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