3863	extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3864	kmp_int32 *p_last, kmp_uint32 *p_lb,
3865	kmp_uint32 *p_ub, kmp_int32 *p_st);
3866	extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3867	kmp_int32 *p_last, kmp_int64 *p_lb,
3868	kmp_int64 *p_ub, kmp_int64 *p_st);
3869	extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3870	kmp_int32 *p_last, kmp_uint64 *p_lb,
3871	kmp_uint64 *p_ub, kmp_int64 *p_st);
3872
3873	extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3874	extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3875	extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3876	extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3877
3878	extern void __kmpc_dispatch_deinit(ident_t *loc, kmp_int32 gtid);
3879
3880	#ifdef KMP_GOMP_COMPAT
3881
3882	extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3883	enum sched_type schedule, kmp_int32 lb,
3884	kmp_int32 ub, kmp_int32 st,
3885	kmp_int32 chunk, int push_ws);
3886	extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3887	enum sched_type schedule, kmp_uint32 lb,
3888	kmp_uint32 ub, kmp_int32 st,
3889	kmp_int32 chunk, int push_ws);
3890	extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3891	enum sched_type schedule, kmp_int64 lb,
3892	kmp_int64 ub, kmp_int64 st,
3893	kmp_int64 chunk, int push_ws);
3894	extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3895	enum sched_type schedule, kmp_uint64 lb,
3896	kmp_uint64 ub, kmp_int64 st,
3897	kmp_int64 chunk, int push_ws);
3898	extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3899	extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3900	extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3901	extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3902
3903	#endif /* KMP_GOMP_COMPAT */
3904
3905	extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3906	extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3907	extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3908	extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3909	extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3910	extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3911	kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3912	void *obj);
3913	extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3914	kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3915
3916	extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3917	int final_spin
3918	#if USE_ITT_BUILD
3919	,
3920	void *itt_sync_obj
3921	#endif
3922	);
3923	extern void __kmp_release_64(kmp_flag_64<> *flag);
3924
3925	extern void __kmp_infinite_loop(void);
3926
3927	extern void __kmp_cleanup(void);
3928
3929	#if KMP_HANDLE_SIGNALS
3930	extern int __kmp_handle_signals;
3931	extern void __kmp_install_signals(int parallel_init);
3932	extern void __kmp_remove_signals(void);
3933	#endif
3934
3935	extern void __kmp_clear_system_time(void);
3936	extern void __kmp_read_system_time(double *delta);
3937
3938	extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3939
3940	extern void __kmp_expand_host_name(char *buffer, size_t size);
3941	extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3942
3943	#if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3944	extern void
3945	__kmp_initialize_system_tick(void); /* Initialize timer tick value */
3946	#endif
3947
3948	extern void
3949	__kmp_runtime_initialize(void); /* machine specific initialization */
3950	extern void __kmp_runtime_destroy(void);
3951
3952	#if KMP_AFFINITY_SUPPORTED
3953	extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3954	kmp_affin_mask_t *mask);
3955	extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3956	kmp_affin_mask_t *mask);
3957	extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3958	extern void __kmp_affinity_uninitialize(void);
3959	extern void __kmp_affinity_set_init_mask(
3960	int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3961	void __kmp_affinity_bind_init_mask(int gtid);
3962	extern void __kmp_affinity_bind_place(int gtid);
3963	extern void __kmp_affinity_determine_capable(const char *env_var);
3964	extern int __kmp_aux_set_affinity(void **mask);
3965	extern int __kmp_aux_get_affinity(void **mask);
3966	extern int __kmp_aux_get_affinity_max_proc();
3967	extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3968	extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3969	extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3970	extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3971	#if KMP_WEIGHTED_ITERATIONS_SUPPORTED
3972	extern int __kmp_get_first_osid_with_ecore(void);
3973	#endif
3974	#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3975	KMP_OS_AIX
3976	extern int kmp_set_thread_affinity_mask_initial(void);
3977	#endif
