xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp.h (revision 401ab69cff8fa2320a9f8ea4baa114a6da6c952b)
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 
31 #define TASK_CURRENT_NOT_QUEUED 0
32 #define TASK_CURRENT_QUEUED 1
33 
34 #ifdef BUILD_TIED_TASK_STACK
35 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37 // Number of entries in each task stack array
38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39 // Mask for determining index into stack block
40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41 #endif // BUILD_TIED_TASK_STACK
42 
43 #define TASK_NOT_PUSHED 1
44 #define TASK_SUCCESSFULLY_PUSHED 0
45 #define TASK_TIED 1
46 #define TASK_UNTIED 0
47 #define TASK_EXPLICIT 1
48 #define TASK_IMPLICIT 0
49 #define TASK_PROXY 1
50 #define TASK_FULL 0
51 #define TASK_DETACHABLE 1
52 #define TASK_UNDETACHABLE 0
53 
54 #define KMP_CANCEL_THREADS
55 #define KMP_THREAD_ATTR
56 
57 // Android does not have pthread_cancel.  Undefine KMP_CANCEL_THREADS if being
58 // built on Android
59 #if defined(__ANDROID__)
60 #undef KMP_CANCEL_THREADS
61 #endif
62 
63 #include <signal.h>
64 #include <stdarg.h>
65 #include <stddef.h>
66 #include <stdio.h>
67 #include <stdlib.h>
68 #include <string.h>
69 #include <limits>
70 #include <type_traits>
71 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72    Microsoft library. Some macros provided below to replace these functions  */
73 #ifndef __ABSOFT_WIN
74 #include <sys/types.h>
75 #endif
76 #include <limits.h>
77 #include <time.h>
78 
79 #include <errno.h>
80 
81 #include "kmp_os.h"
82 
83 #include "kmp_safe_c_api.h"
84 
85 #if KMP_STATS_ENABLED
86 class kmp_stats_list;
87 #endif
88 
89 #if KMP_USE_HIER_SCHED
90 // Only include hierarchical scheduling if affinity is supported
91 #undef KMP_USE_HIER_SCHED
92 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93 #endif
94 
95 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96 #include "hwloc.h"
97 #ifndef HWLOC_OBJ_NUMANODE
98 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99 #endif
100 #ifndef HWLOC_OBJ_PACKAGE
101 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102 #endif
103 #endif
104 
105 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
106 #include <xmmintrin.h>
107 #endif
108 
109 // The below has to be defined before including "kmp_barrier.h".
110 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
111 #define KMP_INTERNAL_FREE(p) free(p)
112 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
113 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
114 
115 #include "kmp_debug.h"
116 #include "kmp_lock.h"
117 #include "kmp_version.h"
118 #include "kmp_barrier.h"
119 #if USE_DEBUGGER
120 #include "kmp_debugger.h"
121 #endif
122 #include "kmp_i18n.h"
123 
124 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
125 
126 #include "kmp_wrapper_malloc.h"
127 #if KMP_OS_UNIX
128 #include <unistd.h>
129 #if !defined NSIG && defined _NSIG
130 #define NSIG _NSIG
131 #endif
132 #endif
133 
134 #if KMP_OS_LINUX
135 #pragma weak clock_gettime
136 #endif
137 
138 #if OMPT_SUPPORT
139 #include "ompt-internal.h"
140 #endif
141 
142 #if OMPD_SUPPORT
143 #include "ompd-specific.h"
144 #endif
145 
146 #ifndef UNLIKELY
147 #define UNLIKELY(x) (x)
148 #endif
149 
150 // Affinity format function
151 #include "kmp_str.h"
152 
153 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
154 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
155 // free lists of limited size.
156 #ifndef USE_FAST_MEMORY
157 #define USE_FAST_MEMORY 3
158 #endif
159 
160 #ifndef KMP_NESTED_HOT_TEAMS
161 #define KMP_NESTED_HOT_TEAMS 0
162 #define USE_NESTED_HOT_ARG(x)
163 #else
164 #if KMP_NESTED_HOT_TEAMS
165 #define USE_NESTED_HOT_ARG(x) , x
166 #else
167 #define USE_NESTED_HOT_ARG(x)
168 #endif
169 #endif
170 
171 // Assume using BGET compare_exchange instruction instead of lock by default.
172 #ifndef USE_CMP_XCHG_FOR_BGET
173 #define USE_CMP_XCHG_FOR_BGET 1
174 #endif
175 
176 // Test to see if queuing lock is better than bootstrap lock for bget
177 // #ifndef USE_QUEUING_LOCK_FOR_BGET
178 // #define USE_QUEUING_LOCK_FOR_BGET
179 // #endif
180 
181 #define KMP_NSEC_PER_SEC 1000000000L
182 #define KMP_USEC_PER_SEC 1000000L
183 
184 /*!
185 @ingroup BASIC_TYPES
186 @{
187 */
188 
189 /*!
190 Values for bit flags used in the ident_t to describe the fields.
191 */
192 enum {
193   /*! Use trampoline for internal microtasks */
194   KMP_IDENT_IMB = 0x01,
195   /*! Use c-style ident structure */
196   KMP_IDENT_KMPC = 0x02,
197   /* 0x04 is no longer used */
198   /*! Entry point generated by auto-parallelization */
199   KMP_IDENT_AUTOPAR = 0x08,
200   /*! Compiler generates atomic reduction option for kmpc_reduce* */
201   KMP_IDENT_ATOMIC_REDUCE = 0x10,
202   /*! To mark a 'barrier' directive in user code */
203   KMP_IDENT_BARRIER_EXPL = 0x20,
204   /*! To Mark implicit barriers. */
205   KMP_IDENT_BARRIER_IMPL = 0x0040,
206   KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
207   KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
208   KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
209 
210   KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
211   KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
212 
213   /*! To mark a static loop in OMPT callbacks */
214   KMP_IDENT_WORK_LOOP = 0x200,
215   /*! To mark a sections directive in OMPT callbacks */
216   KMP_IDENT_WORK_SECTIONS = 0x400,
217   /*! To mark a distribute construct in OMPT callbacks */
218   KMP_IDENT_WORK_DISTRIBUTE = 0x800,
219   /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
220       not currently used. If one day we need more bits, then we can use
221       an invalid combination of hints to mean that another, larger field
222       should be used in a different flag. */
223   KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
224   KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
225   KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
226   KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
227   KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
228   KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
229 };
230 
231 /*!
232  * The ident structure that describes a source location.
233  */
234 typedef struct ident {
235   kmp_int32 reserved_1; /**<  might be used in Fortran; see above  */
236   kmp_int32 flags; /**<  also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
237                       identifies this union member  */
238   kmp_int32 reserved_2; /**<  not really used in Fortran any more; see above */
239 #if USE_ITT_BUILD
240 /*  but currently used for storing region-specific ITT */
241 /*  contextual information. */
242 #endif /* USE_ITT_BUILD */
243   kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++  */
244   char const *psource; /**< String describing the source location.
245                        The string is composed of semi-colon separated fields
246                        which describe the source file, the function and a pair
247                        of line numbers that delimit the construct. */
248   // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
249   kmp_int32 get_openmp_version() {
250     return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
251   }
252 } ident_t;
253 /*!
254 @}
255 */
256 
257 // Some forward declarations.
258 typedef union kmp_team kmp_team_t;
259 typedef struct kmp_taskdata kmp_taskdata_t;
260 typedef union kmp_task_team kmp_task_team_t;
261 typedef union kmp_team kmp_team_p;
262 typedef union kmp_info kmp_info_p;
263 typedef union kmp_root kmp_root_p;
264 
265 template <bool C = false, bool S = true> class kmp_flag_32;
266 template <bool C = false, bool S = true> class kmp_flag_64;
267 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
268 class kmp_flag_oncore;
269 
270 #ifdef __cplusplus
271 extern "C" {
272 #endif
273 
274 /* ------------------------------------------------------------------------ */
275 
276 /* Pack two 32-bit signed integers into a 64-bit signed integer */
277 /* ToDo: Fix word ordering for big-endian machines. */
278 #define KMP_PACK_64(HIGH_32, LOW_32)                                           \
279   ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
280 
281 // Generic string manipulation macros. Assume that _x is of type char *
282 #define SKIP_WS(_x)                                                            \
283   {                                                                            \
284     while (*(_x) == ' ' || *(_x) == '\t')                                      \
285       (_x)++;                                                                  \
286   }
287 #define SKIP_DIGITS(_x)                                                        \
288   {                                                                            \
289     while (*(_x) >= '0' && *(_x) <= '9')                                       \
290       (_x)++;                                                                  \
291   }
292 #define SKIP_TOKEN(_x)                                                         \
293   {                                                                            \
294     while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
295            (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_')                     \
296       (_x)++;                                                                  \
297   }
298 #define SKIP_TO(_x, _c)                                                        \
299   {                                                                            \
300     while (*(_x) != '\0' && *(_x) != (_c))                                     \
301       (_x)++;                                                                  \
302   }
303 
304 /* ------------------------------------------------------------------------ */
305 
306 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
307 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
308 
309 /* ------------------------------------------------------------------------ */
310 /* Enumeration types */
311 
312 enum kmp_state_timer {
313   ts_stop,
314   ts_start,
315   ts_pause,
316 
317   ts_last_state
318 };
319 
320 enum dynamic_mode {
321   dynamic_default,
322 #ifdef USE_LOAD_BALANCE
323   dynamic_load_balance,
324 #endif /* USE_LOAD_BALANCE */
325   dynamic_random,
326   dynamic_thread_limit,
327   dynamic_max
328 };
329 
330 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
331  * not include it here */
332 #ifndef KMP_SCHED_TYPE_DEFINED
333 #define KMP_SCHED_TYPE_DEFINED
334 typedef enum kmp_sched {
335   kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
336   // Note: need to adjust __kmp_sch_map global array in case enum is changed
337   kmp_sched_static = 1, // mapped to kmp_sch_static_chunked           (33)
338   kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked          (35)
339   kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked           (36)
340   kmp_sched_auto = 4, // mapped to kmp_sch_auto                     (38)
341   kmp_sched_upper_std = 5, // upper bound for standard schedules
342   kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
343   kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
344 #if KMP_STATIC_STEAL_ENABLED
345   kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
346 #endif
347   kmp_sched_upper,
348   kmp_sched_default = kmp_sched_static, // default scheduling
349   kmp_sched_monotonic = 0x80000000
350 } kmp_sched_t;
351 #endif
352 
353 /*!
354  @ingroup WORK_SHARING
355  * Describes the loop schedule to be used for a parallel for loop.
356  */
357 enum sched_type : kmp_int32 {
358   kmp_sch_lower = 32, /**< lower bound for unordered values */
359   kmp_sch_static_chunked = 33,
360   kmp_sch_static = 34, /**< static unspecialized */
361   kmp_sch_dynamic_chunked = 35,
362   kmp_sch_guided_chunked = 36, /**< guided unspecialized */
363   kmp_sch_runtime = 37,
364   kmp_sch_auto = 38, /**< auto */
365   kmp_sch_trapezoidal = 39,
366 
367   /* accessible only through KMP_SCHEDULE environment variable */
368   kmp_sch_static_greedy = 40,
369   kmp_sch_static_balanced = 41,
370   /* accessible only through KMP_SCHEDULE environment variable */
371   kmp_sch_guided_iterative_chunked = 42,
372   kmp_sch_guided_analytical_chunked = 43,
373   /* accessible only through KMP_SCHEDULE environment variable */
374   kmp_sch_static_steal = 44,
375 
376   /* static with chunk adjustment (e.g., simd) */
377   kmp_sch_static_balanced_chunked = 45,
378   kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
379   kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
380 
381   /* accessible only through KMP_SCHEDULE environment variable */
382   kmp_sch_upper, /**< upper bound for unordered values */
383 
384   kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
385   kmp_ord_static_chunked = 65,
386   kmp_ord_static = 66, /**< ordered static unspecialized */
387   kmp_ord_dynamic_chunked = 67,
388   kmp_ord_guided_chunked = 68,
389   kmp_ord_runtime = 69,
390   kmp_ord_auto = 70, /**< ordered auto */
391   kmp_ord_trapezoidal = 71,
392   kmp_ord_upper, /**< upper bound for ordered values */
393 
394   /* Schedules for Distribute construct */
395   kmp_distribute_static_chunked = 91, /**< distribute static chunked */
396   kmp_distribute_static = 92, /**< distribute static unspecialized */
397 
398   /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
399      single iteration/chunk, even if the loop is serialized. For the schedule
400      types listed above, the entire iteration vector is returned if the loop is
401      serialized. This doesn't work for gcc/gcomp sections. */
402   kmp_nm_lower = 160, /**< lower bound for nomerge values */
403 
404   kmp_nm_static_chunked =
405       (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
406   kmp_nm_static = 162, /**< static unspecialized */
407   kmp_nm_dynamic_chunked = 163,
408   kmp_nm_guided_chunked = 164, /**< guided unspecialized */
409   kmp_nm_runtime = 165,
410   kmp_nm_auto = 166, /**< auto */
411   kmp_nm_trapezoidal = 167,
412 
413   /* accessible only through KMP_SCHEDULE environment variable */
414   kmp_nm_static_greedy = 168,
415   kmp_nm_static_balanced = 169,
416   /* accessible only through KMP_SCHEDULE environment variable */
417   kmp_nm_guided_iterative_chunked = 170,
418   kmp_nm_guided_analytical_chunked = 171,
419   kmp_nm_static_steal =
420       172, /* accessible only through OMP_SCHEDULE environment variable */
421 
422   kmp_nm_ord_static_chunked = 193,
423   kmp_nm_ord_static = 194, /**< ordered static unspecialized */
424   kmp_nm_ord_dynamic_chunked = 195,
425   kmp_nm_ord_guided_chunked = 196,
426   kmp_nm_ord_runtime = 197,
427   kmp_nm_ord_auto = 198, /**< auto */
428   kmp_nm_ord_trapezoidal = 199,
429   kmp_nm_upper, /**< upper bound for nomerge values */
430 
431   /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
432      we need to distinguish the three possible cases (no modifier, monotonic
433      modifier, nonmonotonic modifier), we need separate bits for each modifier.
434      The absence of monotonic does not imply nonmonotonic, especially since 4.5
435      says that the behaviour of the "no modifier" case is implementation defined
436      in 4.5, but will become "nonmonotonic" in 5.0.
437 
438      Since we're passing a full 32 bit value, we can use a couple of high bits
439      for these flags; out of paranoia we avoid the sign bit.
440 
441      These modifiers can be or-ed into non-static schedules by the compiler to
442      pass the additional information. They will be stripped early in the
443      processing in __kmp_dispatch_init when setting up schedules, so most of the
444      code won't ever see schedules with these bits set.  */
445   kmp_sch_modifier_monotonic =
446       (1 << 29), /**< Set if the monotonic schedule modifier was present */
447   kmp_sch_modifier_nonmonotonic =
448       (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
449 
450 #define SCHEDULE_WITHOUT_MODIFIERS(s)                                          \
451   (enum sched_type)(                                                           \
452       (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
453 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
454 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
455 #define SCHEDULE_HAS_NO_MODIFIERS(s)                                           \
456   (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
457 #define SCHEDULE_GET_MODIFIERS(s)                                              \
458   ((enum sched_type)(                                                          \
459       (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
460 #define SCHEDULE_SET_MODIFIERS(s, m)                                           \
461   (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
462 #define SCHEDULE_NONMONOTONIC 0
463 #define SCHEDULE_MONOTONIC 1
464 
465   kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
466 };
467 
468 // Apply modifiers on internal kind to standard kind
469 static inline void
470 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
471                                enum sched_type internal_kind) {
472   if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
473     *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
474   }
475 }
476 
477 // Apply modifiers on standard kind to internal kind
478 static inline void
479 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
480                                enum sched_type *internal_kind) {
481   if ((int)kind & (int)kmp_sched_monotonic) {
482     *internal_kind = (enum sched_type)((int)*internal_kind |
483                                        (int)kmp_sch_modifier_monotonic);
484   }
485 }
486 
487 // Get standard schedule without modifiers
488 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
489   return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
490 }
491 
492 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
493 typedef union kmp_r_sched {
494   struct {
495     enum sched_type r_sched_type;
496     int chunk;
497   };
498   kmp_int64 sched;
499 } kmp_r_sched_t;
500 
501 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
502 // internal schedule types
503 
504 enum library_type {
505   library_none,
506   library_serial,
507   library_turnaround,
508   library_throughput
509 };
510 
511 #if KMP_OS_LINUX
512 enum clock_function_type {
513   clock_function_gettimeofday,
514   clock_function_clock_gettime
515 };
516 #endif /* KMP_OS_LINUX */
517 
518 #if KMP_MIC_SUPPORTED
519 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
520 #endif
521 
522 /* -- fast reduction stuff ------------------------------------------------ */
523 
524 #undef KMP_FAST_REDUCTION_BARRIER
525 #define KMP_FAST_REDUCTION_BARRIER 1
526 
527 #undef KMP_FAST_REDUCTION_CORE_DUO
528 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
529 #define KMP_FAST_REDUCTION_CORE_DUO 1
530 #endif
531 
532 enum _reduction_method {
533   reduction_method_not_defined = 0,
534   critical_reduce_block = (1 << 8),
535   atomic_reduce_block = (2 << 8),
536   tree_reduce_block = (3 << 8),
537   empty_reduce_block = (4 << 8)
538 };
539 
540 // Description of the packed_reduction_method variable:
541 // The packed_reduction_method variable consists of two enum types variables
542 // that are packed together into 0-th byte and 1-st byte:
543 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
544 // barrier that will be used in fast reduction: bs_plain_barrier or
545 // bs_reduction_barrier
546 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
547 // be used in fast reduction;
548 // Reduction method is of 'enum _reduction_method' type and it's defined the way
549 // so that the bits of 0-th byte are empty, so no need to execute a shift
550 // instruction while packing/unpacking
551 
552 #if KMP_FAST_REDUCTION_BARRIER
553 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
554   ((reduction_method) | (barrier_type))
555 
556 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
557   ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
558 
559 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method)                      \
560   ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
561 #else
562 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
563   (reduction_method)
564 
565 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
566   (packed_reduction_method)
567 
568 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
569 #endif
570 
571 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block)  \
572   ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) ==                       \
573    (which_reduction_block))
574 
575 #if KMP_FAST_REDUCTION_BARRIER
576 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER                               \
577   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
578 
579 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER                                   \
580   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
581 #endif
582 
583 typedef int PACKED_REDUCTION_METHOD_T;
584 
585 /* -- end of fast reduction stuff ----------------------------------------- */
586 
587 #if KMP_OS_WINDOWS
588 #define USE_CBLKDATA
589 #if KMP_MSVC_COMPAT
590 #pragma warning(push)
591 #pragma warning(disable : 271 310)
592 #endif
593 #include <windows.h>
594 #if KMP_MSVC_COMPAT
595 #pragma warning(pop)
596 #endif
597 #endif
598 
599 #if KMP_OS_UNIX
600 #include <dlfcn.h>
601 #include <pthread.h>
602 #endif
603 
604 enum kmp_hw_t : int {
605   KMP_HW_UNKNOWN = -1,
606   KMP_HW_SOCKET = 0,
607   KMP_HW_PROC_GROUP,
608   KMP_HW_NUMA,
609   KMP_HW_DIE,
610   KMP_HW_LLC,
611   KMP_HW_L3,
612   KMP_HW_TILE,
613   KMP_HW_MODULE,
614   KMP_HW_L2,
615   KMP_HW_L1,
616   KMP_HW_CORE,
617   KMP_HW_THREAD,
618   KMP_HW_LAST
619 };
620 
621 typedef enum kmp_hw_core_type_t {
622   KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
623 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
624   KMP_HW_CORE_TYPE_ATOM = 0x20,
625   KMP_HW_CORE_TYPE_CORE = 0x40,
626   KMP_HW_MAX_NUM_CORE_TYPES = 3,