3978	static inline void __kmp_assign_root_init_mask() {
3979	int gtid = __kmp_entry_gtid();
3980	kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3981	if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3982	__kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3983	__kmp_affinity_bind_init_mask(gtid);
3984	r->r.r_affinity_assigned = TRUE;
3985	}
3986	}
3987	static inline void __kmp_reset_root_init_mask(int gtid) {
3988	if (!KMP_AFFINITY_CAPABLE())
3989	return;
3990	kmp_info_t *th = __kmp_threads[gtid];
3991	kmp_root_t *r = th->th.th_root;
3992	if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3993	__kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3994	KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3995	r->r.r_affinity_assigned = FALSE;
3996	}
3997	}
3998	#else /* KMP_AFFINITY_SUPPORTED */
3999	#define __kmp_assign_root_init_mask() /* Nothing */
4000	static inline void __kmp_reset_root_init_mask(int gtid) {}
4001	#endif /* KMP_AFFINITY_SUPPORTED */
4002	// No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
4003	// format string is for affinity, so platforms that do not support
4004	// affinity can still use the other fields, e.g., %n for num_threads
4005	extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
4006	kmp_str_buf_t *buffer);
4007	extern void __kmp_aux_display_affinity(int gtid, const char *format);
4008
4009	extern void __kmp_cleanup_hierarchy();
4010	extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
4011
4012	#if KMP_USE_FUTEX
4013
4014	extern int __kmp_futex_determine_capable(void);
4015
4016	#endif // KMP_USE_FUTEX
4017
4018	extern void __kmp_gtid_set_specific(int gtid);
4019	extern int __kmp_gtid_get_specific(void);
4020
4021	extern double __kmp_read_cpu_time(void);
4022
4023	extern int __kmp_read_system_info(struct kmp_sys_info *info);
4024
4025	#if KMP_USE_MONITOR
4026	extern void __kmp_create_monitor(kmp_info_t *th);
4027	#endif
4028
4029	extern void *__kmp_launch_thread(kmp_info_t *thr);
4030
4031	extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
4032
4033	#if KMP_OS_WINDOWS
4034	extern int __kmp_still_running(kmp_info_t *th);
4035	extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
4036	extern void __kmp_free_handle(kmp_thread_t tHandle);
4037	#endif
4038
4039	#if KMP_USE_MONITOR
4040	extern void __kmp_reap_monitor(kmp_info_t *th);
4041	#endif
4042	extern void __kmp_reap_worker(kmp_info_t *th);
4043	extern void __kmp_terminate_thread(int gtid);
4044
4045	extern int __kmp_try_suspend_mx(kmp_info_t *th);
4046	extern void __kmp_lock_suspend_mx(kmp_info_t *th);
4047	extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
4048
4049	extern void __kmp_elapsed(double *);
4050	extern void __kmp_elapsed_tick(double *);
4051
4052	extern void __kmp_enable(int old_state);
4053	extern void __kmp_disable(int *old_state);
4054
4055	extern void __kmp_thread_sleep(int millis);
4056
4057	extern void __kmp_common_initialize(void);
4058	extern void __kmp_common_destroy(void);
4059	extern void __kmp_common_destroy_gtid(int gtid);
4060
4061	#if KMP_OS_UNIX
4062	extern void __kmp_register_atfork(void);
4063	#endif
4064	extern void __kmp_suspend_initialize(void);
4065	extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
4066	extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
4067
4068	extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4069	int tid);
4070	extern kmp_team_t *
4071	__kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4072	#if OMPT_SUPPORT
4073	ompt_data_t ompt_parallel_data,
4074	#endif
4075	kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
4076	int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
4077	extern void __kmp_free_thread(kmp_info_t *);
4078	extern void __kmp_free_team(kmp_root_t *,
4079	kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
4080	extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
4081
4082	/* ------------------------------------------------------------------------ */
4083
4084	extern void __kmp_initialize_bget(kmp_info_t *th);
4085	extern void __kmp_finalize_bget(kmp_info_t *th);
4086
4087	KMP_EXPORT void *kmpc_malloc(size_t size);
4088	KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
4089	KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
4090	KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
4091	KMP_EXPORT void kmpc_free(void *ptr);
4092
4093	/* declarations for internal use */
4094
4095	extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
4096	size_t reduce_size, void *reduce_data,
4097	void (*reduce)(void *, void *));
4098	extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
4099	extern int __kmp_barrier_gomp_cancel(int gtid);
4100
4101	/*!
4102	* Tell the fork call which compiler generated the fork call, and therefore how
4103	* to deal with the call.