627 #else
628   KMP_HW_MAX_NUM_CORE_TYPES = 1,
629 #endif
630 } kmp_hw_core_type_t;
631 
632 #define KMP_HW_MAX_NUM_CORE_EFFS 8
633 
634 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type)                                   \
635   KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
636 #define KMP_ASSERT_VALID_HW_TYPE(type)                                         \
637   KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
638 
639 #define KMP_FOREACH_HW_TYPE(type)                                              \
640   for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST;                        \
641        type = (kmp_hw_t)((int)type + 1))
642 
643 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
644 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
645 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
646 
647 /* Only Linux* OS and Windows* OS support thread affinity. */
648 #if KMP_AFFINITY_SUPPORTED
649 
650 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
651 #if KMP_OS_WINDOWS
652 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
653 typedef struct GROUP_AFFINITY {
654   KAFFINITY Mask;
655   WORD Group;
656   WORD Reserved[3];
657 } GROUP_AFFINITY;
658 #endif /* _MSC_VER < 1600 */
659 #if KMP_GROUP_AFFINITY
660 extern int __kmp_num_proc_groups;
661 #else
662 static const int __kmp_num_proc_groups = 1;
663 #endif /* KMP_GROUP_AFFINITY */
664 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
665 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
666 
667 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
668 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
669 
670 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
671 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
672 
673 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
674                                              GROUP_AFFINITY *);
675 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
676 #endif /* KMP_OS_WINDOWS */
677 
678 #if KMP_USE_HWLOC
679 extern hwloc_topology_t __kmp_hwloc_topology;
680 extern int __kmp_hwloc_error;
681 #endif
682 
683 extern size_t __kmp_affin_mask_size;
684 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
685 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
686 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
687 #define KMP_CPU_SET_ITERATE(i, mask)                                           \
688   for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
689 #define KMP_CPU_SET(i, mask) (mask)->set(i)
690 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
691 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
692 #define KMP_CPU_ZERO(mask) (mask)->zero()
693 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
694 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
695 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
696 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
697 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
698 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
699 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
700 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
701 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
702 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
703 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
704 #define KMP_CPU_ALLOC_ARRAY(arr, n)                                            \
705   (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
706 #define KMP_CPU_FREE_ARRAY(arr, n)                                             \
707   __kmp_affinity_dispatch->deallocate_mask_array(arr)
708 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
709 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
710 #define __kmp_get_system_affinity(mask, abort_bool)                            \
711   (mask)->get_system_affinity(abort_bool)
712 #define __kmp_set_system_affinity(mask, abort_bool)                            \
713   (mask)->set_system_affinity(abort_bool)
714 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
715 
716 class KMPAffinity {
717 public:
718   class Mask {
719   public:
720     void *operator new(size_t n);
721     void operator delete(void *p);
722     void *operator new[](size_t n);
723     void operator delete[](void *p);
724     virtual ~Mask() {}
725     // Set bit i to 1
726     virtual void set(int i) {}
727     // Return bit i
728     virtual bool is_set(int i) const { return false; }
729     // Set bit i to 0
730     virtual void clear(int i) {}
731     // Zero out entire mask
732     virtual void zero() {}
733     // Copy src into this mask
734     virtual void copy(const Mask *src) {}
735     // this &= rhs
736     virtual void bitwise_and(const Mask *rhs) {}
737     // this |= rhs
738     virtual void bitwise_or(const Mask *rhs) {}
739     // this = ~this
740     virtual void bitwise_not() {}
741     // API for iterating over an affinity mask
742     // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
743     virtual int begin() const { return 0; }
744     virtual int end() const { return 0; }
745     virtual int next(int previous) const { return 0; }
746 #if KMP_OS_WINDOWS
747     virtual int set_process_affinity(bool abort_on_error) const { return -1; }
748 #endif
749     // Set the system's affinity to this affinity mask's value
750     virtual int set_system_affinity(bool abort_on_error) const { return -1; }
751     // Set this affinity mask to the current system affinity
752     virtual int get_system_affinity(bool abort_on_error) { return -1; }
753     // Only 1 DWORD in the mask should have any procs set.
754     // Return the appropriate index, or -1 for an invalid mask.
755     virtual int get_proc_group() const { return -1; }
756     int get_max_cpu() const {
757       int cpu;
758       int max_cpu = -1;
759       KMP_CPU_SET_ITERATE(cpu, this) {
760         if (cpu > max_cpu)
761           max_cpu = cpu;
762       }
763       return max_cpu;
764     }
765   };
766   void *operator new(size_t n);
767   void operator delete(void *p);
768   // Need virtual destructor
769   virtual ~KMPAffinity() = default;
770   // Determine if affinity is capable
771   virtual void determine_capable(const char *env_var) {}
772   // Bind the current thread to os proc
773   virtual void bind_thread(int proc) {}
774   // Factory functions to allocate/deallocate a mask
775   virtual Mask *allocate_mask() { return nullptr; }
776   virtual void deallocate_mask(Mask *m) {}
777   virtual Mask *allocate_mask_array(int num) { return nullptr; }
778   virtual void deallocate_mask_array(Mask *m) {}
779   virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
780   static void pick_api();
781   static void destroy_api();
782   enum api_type {
783     NATIVE_OS
784 #if KMP_USE_HWLOC
785     ,
786     HWLOC
787 #endif
788   };
789   virtual api_type get_api_type() const {
790     KMP_ASSERT(0);
791     return NATIVE_OS;
792   }
793 
794 private:
795   static bool picked_api;
796 };
797 
798 typedef KMPAffinity::Mask kmp_affin_mask_t;
799 extern KMPAffinity *__kmp_affinity_dispatch;
800 
801 // Declare local char buffers with this size for printing debug and info
802 // messages, using __kmp_affinity_print_mask().
803 #define KMP_AFFIN_MASK_PRINT_LEN 1024
804 
805 enum affinity_type {
806   affinity_none = 0,
807   affinity_physical,
808   affinity_logical,
809   affinity_compact,
810   affinity_scatter,
811   affinity_explicit,
812   affinity_balanced,
813   affinity_disabled, // not used outsize the env var parser
814   affinity_default
815 };
816 
817 enum affinity_top_method {
818   affinity_top_method_all = 0, // try all (supported) methods, in order
819 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
820   affinity_top_method_apicid,
821   affinity_top_method_x2apicid,
822   affinity_top_method_x2apicid_1f,
823 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
824   affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
825 #if KMP_GROUP_AFFINITY
826   affinity_top_method_group,
827 #endif /* KMP_GROUP_AFFINITY */
828   affinity_top_method_flat,
829 #if KMP_USE_HWLOC
830   affinity_top_method_hwloc,
831 #endif
832   affinity_top_method_default
833 };
834 
835 #define affinity_respect_mask_default (2)
836 
837 typedef struct kmp_affinity_flags_t {
838   unsigned dups : 1;
839   unsigned verbose : 1;
840   unsigned warnings : 1;
841   unsigned respect : 2;
842   unsigned reset : 1;
843   unsigned initialized : 1;
844   unsigned reserved : 25;
845 } kmp_affinity_flags_t;
846 KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
847 
848 typedef struct kmp_affinity_ids_t {
849   int ids[KMP_HW_LAST];
850   int operator[](size_t idx) const { return ids[idx]; }
851   int &operator[](size_t idx) { return ids[idx]; }
852   kmp_affinity_ids_t &operator=(const kmp_affinity_ids_t &rhs) {
853     for (int i = 0; i < KMP_HW_LAST; ++i)
854       ids[i] = rhs[i];
855     return *this;
856   }
857 } kmp_affinity_ids_t;
858 
859 typedef struct kmp_affinity_attrs_t {
860   int core_type : 8;
861   int core_eff : 8;
862   unsigned valid : 1;
863   unsigned reserved : 15;
864 } kmp_affinity_attrs_t;
865 #define KMP_AFFINITY_ATTRS_UNKNOWN                                             \
866   { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
867 
868 typedef struct kmp_affinity_t {
869   char *proclist;
870   enum affinity_type type;
871   kmp_hw_t gran;
872   int gran_levels;
873   int compact;
874   int offset;
875   kmp_affinity_flags_t flags;
876   unsigned num_masks;
877   kmp_affin_mask_t *masks;
878   kmp_affinity_ids_t *ids;
879   kmp_affinity_attrs_t *attrs;
880   unsigned num_os_id_masks;
881   kmp_affin_mask_t *os_id_masks;
882   const char *env_var;
883 } kmp_affinity_t;
884 
885 #define KMP_AFFINITY_INIT(env)                                                 \
886   {                                                                            \
887     nullptr, affinity_default, KMP_HW_UNKNOWN, -1, 0, 0,                       \
888         {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE}, 0,   \
889         nullptr, nullptr, nullptr, 0, nullptr, env                             \
890   }
891 
892 extern enum affinity_top_method __kmp_affinity_top_method;
893 extern kmp_affinity_t __kmp_affinity;
894 extern kmp_affinity_t __kmp_hh_affinity;
895 extern kmp_affinity_t *__kmp_affinities[2];
896 
897 extern void __kmp_affinity_bind_thread(int which);
898 
899 extern kmp_affin_mask_t *__kmp_affin_fullMask;
900 extern kmp_affin_mask_t *__kmp_affin_origMask;
901 extern char *__kmp_cpuinfo_file;
902 
903 #endif /* KMP_AFFINITY_SUPPORTED */
904 
905 // This needs to be kept in sync with the values in omp.h !!!