4104	*/
4105	enum fork_context_e {
4106	fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
4107	microtask internally. */
4108	fork_context_intel, /**< Called from Intel generated code. */
4109	fork_context_last
4110	};
4111	extern int __kmp_fork_call(ident_t *loc, int gtid,
4112	enum fork_context_e fork_context, kmp_int32 argc,
4113	microtask_t microtask, launch_t invoker,
4114	kmp_va_list ap);
4115
4116	extern void __kmp_join_call(ident_t *loc, int gtid
4117	#if OMPT_SUPPORT
4118	,
4119	enum fork_context_e fork_context
4120	#endif
4121	,
4122	int exit_teams = 0);
4123
4124	extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
4125	extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
4126	extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
4127	extern int __kmp_invoke_task_func(int gtid);
4128	extern void __kmp_run_before_invoked_task(int gtid, int tid,
4129	kmp_info_t *this_thr,
4130	kmp_team_t *team);
4131	extern void __kmp_run_after_invoked_task(int gtid, int tid,
4132	kmp_info_t *this_thr,
4133	kmp_team_t *team);
4134
4135	// should never have been exported
4136	KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
4137	extern int __kmp_invoke_teams_master(int gtid);
4138	extern void __kmp_teams_master(int gtid);
4139	extern int __kmp_aux_get_team_num();
4140	extern int __kmp_aux_get_num_teams();
4141	extern void __kmp_save_internal_controls(kmp_info_t *thread);
4142	extern void __kmp_user_set_library(enum library_type arg);
4143	extern void __kmp_aux_set_library(enum library_type arg);
4144	extern void __kmp_aux_set_stacksize(size_t arg);
4145	extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
4146	extern void __kmp_aux_set_defaults(char const *str, size_t len);
4147
4148	/* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
4149	void kmpc_set_blocktime(int arg);
4150	void ompc_set_nested(int flag);
4151	void ompc_set_dynamic(int flag);
4152	void ompc_set_num_threads(int arg);
4153
4154	extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
4155	kmp_team_t *team, int tid);
4156	extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
4157	extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4158	kmp_tasking_flags_t *flags,
4159	size_t sizeof_kmp_task_t,
4160	size_t sizeof_shareds,
4161	kmp_routine_entry_t task_entry);
4162	extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4163	kmp_team_t *team, int tid,
4164	int set_curr_task);
4165	extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4166	extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4167
4168	extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4169	int gtid,
4170	kmp_task_t *task);
4171	extern void __kmp_fulfill_event(kmp_event_t *event);
4172
4173	extern void __kmp_free_task_team(kmp_info_t *thread,
4174	kmp_task_team_t *task_team);
4175	extern void __kmp_reap_task_teams(void);
4176	extern void __kmp_push_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4177	extern void __kmp_pop_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4178	extern void __kmp_wait_to_unref_task_teams(void);
4179	extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team);
4180	extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4181	extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4182	#if USE_ITT_BUILD
4183	,
4184	void *itt_sync_obj
4185	#endif /* USE_ITT_BUILD */
4186	,
4187	int wait = 1);
4188	extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4189	int gtid);
4190	#if KMP_DEBUG
4191	#define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr) \
4192	KMP_DEBUG_ASSERT( \
4193	__kmp_tasking_mode != tskm_task_teams || team->t.t_nproc == 1 || \
4194	thr->th.th_task_team == team->t.t_task_team[thr->th.th_task_state])
4195	#else
4196	#define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr) /* Nothing */
4197	#endif
4198
4199	extern int __kmp_is_address_mapped(void *addr);
4200	extern kmp_uint64 __kmp_hardware_timestamp(void);
4201
4202	#if KMP_OS_UNIX
4203	extern int __kmp_read_from_file(char const *path, char const *format, ...);
4204	#endif
4205
4206	/* ------------------------------------------------------------------------ */
4207	//
4208	// Assembly routines that have no compiler intrinsic replacement
4209	//
4210
4211	extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4212	void *argv[]
4213	#if OMPT_SUPPORT
4214	,
4215	void **exit_frame_ptr
4216	#endif
4217	);
4218
4219	/* ------------------------------------------------------------------------ */
4220
4221	KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4222	KMP_EXPORT void __kmpc_end(ident_t *);
4223
4224	KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4225	kmpc_ctor_vec ctor,
4226	kmpc_cctor_vec cctor,
4227	kmpc_dtor_vec dtor,
4228	size_t vector_length);
4229	KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4230	kmpc_ctor ctor, kmpc_cctor cctor,
4231	kmpc_dtor dtor);
4232	KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4233	void *data, size_t size);
4234
4235	KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4236	KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4237	KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4238	KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4239
4240	KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4241	KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4242	kmpc_micro microtask, ...);
4243	KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4244	kmpc_micro microtask, kmp_int32 cond,
4245	void *args);
4246
4247	KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4248	KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4249
4250	KMP_EXPORT void __kmpc_flush(ident_t *);
4251	KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4252	KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4253	KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4254	KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4255	kmp_int32 filter);
4256	KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4257	KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4258	KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4259	KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4260	kmp_critical_name *);
4261	KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4262	kmp_critical_name *);
4263	KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4264	kmp_critical_name *, uint32_t hint);
4265
4266	KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4267	KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4268
4269	KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4270	kmp_int32 global_tid);
4271
4272	KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4273	KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4274
4275	KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4276	KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4277	kmp_int32 numberOfSections);
4278	KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4279
4280	KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4281	kmp_int32 schedtype, kmp_int32 *plastiter,
4282	kmp_int *plower, kmp_int *pupper,
4283	kmp_int *pstride, kmp_int incr,
4284	kmp_int chunk);
4285
4286	KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4287
4288	KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4289	size_t cpy_size, void *cpy_data,
4290	void (*cpy_func)(void *, void *),
4291	kmp_int32 didit);
4292
4293	KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4294	void *cpy_data);
4295
4296	extern void KMPC_SET_NUM_THREADS(int arg);
4297	extern void KMPC_SET_DYNAMIC(int flag);
4298	extern void KMPC_SET_NESTED(int flag);
4299
4300	/* OMP 3.0 tasking interface routines */
4301	KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4302	kmp_task_t *new_task);
4303	KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4304	kmp_int32 flags,
4305	size_t sizeof_kmp_task_t,
4306	size_t sizeof_shareds,
4307	kmp_routine_entry_t task_entry);
4308	KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4309	ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4310	size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4311	KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4312	kmp_task_t *task);
4313	KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4314	kmp_task_t *task);
4315	KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4316	kmp_task_t *new_task);
4317	KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4318	KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4319	int end_part);
4320
4321	#if TASK_UNUSED
4322	void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4323	void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4324	kmp_task_t *task);
4325	#endif // TASK_UNUSED
4326
4327	/* ------------------------------------------------------------------------ */
4328
4329	KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4330	KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4331
4332	KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4333	ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4334	kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4335	kmp_depend_info_t *noalias_dep_list);
4336
4337	KMP_EXPORT kmp_base_depnode_t *__kmpc_task_get_depnode(kmp_task_t *task);
4338
4339	KMP_EXPORT kmp_depnode_list_t *__kmpc_task_get_successors(kmp_task_t *task);
4340
4341	KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4342	kmp_int32 ndeps,
4343	kmp_depend_info_t *dep_list,
4344	kmp_int32 ndeps_noalias,
4345	kmp_depend_info_t *noalias_dep_list);
4346	/* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4347	* Placeholder for taskwait with nowait clause.*/
4348	KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4349	kmp_int32 ndeps,
4350	kmp_depend_info_t *dep_list,
4351	kmp_int32 ndeps_noalias,
4352	kmp_depend_info_t *noalias_dep_list,
4353	kmp_int32 has_no_wait);
4354
4355	extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4356	bool serialize_immediate);
4357
4358	KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4359	kmp_int32 cncl_kind);
4360	KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4361	kmp_int32 cncl_kind);
4362	KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4363	KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4364
4365	KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4366	KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4367	KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4368	kmp_int32 if_val, kmp_uint64 *lb,
4369	kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4370	kmp_int32 sched, kmp_uint64 grainsize,
4371	void *task_dup);
4372	KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4373	kmp_task_t *task, kmp_int32 if_val,
4374	kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4375	kmp_int32 nogroup, kmp_int32 sched,
4376	kmp_uint64 grainsize, kmp_int32 modifier,
4377	void *task_dup);
4378	KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4379	KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4380	KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4381	KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4382	int is_ws, int num,
4383	void *data);
4384	KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4385	int num, void *data);
4386	KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4387	int is_ws);
4388	KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4389	ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4390	kmp_task_affinity_info_t *affin_list);
4391	KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4392	KMP_EXPORT int __kmp_get_max_teams(void);
4393	KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4394	KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4395
4396	/* Interface target task integration */
4397	KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4398	KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4399
4400	/* Lock interface routines (fast versions with gtid passed in) */
4401	KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4402	void **user_lock);
4403	KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4404	void **user_lock);
4405	KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4406	void **user_lock);
4407	KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4408	void **user_lock);
4409	KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4410	KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4411	void **user_lock);
4412	KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4413	void **user_lock);
4414	KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4415	void **user_lock);
4416	KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4417	KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4418	void **user_lock);
4419
4420	KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4421	void **user_lock, uintptr_t hint);
4422	KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4423	void **user_lock,