906 typedef enum kmp_proc_bind_t {
907   proc_bind_false = 0,
908   proc_bind_true,
909   proc_bind_primary,
910   proc_bind_close,
911   proc_bind_spread,
912   proc_bind_intel, // use KMP_AFFINITY interface
913   proc_bind_default
914 } kmp_proc_bind_t;
915 
916 typedef struct kmp_nested_proc_bind_t {
917   kmp_proc_bind_t *bind_types;
918   int size;
919   int used;
920 } kmp_nested_proc_bind_t;
921 
922 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
923 extern kmp_proc_bind_t __kmp_teams_proc_bind;
924 
925 extern int __kmp_display_affinity;
926 extern char *__kmp_affinity_format;
927 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
928 #if OMPT_SUPPORT
929 extern int __kmp_tool;
930 extern char *__kmp_tool_libraries;
931 #endif // OMPT_SUPPORT
932 
933 #if KMP_AFFINITY_SUPPORTED
934 #define KMP_PLACE_ALL (-1)
935 #define KMP_PLACE_UNDEFINED (-2)
936 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
937 #define KMP_AFFINITY_NON_PROC_BIND                                             \
938   ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false ||                 \
939     __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) &&                \
940    (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
941 #endif /* KMP_AFFINITY_SUPPORTED */
942 
943 extern int __kmp_affinity_num_places;
944 
945 typedef enum kmp_cancel_kind_t {
946   cancel_noreq = 0,
947   cancel_parallel = 1,
948   cancel_loop = 2,
949   cancel_sections = 3,
950   cancel_taskgroup = 4
951 } kmp_cancel_kind_t;
952 
953 // KMP_HW_SUBSET support:
954 typedef struct kmp_hws_item {
955   int num;
956   int offset;
957 } kmp_hws_item_t;
958 
959 extern kmp_hws_item_t __kmp_hws_socket;
960 extern kmp_hws_item_t __kmp_hws_die;
961 extern kmp_hws_item_t __kmp_hws_node;
962 extern kmp_hws_item_t __kmp_hws_tile;
963 extern kmp_hws_item_t __kmp_hws_core;
964 extern kmp_hws_item_t __kmp_hws_proc;
965 extern int __kmp_hws_requested;
966 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
967 
968 /* ------------------------------------------------------------------------ */
969 
970 #define KMP_PAD(type, sz)                                                      \
971   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
972 
973 // We need to avoid using -1 as a GTID as +1 is added to the gtid
974 // when storing it in a lock, and the value 0 is reserved.
975 #define KMP_GTID_DNE (-2) /* Does not exist */
976 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
977 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
978 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
979 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
980 
981 /* OpenMP 5.0 Memory Management support */
982 
983 #ifndef __OMP_H
984 // Duplicate type definitions from omp.h
985 typedef uintptr_t omp_uintptr_t;
986 
987 typedef enum {
988   omp_atk_sync_hint = 1,
989   omp_atk_alignment = 2,
990   omp_atk_access = 3,
991   omp_atk_pool_size = 4,
992   omp_atk_fallback = 5,
993   omp_atk_fb_data = 6,
994   omp_atk_pinned = 7,
995   omp_atk_partition = 8
996 } omp_alloctrait_key_t;
997 
998 typedef enum {
999   omp_atv_false = 0,
1000   omp_atv_true = 1,
1001   omp_atv_contended = 3,
1002   omp_atv_uncontended = 4,
1003   omp_atv_serialized = 5,
1004   omp_atv_sequential = omp_atv_serialized, // (deprecated)
1005   omp_atv_private = 6,
1006   omp_atv_all = 7,
1007   omp_atv_thread = 8,
1008   omp_atv_pteam = 9,
1009   omp_atv_cgroup = 10,
1010   omp_atv_default_mem_fb = 11,
1011   omp_atv_null_fb = 12,
1012   omp_atv_abort_fb = 13,
1013   omp_atv_allocator_fb = 14,
1014   omp_atv_environment = 15,
1015   omp_atv_nearest = 16,
1016   omp_atv_blocked = 17,
1017   omp_atv_interleaved = 18
1018 } omp_alloctrait_value_t;
1019 #define omp_atv_default ((omp_uintptr_t)-1)
1020 
1021 typedef void *omp_memspace_handle_t;
1022 extern omp_memspace_handle_t const omp_default_mem_space;
1023 extern omp_memspace_handle_t const omp_large_cap_mem_space;
1024 extern omp_memspace_handle_t const omp_const_mem_space;
1025 extern omp_memspace_handle_t const omp_high_bw_mem_space;
1026 extern omp_memspace_handle_t const omp_low_lat_mem_space;
1027 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1028 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1029 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1030 
1031 typedef struct {
1032   omp_alloctrait_key_t key;
1033   omp_uintptr_t value;
1034 } omp_alloctrait_t;
1035 
1036 typedef void *omp_allocator_handle_t;
1037 extern omp_allocator_handle_t const omp_null_allocator;
1038 extern omp_allocator_handle_t const omp_default_mem_alloc;
1039 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1040 extern omp_allocator_handle_t const omp_const_mem_alloc;
1041 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1042 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1043 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1044 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1045 extern omp_allocator_handle_t const omp_thread_mem_alloc;
1046 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1047 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1048 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1049 extern omp_allocator_handle_t const kmp_max_mem_alloc;
1050 extern omp_allocator_handle_t __kmp_def_allocator;
1051 
1052 // end of duplicate type definitions from omp.h
1053 #endif
1054 
1055 extern int __kmp_memkind_available;
1056 
1057 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1058 
1059 typedef struct kmp_allocator_t {
1060   omp_memspace_handle_t memspace;
1061   void **memkind; // pointer to memkind
1062   size_t alignment;
1063   omp_alloctrait_value_t fb;
1064   kmp_allocator_t *fb_data;
1065   kmp_uint64 pool_size;
1066   kmp_uint64 pool_used;
1067   bool pinned;
1068 } kmp_allocator_t;
1069 
1070 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1071                                                     omp_memspace_handle_t,
1072                                                     int ntraits,
1073                                                     omp_alloctrait_t traits[]);
1074 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1075 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1076 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1077 // external interfaces, may be used by compiler
1078 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1079 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1080                                   omp_allocator_handle_t al);
1081 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1082                            omp_allocator_handle_t al);
1083 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1084                             omp_allocator_handle_t al,
1085                             omp_allocator_handle_t free_al);
1086 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1087 // internal interfaces, contain real implementation
1088 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1089                          omp_allocator_handle_t al);
1090 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1091                           omp_allocator_handle_t al);
1092 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1093                            omp_allocator_handle_t al,
1094                            omp_allocator_handle_t free_al);
1095 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1096 
1097 extern void __kmp_init_memkind();
1098 extern void __kmp_fini_memkind();
1099 extern void __kmp_init_target_mem();
1100 
1101 /* ------------------------------------------------------------------------ */
1102 
1103 #define KMP_UINT64_MAX                                                         \
1104   (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1105 
1106 #define KMP_MIN_NTH 1
1107 
1108 #ifndef KMP_MAX_NTH
1109 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1110 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1111 #else
1112 #define KMP_MAX_NTH INT_MAX
1113 #endif
1114 #endif /* KMP_MAX_NTH */
1115 
1116 #ifdef PTHREAD_STACK_MIN
1117 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1118 #else
1119 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1120 #endif
1121 
1122 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1123 
1124 #if KMP_ARCH_X86
1125 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1126 #elif KMP_ARCH_X86_64
1127 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1128 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1129 #else
1130 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1131 #endif
1132 
1133 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1134 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1135 #define KMP_MAX_MALLOC_POOL_INCR                                               \
1136   (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1137 
1138 #define KMP_MIN_STKOFFSET (0)
1139 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1140 #if KMP_OS_DARWIN
1141 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1142 #else
1143 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1144 #endif
1145 
1146 #define KMP_MIN_STKPADDING (0)
1147 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1148 
1149 #define KMP_BLOCKTIME_MULTIPLIER                                               \
1150   (1000) /* number of blocktime units per second */
1151 #define KMP_MIN_BLOCKTIME (0)
1152 #define KMP_MAX_BLOCKTIME                                                      \
1153   (INT_MAX) /* Must be this for "infinite" setting the work */
1154 
1155 /* __kmp_blocktime is in milliseconds */
1156 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200))
1157 
1158 #if KMP_USE_MONITOR
1159 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1160 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1161 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1162 
1163 /* Calculate new number of monitor wakeups for a specific block time based on
1164    previous monitor_wakeups. Only allow increasing number of wakeups */
1165 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups)                 \
1166   (((blocktime) == KMP_MAX_BLOCKTIME)   ? (monitor_wakeups)                    \
1167    : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS              \
1168    : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime)))            \
1169        ? (monitor_wakeups)                                                     \
1170        : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1171 
1172 /* Calculate number of intervals for a specific block time based on
1173    monitor_wakeups */
1174 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups)               \
1175   (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) /        \
1176    (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1177 #else
1178 #define KMP_BLOCKTIME(team, tid)                                               \
1179   (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1180 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1181 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1182 extern kmp_uint64 __kmp_ticks_per_msec;
1183 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1184 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1185 #else
1186 #define KMP_NOW() __kmp_hardware_timestamp()
1187 #endif
1188 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1189 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1190   (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1191 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1192 #else
1193 // System time is retrieved sporadically while blocking.