4424	uintptr_t hint);
4425
4426	#if OMPX_TASKGRAPH
4427	// Taskgraph's Record & Replay mechanism
4428	// __kmp_tdg_is_recording: check whether a given TDG is recording
4429	// status: the tdg's current status
4430	static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4431	return status == KMP_TDG_RECORDING;
4432	}
4433
4434	KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4435	kmp_int32 input_flags,
4436	kmp_int32 tdg_id);
4437	KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4438	kmp_int32 input_flags, kmp_int32 tdg_id);
4439	#endif
4440	/* Interface to fast scalable reduce methods routines */
4441
4442	KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4443	ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4444	void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4445	kmp_critical_name *lck);
4446	KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4447	kmp_critical_name *lck);
4448	KMP_EXPORT kmp_int32 __kmpc_reduce(
4449	ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4450	void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4451	kmp_critical_name *lck);
4452	KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4453	kmp_critical_name *lck);
4454
4455	/* Internal fast reduction routines */
4456
4457	extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4458	ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4459	void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4460	kmp_critical_name *lck);
4461
4462	// this function is for testing set/get/determine reduce method
4463	KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4464
4465	KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4466	KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4467
4468	// C++ port
4469	// missing 'extern "C"' declarations
4470
4471	KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4472	KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4473	KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4474	kmp_int32 num_threads);
4475	KMP_EXPORT void __kmpc_push_num_threads_strict(ident_t *loc,
4476	kmp_int32 global_tid,
4477	kmp_int32 num_threads,
4478	int severity,
4479	const char *message);
4480
4481	KMP_EXPORT void __kmpc_push_num_threads_list(ident_t *loc, kmp_int32 global_tid,
4482	kmp_uint32 list_length,
4483	kmp_int32 *num_threads_list);
4484	KMP_EXPORT void __kmpc_push_num_threads_list_strict(
4485	ident_t *loc, kmp_int32 global_tid, kmp_uint32 list_length,
4486	kmp_int32 *num_threads_list, int severity, const char *message);
4487
4488	KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4489	int proc_bind);
4490	KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4491	kmp_int32 num_teams,
4492	kmp_int32 num_threads);
4493	KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4494	kmp_int32 thread_limit);
4495	/* Function for OpenMP 5.1 num_teams clause */
4496	KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4497	kmp_int32 num_teams_lb,
4498	kmp_int32 num_teams_ub,
4499	kmp_int32 num_threads);
4500	KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4501	kmpc_micro microtask, ...);
4502	struct kmp_dim { // loop bounds info casted to kmp_int64
4503	kmp_int64 lo; // lower
4504	kmp_int64 up; // upper
4505	kmp_int64 st; // stride
4506	};
4507	KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4508	kmp_int32 num_dims,
4509	const struct kmp_dim *dims);
4510	KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4511	const kmp_int64 *vec);
4512	KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4513	const kmp_int64 *vec);
4514	KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4515
4516	KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4517	void *data, size_t size,
4518	void ***cache);
4519
4520	// The routines below are not exported.
4521	// Consider making them 'static' in corresponding source files.
4522	void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4523	void *data_addr, size_t pc_size);
4524	struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4525	void *data_addr,
4526	size_t pc_size);
4527	void __kmp_threadprivate_resize_cache(int newCapacity);
4528	void __kmp_cleanup_threadprivate_caches();
4529
4530	// ompc_, kmpc_ entries moved from omp.h.
4531	#if KMP_OS_WINDOWS
4532	#define KMPC_CONVENTION __cdecl
4533	#else
4534	#define KMPC_CONVENTION
4535	#endif
4536
4537	#ifndef __OMP_H
4538	typedef enum omp_sched_t {
4539	omp_sched_static = 1,
4540	omp_sched_dynamic = 2,
4541	omp_sched_guided = 3,
4542	omp_sched_auto = 4
4543	} omp_sched_t;
4544	typedef void *kmp_affinity_mask_t;
4545	#endif
4546
4547	KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4548	KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4549	KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4550	KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4551	KMP_EXPORT int KMPC_CONVENTION
4552	kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4553	KMP_EXPORT int KMPC_CONVENTION
4554	kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4555	KMP_EXPORT int KMPC_CONVENTION
4556	kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4557
4558	KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4559	KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4560	KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4561	KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4562	KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4563	void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4564	size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4565	void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4566	size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4567	char const *format);
4568
4569	enum kmp_target_offload_kind {
4570	tgt_disabled = 0,
4571	tgt_default = 1,
4572	tgt_mandatory = 2
4573	};
4574	typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4575	// Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4576	extern kmp_target_offload_kind_t __kmp_target_offload;
4577	extern int __kmpc_get_target_offload();
4578
4579	// Constants used in libomptarget
4580	#define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4581	#define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4582
4583	// OMP Pause Resource
4584
4585	// The following enum is used both to set the status in __kmp_pause_status, and
4586	// as the internal equivalent of the externally-visible omp_pause_resource_t.
4587	typedef enum kmp_pause_status_t {
4588	kmp_not_paused = 0, // status is not paused, or, requesting resume
4589	kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4590	kmp_hard_paused = 2, // status is hard-paused, or, requesting hard pause
4591	kmp_stop_tool_paused = 3 // requesting stop_tool pause
4592	} kmp_pause_status_t;
4593
4594	// This stores the pause state of the runtime
4595	extern kmp_pause_status_t __kmp_pause_status;
4596	extern int __kmpc_pause_resource(kmp_pause_status_t level);
4597	extern int __kmp_pause_resource(kmp_pause_status_t level);
4598	// Soft resume sets __kmp_pause_status, and wakes up all threads.
4599	extern void __kmp_resume_if_soft_paused();
4600	// Hard resume simply resets the status to not paused. Library will appear to
4601	// be uninitialized after hard pause. Let OMP constructs trigger required
4602	// initializations.
4603	static inline void __kmp_resume_if_hard_paused() {
4604	if (__kmp_pause_status == kmp_hard_paused) {
4605	__kmp_pause_status = kmp_not_paused;
4606	}
4607	}
4608
4609	extern void __kmp_omp_display_env(int verbose);
4610
4611	// 1: it is initializing hidden helper team
4612	extern volatile int __kmp_init_hidden_helper;
4613	// 1: the hidden helper team is done
4614	extern volatile int __kmp_hidden_helper_team_done;
4615	// 1: enable hidden helper task
4616	extern kmp_int32 __kmp_enable_hidden_helper;
4617	// Main thread of hidden helper team
4618	extern kmp_info_t *__kmp_hidden_helper_main_thread;
4619	// Descriptors for the hidden helper threads
4620	extern kmp_info_t **__kmp_hidden_helper_threads;
4621	// Number of hidden helper threads
4622	extern kmp_int32 __kmp_hidden_helper_threads_num;
4623	// Number of hidden helper tasks that have not been executed yet
4624	extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4625
4626	extern void __kmp_hidden_helper_initialize();
4627	extern void __kmp_hidden_helper_threads_initz_routine();
4628	extern void __kmp_do_initialize_hidden_helper_threads();
4629	extern void __kmp_hidden_helper_threads_initz_wait();
4630	extern void __kmp_hidden_helper_initz_release();
4631	extern void __kmp_hidden_helper_threads_deinitz_wait();
4632	extern void __kmp_hidden_helper_threads_deinitz_release();
4633	extern void __kmp_hidden_helper_main_thread_wait();
4634	extern void __kmp_hidden_helper_worker_thread_wait();
4635	extern void __kmp_hidden_helper_worker_thread_signal();
4636	extern void __kmp_hidden_helper_main_thread_release();
4637
4638	// Check whether a given thread is a hidden helper thread
4639	#define KMP_HIDDEN_HELPER_THREAD(gtid) \
4640	((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4641
4642	#define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4643	((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4644
4645	#define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4646	((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4647
4648	#define KMP_HIDDEN_HELPER_TEAM(team) \
4649	(team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4650
4651	// Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4652	// main thread, is skipped.
4653	#define KMP_GTID_TO_SHADOW_GTID(gtid) \
4654	((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4655
4656	// Return the adjusted gtid value by subtracting from gtid the number
4657	// of hidden helper threads. This adjusted value is the gtid the thread would
4658	// have received if there were no hidden helper threads.