1194 extern kmp_uint64 __kmp_now_nsec();
1195 #define KMP_NOW() __kmp_now_nsec()
1196 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1197 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1198   (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1199 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1200 #endif
1201 #endif // KMP_USE_MONITOR
1202 
1203 #define KMP_MIN_STATSCOLS 40
1204 #define KMP_MAX_STATSCOLS 4096
1205 #define KMP_DEFAULT_STATSCOLS 80
1206 
1207 #define KMP_MIN_INTERVAL 0
1208 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1209 #define KMP_DEFAULT_INTERVAL 0
1210 
1211 #define KMP_MIN_CHUNK 1
1212 #define KMP_MAX_CHUNK (INT_MAX - 1)
1213 #define KMP_DEFAULT_CHUNK 1
1214 
1215 #define KMP_MIN_DISP_NUM_BUFF 1
1216 #define KMP_DFLT_DISP_NUM_BUFF 7
1217 #define KMP_MAX_DISP_NUM_BUFF 4096
1218 
1219 #define KMP_MAX_ORDERED 8
1220 
1221 #define KMP_MAX_FIELDS 32
1222 
1223 #define KMP_MAX_BRANCH_BITS 31
1224 
1225 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1226 
1227 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1228 
1229 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1230 
1231 /* Minimum number of threads before switch to TLS gtid (experimentally
1232    determined) */
1233 /* josh TODO: what about OS X* tuning? */
1234 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1235 #define KMP_TLS_GTID_MIN 5
1236 #else
1237 #define KMP_TLS_GTID_MIN INT_MAX
1238 #endif
1239 
1240 #define KMP_MASTER_TID(tid) (0 == (tid))
1241 #define KMP_WORKER_TID(tid) (0 != (tid))
1242 
1243 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1244 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1245 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1246 
1247 #ifndef TRUE
1248 #define FALSE 0
1249 #define TRUE (!FALSE)
1250 #endif
1251 
1252 /* NOTE: all of the following constants must be even */
1253 
1254 #if KMP_OS_WINDOWS
1255 #define KMP_INIT_WAIT 64U /* initial number of spin-tests   */
1256 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1257 #elif KMP_OS_LINUX
1258 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1259 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1260 #elif KMP_OS_DARWIN
1261 /* TODO: tune for KMP_OS_DARWIN */
1262 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1263 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1264 #elif KMP_OS_DRAGONFLY
1265 /* TODO: tune for KMP_OS_DRAGONFLY */
1266 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1267 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1268 #elif KMP_OS_FREEBSD
1269 /* TODO: tune for KMP_OS_FREEBSD */
1270 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1271 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1272 #elif KMP_OS_NETBSD
1273 /* TODO: tune for KMP_OS_NETBSD */
1274 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1275 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1276 #elif KMP_OS_HURD
1277 /* TODO: tune for KMP_OS_HURD */
1278 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1279 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1280 #elif KMP_OS_OPENBSD
1281 /* TODO: tune for KMP_OS_OPENBSD */
1282 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1283 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1284 #endif
1285 
1286 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1287 typedef struct kmp_cpuid {
1288   kmp_uint32 eax;
1289   kmp_uint32 ebx;
1290   kmp_uint32 ecx;
1291   kmp_uint32 edx;
1292 } kmp_cpuid_t;
1293 
1294 typedef struct kmp_cpuinfo_flags_t {
1295   unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1296   unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1297   unsigned hybrid : 1;
1298   unsigned reserved : 29; // Ensure size of 32 bits
1299 } kmp_cpuinfo_flags_t;
1300 
1301 typedef struct kmp_cpuinfo {
1302   int initialized; // If 0, other fields are not initialized.
1303   int signature; // CPUID(1).EAX
1304   int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1305   int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1306   // Model << 4 ) + Model)
1307   int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1308   kmp_cpuinfo_flags_t flags;
1309   int apic_id;
1310   int physical_id;
1311   int logical_id;
1312   kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1313   char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1314 } kmp_cpuinfo_t;
1315 
1316 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1317 
1318 #if KMP_OS_UNIX
1319 // subleaf is only needed for cache and topology discovery and can be set to
1320 // zero in most cases
1321 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1322   __asm__ __volatile__("cpuid"
1323                        : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1324                        : "a"(leaf), "c"(subleaf));
1325 }
1326 // Load p into FPU control word
1327 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1328   __asm__ __volatile__("fldcw %0" : : "m"(*p));
1329 }
1330 // Store FPU control word into p
1331 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1332   __asm__ __volatile__("fstcw %0" : "=m"(*p));
1333 }
1334 static inline void __kmp_clear_x87_fpu_status_word() {
1335 #if KMP_MIC
1336   // 32-bit protected mode x87 FPU state
1337   struct x87_fpu_state {
1338     unsigned cw;
1339     unsigned sw;
1340     unsigned tw;
1341     unsigned fip;
1342     unsigned fips;
1343     unsigned fdp;
1344     unsigned fds;
1345   };
1346   struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1347   __asm__ __volatile__("fstenv %0\n\t" // store FP env
1348                        "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1349                        "fldenv %0\n\t" // load FP env back
1350                        : "+m"(fpu_state), "+m"(fpu_state.sw));
1351 #else
1352   __asm__ __volatile__("fnclex");
1353 #endif // KMP_MIC
1354 }
1355 #if __SSE__
1356 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1357 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1358 #else
1359 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1360 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1361 #endif
1362 #else
1363 // Windows still has these as external functions in assembly file
1364 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1365 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1366 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1367 extern void __kmp_clear_x87_fpu_status_word();
1368 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1369 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1370 #endif // KMP_OS_UNIX
1371 
1372 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1373 
1374 // User-level Monitor/Mwait
1375 #if KMP_HAVE_UMWAIT
1376 // We always try for UMWAIT first
1377 #if KMP_HAVE_WAITPKG_INTRINSICS
1378 #if KMP_HAVE_IMMINTRIN_H
1379 #include <immintrin.h>
1380 #elif KMP_HAVE_INTRIN_H
1381 #include <intrin.h>
1382 #endif
1383 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1384 
1385 KMP_ATTRIBUTE_TARGET_WAITPKG
1386 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1387 #if !KMP_HAVE_WAITPKG_INTRINSICS
1388   uint32_t timeHi = uint32_t(counter >> 32);
1389   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1390   char flag;
1391   __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1392                    "setb   %0"
1393                    // The "=q" restraint means any register accessible as rl
1394                    //   in 32-bit mode: a, b, c, and d;
1395                    //   in 64-bit mode: any integer register
1396                    : "=q"(flag)
1397                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1398                    :);
1399   return flag;
1400 #else
1401   return _tpause(hint, counter);
1402 #endif
1403 }
1404 KMP_ATTRIBUTE_TARGET_WAITPKG
1405 static inline void __kmp_umonitor(void *cacheline) {
1406 #if !KMP_HAVE_WAITPKG_INTRINSICS
1407   __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1408                    :
1409                    : "a"(cacheline)
1410                    :);
1411 #else
1412   _umonitor(cacheline);
1413 #endif
1414 }
1415 KMP_ATTRIBUTE_TARGET_WAITPKG
1416 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1417 #if !KMP_HAVE_WAITPKG_INTRINSICS
1418   uint32_t timeHi = uint32_t(counter >> 32);
1419   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1420   char flag;
1421   __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1422                    "setb   %0"
1423                    // The "=q" restraint means any register accessible as rl
1424                    //   in 32-bit mode: a, b, c, and d;
1425                    //   in 64-bit mode: any integer register
1426                    : "=q"(flag)
1427                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1428                    :);
1429   return flag;
1430 #else
1431   return _umwait(hint, counter);
1432 #endif
1433 }
1434 #elif KMP_HAVE_MWAIT
1435 #if KMP_OS_UNIX
1436 #include <pmmintrin.h>
1437 #else
1438 #include <intrin.h>
1439 #endif
1440 #if KMP_OS_UNIX
1441 __attribute__((target("sse3")))
1442 #endif
1443 static inline void
1444 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1445   _mm_monitor(cacheline, extensions, hints);
1446 }
1447 #if KMP_OS_UNIX
1448 __attribute__((target("sse3")))
1449 #endif
1450 static inline void
1451 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1452   _mm_mwait(extensions, hints);
1453 }
1454 #endif // KMP_HAVE_UMWAIT
1455 
1456 #if KMP_ARCH_X86
1457 extern void __kmp_x86_pause(void);
1458 #elif KMP_MIC
1459 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1460 // regression after removal of extra PAUSE from spin loops. Changing
1461 // the delay from 100 to 300 showed even better performance than double PAUSE
1462 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1463 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1464 #else
1465 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1466 #endif
1467 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1468 #elif KMP_ARCH_PPC64
1469 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1470 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1471 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1472 #define KMP_CPU_PAUSE()                                                        \
1473   do {                                                                         \
1474     KMP_PPC64_PRI_LOW();                                                       \
1475     KMP_PPC64_PRI_MED();                                                       \
1476     KMP_PPC64_PRI_LOC_MB();                                                    \
1477   } while (0)
1478 #else
1479 #define KMP_CPU_PAUSE() /* nothing to do */
1480 #endif
1481 
1482 #define KMP_INIT_YIELD(count)                                                  \
1483   { (count) = __kmp_yield_init; }
1484 
1485 #define KMP_INIT_BACKOFF(time)                                                 \
1486   { (time) = __kmp_pause_init; }
1487 
1488 #define KMP_OVERSUBSCRIBED                                                     \
1489   (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1490 
1491 #define KMP_TRY_YIELD                                                          \
1492   ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1493 
1494 #define KMP_TRY_YIELD_OVERSUB                                                  \
1495   ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1496 
1497 #define KMP_YIELD(cond)                                                        \
1498   {                                                                            \
1499     KMP_CPU_PAUSE();                                                           \
1500     if ((cond) && (KMP_TRY_YIELD))                                             \
1501       __kmp_yield();                                                           \
1502   }
1503 
1504 #define KMP_YIELD_OVERSUB()                                                    \
1505   {                                                                            \
1506     KMP_CPU_PAUSE();                                                           \
1507     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1508       __kmp_yield();                                                           \
1509   }
1510 
1511 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1512 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1513 #define KMP_YIELD_SPIN(count)                                                  \
1514   {                                                                            \
1515     KMP_CPU_PAUSE();                                                           \
1516     if (KMP_TRY_YIELD) {                                                       \
1517       (count) -= 2;                                                            \
1518       if (!(count)) {                                                          \
1519         __kmp_yield();                                                         \
1520         (count) = __kmp_yield_next;                                            \
1521       }                                                                        \
1522     }                                                                          \
1523   }
1524 
1525 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1526 // (C0.2) state, which improves performance of other SMT threads on the same
1527 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1528 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1529 // available, fall back to the regular CPU pause and yield combination.
1530 #if KMP_HAVE_UMWAIT
1531 #define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1532 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1533   {                                                                            \
1534     if (__kmp_tpause_enabled) {                                                \
1535       if (KMP_OVERSUBSCRIBED) {                                                \
1536         __kmp_tpause(0, (time));                                               \
1537       } else {                                                                 \
1538         __kmp_tpause(__kmp_tpause_hint, (time));                               \
1539       }                                                                        \
1540       (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK;                          \
1541     } else {                                                                   \
1542       KMP_CPU_PAUSE();                                                         \
1543       if ((KMP_TRY_YIELD_OVERSUB)) {                                           \
1544         __kmp_yield();                                                         \
1545       } else if (__kmp_use_yield == 1) {                                       \
1546         (count) -= 2;                                                          \
1547         if (!(count)) {                                                        \
1548           __kmp_yield();                                                       \
1549           (count) = __kmp_yield_next;                                          \
1550         }                                                                      \
1551       }                                                                        \
1552     }                                                                          \
1553   }
1554 #else
1555 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1556   {                                                                            \
1557     KMP_CPU_PAUSE();                                                           \
1558     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1559       __kmp_yield();                                                           \
1560     else if (__kmp_use_yield == 1) {                                           \
1561       (count) -= 2;                                                            \
1562       if (!(count)) {                                                          \
1563         __kmp_yield();                                                         \
1564         (count) = __kmp_yield_next;                                            \
1565       }                                                                        \
1566     }                                                                          \
1567   }
1568 #endif // KMP_HAVE_UMWAIT
1569 
1570 /* ------------------------------------------------------------------------ */
1571 /* Support datatypes for the orphaned construct nesting checks.             */
1572 /* ------------------------------------------------------------------------ */
1573 
1574 /* When adding to this enum, add its corresponding string in cons_text_c[]
1575  * array in kmp_error.cpp */
1576 enum cons_type {
1577   ct_none,
1578   ct_parallel,
1579   ct_pdo,
1580   ct_pdo_ordered,
1581   ct_psections,
1582   ct_psingle,
1583   ct_critical,
1584   ct_ordered_in_parallel,
1585   ct_ordered_in_pdo,
1586   ct_master,
1587   ct_reduce,
1588   ct_barrier,
1589   ct_masked
1590 };
1591 
1592 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1593 
1594 struct cons_data {
1595   ident_t const *ident;
1596   enum cons_type type;
1597   int prev;
1598   kmp_user_lock_p
1599       name; /* address exclusively for critical section name comparison */
1600 };
1601 
1602 struct cons_header {
1603   int p_top, w_top, s_top;
1604   int stack_size, stack_top;
1605   struct cons_data *stack_data;
1606 };
1607 
1608 struct kmp_region_info {
1609   char *text;
1610   int offset[KMP_MAX_FIELDS];
1611   int length[KMP_MAX_FIELDS];
1612 };
1613 
1614 /* ---------------------------------------------------------------------- */
1615 /* ---------------------------------------------------------------------- */
1616 
1617 #if KMP_OS_WINDOWS
1618 typedef HANDLE kmp_thread_t;
1619 typedef DWORD kmp_key_t;
1620 #endif /* KMP_OS_WINDOWS */
1621 
1622 #if KMP_OS_UNIX
1623 typedef pthread_t kmp_thread_t;
1624 typedef pthread_key_t kmp_key_t;
1625 #endif
1626 
1627 extern kmp_key_t __kmp_gtid_threadprivate_key;
1628 
1629 typedef struct kmp_sys_info {
1630   long maxrss; /* the maximum resident set size utilized (in kilobytes)     */
1631   long minflt; /* the number of page faults serviced without any I/O        */
1632   long majflt; /* the number of page faults serviced that required I/O      */
1633   long nswap; /* the number of times a process was "swapped" out of memory */
1634   long inblock; /* the number of times the file system had to perform input  */
1635   long oublock; /* the number of times the file system had to perform output */
1636   long nvcsw; /* the number of times a context switch was voluntarily      */
1637   long nivcsw; /* the number of times a context switch was forced           */
1638 } kmp_sys_info_t;
1639 
1640 #if USE_ITT_BUILD
1641 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1642 // required type here. Later we will check the type meets requirements.
1643 typedef int kmp_itt_mark_t;
1644 #define KMP_ITT_DEBUG 0
1645 #endif /* USE_ITT_BUILD */
1646 
1647 typedef kmp_int32 kmp_critical_name[8];
1648 
1649 /*!
1650 @ingroup PARALLEL
1651 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1652 The arguments to the outlined function are
1653 @param global_tid the global thread identity of the thread executing the
1654 function.
1655 @param bound_tid  the local identity of the thread executing the function
1656 @param ... pointers to shared variables accessed by the function.
1657 */
1658 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1659 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1660                                  ...);
1661 
1662 /*!
1663 @ingroup THREADPRIVATE
1664 @{
1665 */
1666 /* ---------------------------------------------------------------------------
1667  */
1668 /* Threadprivate initialization/finalization function declarations */
1669 
1670 /*  for non-array objects:  __kmpc_threadprivate_register()  */
1671 
1672 /*!
1673  Pointer to the constructor function.
1674  The first argument is the <tt>this</tt> pointer
1675 */
1676 typedef void *(*kmpc_ctor)(void *);
1677 
1678 /*!
1679  Pointer to the destructor function.
1680  The first argument is the <tt>this</tt> pointer
1681 */
1682 typedef void (*kmpc_dtor)(
1683     void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1684                               compiler */
1685 /*!
1686  Pointer to an alternate constructor.
1687  The first argument is the <tt>this</tt> pointer.
1688 */
1689 typedef void *(*kmpc_cctor)(void *, void *);
1690 
1691 /* for array objects: __kmpc_threadprivate_register_vec() */
1692 /* First arg: "this" pointer */
1693 /* Last arg: number of array elements */
1694 /*!
1695  Array constructor.
1696  First argument is the <tt>this</tt> pointer
1697  Second argument the number of array elements.
1698 */
1699 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1700 /*!
1701  Pointer to the array destructor function.
1702  The first argument is the <tt>this</tt> pointer
1703  Second argument the number of array elements.
1704 */
1705 typedef void (*kmpc_dtor_vec)(void *, size_t);
1706 /*!
1707  Array constructor.
1708  First argument is the <tt>this</tt> pointer
1709  Third argument the number of array elements.
1710 */
1711 typedef void *(*kmpc_cctor_vec)(void *, void *,
1712                                 size_t); /* function unused by compiler */
1713 
1714 /*!