4659	static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4660	int adjusted_gtid = gtid;
4661	if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4662	gtid - __kmp_hidden_helper_threads_num >= 0) {
4663	adjusted_gtid -= __kmp_hidden_helper_threads_num;
4664	}
4665	return adjusted_gtid;
4666	}
4667
4668	// Support for error directive
4669	typedef enum kmp_severity_t {
4670	severity_warning = 1,
4671	severity_fatal = 2
4672	} kmp_severity_t;
4673	extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4674
4675	// Support for scope directive
4676	KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4677	KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4678
4679	#ifdef __cplusplus
4680	}
4681	#endif
4682
4683	template <bool C, bool S>
4684	extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4685	template <bool C, bool S>
4686	extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4687	template <bool C, bool S>
4688	extern void __kmp_atomic_suspend_64(int th_gtid,
4689	kmp_atomic_flag_64<C, S> *flag);
4690	extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4691	#if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4692	template <bool C, bool S>
4693	extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4694	template <bool C, bool S>
4695	extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4696	template <bool C, bool S>
4697	extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4698	extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4699	#endif
4700	template <bool C, bool S>
4701	extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4702	template <bool C, bool S>
4703	extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4704	template <bool C, bool S>
4705	extern void __kmp_atomic_resume_64(int target_gtid,
4706	kmp_atomic_flag_64<C, S> *flag);
4707	extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4708
4709	template <bool C, bool S>
4710	int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4711	kmp_flag_32<C, S> *flag, int final_spin,
4712	int *thread_finished,
4713	#if USE_ITT_BUILD
4714	void *itt_sync_obj,
4715	#endif /* USE_ITT_BUILD */
4716	kmp_int32 is_constrained);
4717	template <bool C, bool S>
4718	int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4719	kmp_flag_64<C, S> *flag, int final_spin,
4720	int *thread_finished,
4721	#if USE_ITT_BUILD
4722	void *itt_sync_obj,
4723	#endif /* USE_ITT_BUILD */
4724	kmp_int32 is_constrained);
4725	template <bool C, bool S>
4726	int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4727	kmp_atomic_flag_64<C, S> *flag,
4728	int final_spin, int *thread_finished,
4729	#if USE_ITT_BUILD
4730	void *itt_sync_obj,
4731	#endif /* USE_ITT_BUILD */
4732	kmp_int32 is_constrained);
4733	int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4734	kmp_flag_oncore *flag, int final_spin,
4735	int *thread_finished,
4736	#if USE_ITT_BUILD
4737	void *itt_sync_obj,
4738	#endif /* USE_ITT_BUILD */
4739	kmp_int32 is_constrained);
4740
4741	extern int __kmp_nesting_mode;
4742	extern int __kmp_nesting_mode_nlevels;
4743	extern int *__kmp_nesting_nth_level;
4744	extern void __kmp_init_nesting_mode();
4745	extern void __kmp_set_nesting_mode_threads();
4746
4747	/// This class safely opens and closes a C-style FILE* object using RAII
4748	/// semantics. There are also methods which allow using stdout or stderr as
4749	/// the underlying FILE* object. With the implicit conversion operator to
4750	/// FILE*, an object with this type can be used in any function which takes
4751	/// a FILE* object e.g., fprintf().
4752	/// No close method is needed at use sites.
4753	class kmp_safe_raii_file_t {
4754	FILE *f;
4755
4756	void close() {
4757	if (f && f != stdout && f != stderr) {
4758	fclose(stream: f);
4759	f = nullptr;
4760	}
4761	}
4762
4763	public:
4764	kmp_safe_raii_file_t() : f(nullptr) {}
4765	kmp_safe_raii_file_t(const char *filename, const char *mode,
4766	const char *env_var = nullptr)
4767	: f(nullptr) {
4768	open(filename, mode, env_var);
4769	}
4770	kmp_safe_raii_file_t(const kmp_safe_raii_file_t &other) = delete;
4771	kmp_safe_raii_file_t &operator=(const kmp_safe_raii_file_t &other) = delete;
4772	~kmp_safe_raii_file_t() { close(); }
4773
4774	/// Open filename using mode. This is automatically closed in the destructor.
4775	/// The env_var parameter indicates the environment variable the filename
4776	/// came from if != nullptr.
4777	void open(const char *filename, const char *mode,
4778	const char *env_var = nullptr) {
4779	KMP_ASSERT(!f);
4780	f = fopen(filename: filename, modes: mode);
4781	if (!f) {
4782	int code = errno;
4783	if (env_var) {
4784	__kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4785	KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4786	} else {
4787	__kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4788	__kmp_msg_null);
4789	}
4790	}
4791	}
4792	/// Instead of erroring out, return non-zero when
4793	/// unsuccessful fopen() for any reason
4794	int try_open(const char *filename, const char *mode) {
4795	KMP_ASSERT(!f);
4796	f = fopen(filename: filename, modes: mode);
4797	if (!f)
4798	return errno;
4799	return 0;
4800	}
4801	/// Set the FILE* object to stdout and output there
4802	/// No open call should happen before this call.
4803	void set_stdout() {
4804	KMP_ASSERT(!f);
4805	f = stdout;
4806	}
4807	/// Set the FILE* object to stderr and output there
4808	/// No open call should happen before this call.
4809	void set_stderr() {
4810	KMP_ASSERT(!f);
4811	f = stderr;
4812	}
4813	operator bool() { return bool(f); }
4814	operator FILE *() { return f; }
4815	};
4816
4817	template <typename SourceType, typename TargetType,
4818	bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4819	bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4820	bool isSourceSigned = std::is_signed<SourceType>::value,
4821	bool isTargetSigned = std::is_signed<TargetType>::value>
4822	struct kmp_convert {};
4823
4824	// Both types are signed; Source smaller
4825	template <typename SourceType, typename TargetType>
4826	struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4827	static TargetType to(SourceType src) { return (TargetType)src; }
4828	};
4829	// Source equal
4830	template <typename SourceType, typename TargetType>
4831	struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4832	static TargetType to(SourceType src) { return src; }
4833	};
4834	// Source bigger
4835	template <typename SourceType, typename TargetType>
4836	struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4837	static TargetType to(SourceType src) {
4838	KMP_ASSERT(src <= static_cast<SourceType>(
4839	(std::numeric_limits<TargetType>::max)()));
4840	KMP_ASSERT(src >= static_cast<SourceType>(
4841	(std::numeric_limits<TargetType>::min)()));
4842	return (TargetType)src;
4843	}
4844	};
4845
4846	// Source signed, Target unsigned
4847	// Source smaller
4848	template <typename SourceType, typename TargetType>
4849	struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4850	static TargetType to(SourceType src) {
4851	KMP_ASSERT(src >= 0);
4852	return (TargetType)src;
4853	}
4854	};
4855	// Source equal
4856	template <typename SourceType, typename TargetType>
4857	struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4858	static TargetType to(SourceType src) {
4859	KMP_ASSERT(src >= 0);
4860	return (TargetType)src;
4861	}
4862	};
4863	// Source bigger
4864	template <typename SourceType, typename TargetType>
4865	struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4866	static TargetType to(SourceType src) {
4867	KMP_ASSERT(src >= 0);
4868	KMP_ASSERT(src <= static_cast<SourceType>(
4869	(std::numeric_limits<TargetType>::max)()));
4870	return (TargetType)src;
4871	}
4872	};
4873
4874	// Source unsigned, Target signed
4875	// Source smaller
4876	template <typename SourceType, typename TargetType>
4877	struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4878	static TargetType to(SourceType src) { return (TargetType)src; }
4879	};
4880	// Source equal
4881	template <typename SourceType, typename TargetType>
4882	struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4883	static TargetType to(SourceType src) {
4884	KMP_ASSERT(src <= static_cast<SourceType>(
4885	(std::numeric_limits<TargetType>::max)()));
4886	return (TargetType)src;
4887	}
4888	};
4889	// Source bigger
4890	template <typename SourceType, typename TargetType>
4891	struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4892	static TargetType to(SourceType src) {
4893	KMP_ASSERT(src <= static_cast<SourceType>(
4894	(std::numeric_limits<TargetType>::max)()));
4895	return (TargetType)src;
4896	}
4897	};
4898
4899	// Source unsigned, Target unsigned
4900	// Source smaller
4901	template <typename SourceType, typename TargetType>
4902	struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4903	static TargetType to(SourceType src) { return (TargetType)src; }
4904	};
4905	// Source equal
4906	template <typename SourceType, typename TargetType>
4907	struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4908	static TargetType to(SourceType src) { return src; }
4909	};
4910	// Source bigger
4911	template <typename SourceType, typename TargetType>
4912	struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4913	static TargetType to(SourceType src) {
4914	KMP_ASSERT(src <= static_cast<SourceType>(
4915	(std::numeric_limits<TargetType>::max)()));
4916	return (TargetType)src;
4917	}
4918	};
4919
4920	template <typename T1, typename T2>
4921	static inline void __kmp_type_convert(T1 src, T2 *dest) {
4922	*dest = kmp_convert<T1, T2>::to(src);
4923	}
4924
4925	#endif /* KMP_H */
4926