1715 @}
1716 */
1717 
1718 /* keeps tracked of threadprivate cache allocations for cleanup later */
1719 typedef struct kmp_cached_addr {
1720   void **addr; /* address of allocated cache */
1721   void ***compiler_cache; /* pointer to compiler's cache */
1722   void *data; /* pointer to global data */
1723   struct kmp_cached_addr *next; /* pointer to next cached address */
1724 } kmp_cached_addr_t;
1725 
1726 struct private_data {
1727   struct private_data *next; /* The next descriptor in the list      */
1728   void *data; /* The data buffer for this descriptor  */
1729   int more; /* The repeat count for this descriptor */
1730   size_t size; /* The data size for this descriptor    */
1731 };
1732 
1733 struct private_common {
1734   struct private_common *next;
1735   struct private_common *link;
1736   void *gbl_addr;
1737   void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1738   size_t cmn_size;
1739 };
1740 
1741 struct shared_common {
1742   struct shared_common *next;
1743   struct private_data *pod_init;
1744   void *obj_init;
1745   void *gbl_addr;
1746   union {
1747     kmpc_ctor ctor;
1748     kmpc_ctor_vec ctorv;
1749   } ct;
1750   union {
1751     kmpc_cctor cctor;
1752     kmpc_cctor_vec cctorv;
1753   } cct;
1754   union {
1755     kmpc_dtor dtor;
1756     kmpc_dtor_vec dtorv;
1757   } dt;
1758   size_t vec_len;
1759   int is_vec;
1760   size_t cmn_size;
1761 };
1762 
1763 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1764 #define KMP_HASH_TABLE_SIZE                                                    \
1765   (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1766 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1767 #define KMP_HASH(x)                                                            \
1768   ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1769 
1770 struct common_table {
1771   struct private_common *data[KMP_HASH_TABLE_SIZE];
1772 };
1773 
1774 struct shared_table {
1775   struct shared_common *data[KMP_HASH_TABLE_SIZE];
1776 };
1777 
1778 /* ------------------------------------------------------------------------ */
1779 
1780 #if KMP_USE_HIER_SCHED
1781 // Shared barrier data that exists inside a single unit of the scheduling
1782 // hierarchy
1783 typedef struct kmp_hier_private_bdata_t {
1784   kmp_int32 num_active;
1785   kmp_uint64 index;
1786   kmp_uint64 wait_val[2];
1787 } kmp_hier_private_bdata_t;
1788 #endif
1789 
1790 typedef struct kmp_sched_flags {
1791   unsigned ordered : 1;
1792   unsigned nomerge : 1;
1793   unsigned contains_last : 1;
1794 #if KMP_USE_HIER_SCHED
1795   unsigned use_hier : 1;
1796   unsigned unused : 28;
1797 #else
1798   unsigned unused : 29;
1799 #endif
1800 } kmp_sched_flags_t;
1801 
1802 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1803 
1804 #if KMP_STATIC_STEAL_ENABLED
1805 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1806   kmp_int32 count;
1807   kmp_int32 ub;
1808   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1809   kmp_int32 lb;
1810   kmp_int32 st;
1811   kmp_int32 tc;
1812   kmp_lock_t *steal_lock; // lock used for chunk stealing
1813   // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1814   //    a) parm3 is properly aligned and
1815   //    b) all parm1-4 are on the same cache line.
1816   // Because of parm1-4 are used together, performance seems to be better
1817   // if they are on the same cache line (not measured though).
1818 
1819   struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1820     kmp_int32 parm1; //     structures in kmp_dispatch.cpp. This should
1821     kmp_int32 parm2; //     make no real change at least while padding is off.
1822     kmp_int32 parm3;
1823     kmp_int32 parm4;
1824   };
1825 
1826   kmp_uint32 ordered_lower;
1827   kmp_uint32 ordered_upper;
1828 #if KMP_OS_WINDOWS
1829   kmp_int32 last_upper;
1830 #endif /* KMP_OS_WINDOWS */
1831 } dispatch_private_info32_t;
1832 
1833 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1834   kmp_int64 count; // current chunk number for static & static-steal scheduling
1835   kmp_int64 ub; /* upper-bound */
1836   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1837   kmp_int64 lb; /* lower-bound */
1838   kmp_int64 st; /* stride */
1839   kmp_int64 tc; /* trip count (number of iterations) */
1840   kmp_lock_t *steal_lock; // lock used for chunk stealing
1841   /* parm[1-4] are used in different ways by different scheduling algorithms */
1842 
1843   // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1844   //    a) parm3 is properly aligned and
1845   //    b) all parm1-4 are in the same cache line.
1846   // Because of parm1-4 are used together, performance seems to be better
1847   // if they are in the same line (not measured though).
1848 
1849   struct KMP_ALIGN(32) {
1850     kmp_int64 parm1;
1851     kmp_int64 parm2;
1852     kmp_int64 parm3;
1853     kmp_int64 parm4;
1854   };
1855 
1856   kmp_uint64 ordered_lower;
1857   kmp_uint64 ordered_upper;
1858 #if KMP_OS_WINDOWS
1859   kmp_int64 last_upper;
1860 #endif /* KMP_OS_WINDOWS */
1861 } dispatch_private_info64_t;
1862 #else /* KMP_STATIC_STEAL_ENABLED */
1863 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1864   kmp_int32 lb;
1865   kmp_int32 ub;
1866   kmp_int32 st;
1867   kmp_int32 tc;
1868 
1869   kmp_int32 parm1;
1870   kmp_int32 parm2;
1871   kmp_int32 parm3;
1872   kmp_int32 parm4;
1873 
1874   kmp_int32 count;
1875 
1876   kmp_uint32 ordered_lower;
1877   kmp_uint32 ordered_upper;
1878 #if KMP_OS_WINDOWS
1879   kmp_int32 last_upper;
1880 #endif /* KMP_OS_WINDOWS */
1881 } dispatch_private_info32_t;
1882 
1883 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1884   kmp_int64 lb; /* lower-bound */
1885   kmp_int64 ub; /* upper-bound */
1886   kmp_int64 st; /* stride */
1887   kmp_int64 tc; /* trip count (number of iterations) */
1888 
1889   /* parm[1-4] are used in different ways by different scheduling algorithms */
1890   kmp_int64 parm1;
1891   kmp_int64 parm2;
1892   kmp_int64 parm3;
1893   kmp_int64 parm4;
1894 
1895   kmp_int64 count; /* current chunk number for static scheduling */
1896 
1897   kmp_uint64 ordered_lower;
1898   kmp_uint64 ordered_upper;
1899 #if KMP_OS_WINDOWS
1900   kmp_int64 last_upper;
1901 #endif /* KMP_OS_WINDOWS */
1902 } dispatch_private_info64_t;
1903 #endif /* KMP_STATIC_STEAL_ENABLED */
1904 
1905 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1906   union private_info {
1907     dispatch_private_info32_t p32;
1908     dispatch_private_info64_t p64;
1909   } u;
1910   enum sched_type schedule; /* scheduling algorithm */
1911   kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1912   std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1913   kmp_int32 ordered_bumped;
1914   // Stack of buffers for nest of serial regions
1915   struct dispatch_private_info *next;
1916   kmp_int32 type_size; /* the size of types in private_info */
1917 #if KMP_USE_HIER_SCHED
1918   kmp_int32 hier_id;
1919   void *parent; /* hierarchical scheduling parent pointer */
1920 #endif
1921   enum cons_type pushed_ws;
1922 } dispatch_private_info_t;
1923 
1924 typedef struct dispatch_shared_info32 {
1925   /* chunk index under dynamic, number of idle threads under static-steal;
1926      iteration index otherwise */
1927   volatile kmp_uint32 iteration;
1928   volatile kmp_int32 num_done;
1929   volatile kmp_uint32 ordered_iteration;
1930   // Dummy to retain the structure size after making ordered_iteration scalar
1931   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1932 } dispatch_shared_info32_t;
1933 
1934 typedef struct dispatch_shared_info64 {
1935   /* chunk index under dynamic, number of idle threads under static-steal;
1936      iteration index otherwise */
1937   volatile kmp_uint64 iteration;
1938   volatile kmp_int64 num_done;
1939   volatile kmp_uint64 ordered_iteration;
1940   // Dummy to retain the structure size after making ordered_iteration scalar
1941   kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1942 } dispatch_shared_info64_t;
1943 
1944 typedef struct dispatch_shared_info {
1945   union shared_info {
1946     dispatch_shared_info32_t s32;
1947     dispatch_shared_info64_t s64;
1948   } u;
1949   volatile kmp_uint32 buffer_index;
1950   volatile kmp_int32 doacross_buf_idx; // teamwise index
1951   volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1952   kmp_int32 doacross_num_done; // count finished threads
1953 #if KMP_USE_HIER_SCHED
1954   void *hier;
1955 #endif
1956 #if KMP_USE_HWLOC
1957   // When linking with libhwloc, the ORDERED EPCC test slows down on big
1958   // machines (> 48 cores). Performance analysis showed that a cache thrash
1959   // was occurring and this padding helps alleviate the problem.
1960   char padding[64];
1961 #endif
1962 } dispatch_shared_info_t;
1963 
1964 typedef struct kmp_disp {
1965   /* Vector for ORDERED SECTION */
1966   void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1967   /* Vector for END ORDERED SECTION */
1968   void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1969 
1970   dispatch_shared_info_t *th_dispatch_sh_current;
1971   dispatch_private_info_t *th_dispatch_pr_current;
1972 
1973   dispatch_private_info_t *th_disp_buffer;
1974   kmp_uint32 th_disp_index;
1975   kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1976   volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1977   kmp_int64 *th_doacross_info; // info on loop bounds
1978 #if KMP_USE_INTERNODE_ALIGNMENT
1979   char more_padding[INTERNODE_CACHE_LINE];
1980 #endif
1981 } kmp_disp_t;
1982 
1983 /* ------------------------------------------------------------------------ */
1984 /* Barrier stuff */
1985 
1986 /* constants for barrier state update */
1987 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1988 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1989 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1990 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1991 
1992 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1993 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1994 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1995 
1996 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1997 #error "Barrier sleep bit must be smaller than barrier bump bit"
1998 #endif
1999 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2000 #error "Barrier unused bit must be smaller than barrier bump bit"
2001 #endif
2002 
2003 // Constants for release barrier wait state: currently, hierarchical only
2004 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2005 #define KMP_BARRIER_OWN_FLAG                                                   \
2006   1 // Normal state; worker waiting on own b_go flag in release
2007 #define KMP_BARRIER_PARENT_FLAG                                                \
2008   2 // Special state; worker waiting on parent's b_go flag in release
2009 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG                                         \
2010   3 // Special state; tells worker to shift from parent to own b_go
2011 #define KMP_BARRIER_SWITCHING                                                  \
2012   4 // Special state; worker resets appropriate flag on wake-up
2013 
2014 #define KMP_NOT_SAFE_TO_REAP                                                   \
2015   0 // Thread th_reap_state: not safe to reap (tasking)
2016 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2017 
2018 // The flag_type describes the storage used for the flag.
2019 enum flag_type {
2020   flag32, /**< atomic 32 bit flags */
2021   flag64, /**< 64 bit flags */
2022   atomic_flag64, /**< atomic 64 bit flags */
2023   flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */
2024   flag_unset
2025 };
2026 
2027 enum barrier_type {
2028   bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2029                            barriers if enabled) */
2030   bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2031 #if KMP_FAST_REDUCTION_BARRIER
2032   bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2033 #endif // KMP_FAST_REDUCTION_BARRIER
2034   bs_last_barrier /* Just a placeholder to mark the end */
2035 };
2036 
2037 // to work with reduction barriers just like with plain barriers
2038 #if !KMP_FAST_REDUCTION_BARRIER
2039 #define bs_reduction_barrier bs_plain_barrier
2040 #endif // KMP_FAST_REDUCTION_BARRIER
2041 
2042 typedef enum kmp_bar_pat { /* Barrier communication patterns */
2043                            bp_linear_bar =
2044                                0, /* Single level (degenerate) tree */
2045                            bp_tree_bar =
2046                                1, /* Balanced tree with branching factor 2^n */
2047                            bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2048                                                 branching factor 2^n */
2049                            bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2050                            bp_dist_bar = 4, /* Distributed barrier */
2051                            bp_last_bar /* Placeholder to mark the end */
2052 } kmp_bar_pat_e;
2053 
2054 #define KMP_BARRIER_ICV_PUSH 1
2055 
2056 /* Record for holding the values of the internal controls stack records */
2057 typedef struct kmp_internal_control {
2058   int serial_nesting_level; /* corresponds to the value of the
2059                                th_team_serialized field */
2060   kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2061                        thread) */
2062   kmp_int8
2063       bt_set; /* internal control for whether blocktime is explicitly set */
2064   int blocktime; /* internal control for blocktime */
2065 #if KMP_USE_MONITOR
2066   int bt_intervals; /* internal control for blocktime intervals */
2067 #endif
2068   int nproc; /* internal control for #threads for next parallel region (per
2069                 thread) */
2070   int thread_limit; /* internal control for thread-limit-var */
2071   int max_active_levels; /* internal control for max_active_levels */
2072   kmp_r_sched_t
2073       sched; /* internal control for runtime schedule {sched,chunk} pair */
2074   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
2075   kmp_int32 default_device; /* internal control for default device */
2076   struct kmp_internal_control *next;
2077 } kmp_internal_control_t;
2078 
2079 static inline void copy_icvs(kmp_internal_control_t *dst,
2080                              kmp_internal_control_t *src) {
2081   *dst = *src;
2082 }
2083 
2084 /* Thread barrier needs volatile barrier fields */
2085 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2086   // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2087   // uses of it). It is not explicitly aligned below, because we *don't* want
2088   // it to be padded -- instead, we fit b_go into the same cache line with
2089   // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2090   kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2091   // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2092   // same NGO store
2093   volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2094   KMP_ALIGN_CACHE volatile kmp_uint64
2095       b_arrived; // STATE => task reached synch point.
2096   kmp_uint32 *skip_per_level;
2097   kmp_uint32 my_level;
2098   kmp_int32 parent_tid;
2099   kmp_int32 old_tid;
2100   kmp_uint32 depth;
2101   struct kmp_bstate *parent_bar;
2102   kmp_team_t *team;
2103   kmp_uint64 leaf_state;
2104   kmp_uint32 nproc;
2105   kmp_uint8 base_leaf_kids;
2106   kmp_uint8 leaf_kids;
2107   kmp_uint8 offset;
2108   kmp_uint8 wait_flag;
2109   kmp_uint8 use_oncore_barrier;
2110 #if USE_DEBUGGER
2111   // The following field is intended for the debugger solely. Only the worker
2112   // thread itself accesses this field: the worker increases it by 1 when it
2113   // arrives to a barrier.
2114   KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2115 #endif /* USE_DEBUGGER */
2116 } kmp_bstate_t;
2117 
2118 union KMP_ALIGN_CACHE kmp_barrier_union {
2119   double b_align; /* use worst case alignment */
2120   char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2121   kmp_bstate_t bb;
2122 };
2123 
2124 typedef union kmp_barrier_union kmp_balign_t;
2125 
2126 /* Team barrier needs only non-volatile arrived counter */
2127 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2128   double b_align; /* use worst case alignment */
2129   char b_pad[CACHE_LINE];
2130   struct {
2131     kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2132 #if USE_DEBUGGER
2133     // The following two fields are indended for the debugger solely. Only
2134     // primary thread of the team accesses these fields: the first one is
2135     // increased by 1 when the primary thread arrives to a barrier, the second
2136     // one is increased by one when all the threads arrived.
2137     kmp_uint b_master_arrived;
2138     kmp_uint b_team_arrived;
2139 #endif
2140   };
2141 };
2142 
2143 typedef union kmp_barrier_team_union kmp_balign_team_t;
2144 
2145 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2146    threads when a condition changes.  This is to workaround an NPTL bug where
2147    padding was added to pthread_cond_t which caused the initialization routine
2148    to write outside of the structure if compiled on pre-NPTL threads.  */
2149 #if KMP_OS_WINDOWS
2150 typedef struct kmp_win32_mutex {
2151   /* The Lock */
2152   CRITICAL_SECTION cs;
2153 } kmp_win32_mutex_t;
2154 
2155 typedef struct kmp_win32_cond {
2156   /* Count of the number of waiters. */
2157   int waiters_count_;
2158 
2159   /* Serialize access to <waiters_count_> */
2160   kmp_win32_mutex_t waiters_count_lock_;
2161 
2162   /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2163   int release_count_;
2164 
2165   /* Keeps track of the current "generation" so that we don't allow */
2166   /* one thread to steal all the "releases" from the broadcast. */
2167   int wait_generation_count_;
2168 
2169   /* A manual-reset event that's used to block and release waiting threads. */
2170   HANDLE event_;
2171 } kmp_win32_cond_t;
2172 #endif
2173 
2174 #if KMP_OS_UNIX
2175 
2176 union KMP_ALIGN_CACHE kmp_cond_union {
2177   double c_align;
2178   char c_pad[CACHE_LINE];
2179   pthread_cond_t c_cond;
2180 };
2181 
2182 typedef union kmp_cond_union kmp_cond_align_t;
2183 
2184 union KMP_ALIGN_CACHE kmp_mutex_union {
2185   double m_align;
2186   char m_pad[CACHE_LINE];
2187   pthread_mutex_t m_mutex;
2188 };
2189 
2190 typedef union kmp_mutex_union kmp_mutex_align_t;
2191 
2192 #endif /* KMP_OS_UNIX */
2193 
2194 typedef struct kmp_desc_base {
2195   void *ds_stackbase;
2196   size_t ds_stacksize;
2197   int ds_stackgrow;
2198   kmp_thread_t ds_thread;
2199   volatile int ds_tid;
2200   int ds_gtid;
2201 #if KMP_OS_WINDOWS
2202   volatile int ds_alive;
2203   DWORD ds_thread_id;
2204 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2205    However, debugger support (libomp_db) cannot work with handles, because they
2206    uncomparable. For example, debugger requests info about thread with handle h.
2207    h is valid within debugger process, and meaningless within debugee process.
2208    Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2209    within debugee process, but it is a *new* handle which does *not* equal to
2210    any other handle in debugee... The only way to compare handles is convert
2211    them to system-wide ids. GetThreadId() function is available only in
2212    Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2213    on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2214    thread id by call to GetCurrentThreadId() from within the thread and save it
2215    to let libomp_db identify threads.  */
2216 #endif /* KMP_OS_WINDOWS */
2217 } kmp_desc_base_t;
2218 
2219 typedef union KMP_ALIGN_CACHE kmp_desc {
2220   double ds_align; /* use worst case alignment */
2221   char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2222   kmp_desc_base_t ds;
2223 } kmp_desc_t;
2224 
2225 typedef struct kmp_local {
2226   volatile int this_construct; /* count of single's encountered by thread */
2227   void *reduce_data;
2228 #if KMP_USE_BGET
2229   void *bget_data;
2230   void *bget_list;
2231 #if !USE_CMP_XCHG_FOR_BGET
2232 #ifdef USE_QUEUING_LOCK_FOR_BGET
2233   kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2234 #else
2235   kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2236 // bootstrap lock so we can use it at library
2237 // shutdown.
2238 #endif /* USE_LOCK_FOR_BGET */
2239 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2240 #endif /* KMP_USE_BGET */
2241 
2242   PACKED_REDUCTION_METHOD_T
2243   packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2244                               __kmpc_end_reduce*() */
2245 
2246 } kmp_local_t;
2247 
2248 #define KMP_CHECK_UPDATE(a, b)                                                 \
2249   if ((a) != (b))                                                              \
2250   (a) = (b)
2251 #define KMP_CHECK_UPDATE_SYNC(a, b)                                            \
2252   if ((a) != (b))                                                              \
2253   TCW_SYNC_PTR((a), (b))
2254 
2255 #define get__blocktime(xteam, xtid)                                            \
2256   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2257 #define get__bt_set(xteam, xtid)                                               \
2258   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2259 #if KMP_USE_MONITOR
2260 #define get__bt_intervals(xteam, xtid)                                         \
2261   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2262 #endif
2263 
2264 #define get__dynamic_2(xteam, xtid)                                            \
2265   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2266 #define get__nproc_2(xteam, xtid)                                              \
2267   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2268 #define get__sched_2(xteam, xtid)                                              \
2269   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2270 
2271 #define set__blocktime_team(xteam, xtid, xval)                                 \
2272   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) =     \
2273        (xval))
2274 
2275 #if KMP_USE_MONITOR
2276 #define set__bt_intervals_team(xteam, xtid, xval)                              \
2277   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) =  \
2278        (xval))
2279 #endif
2280 
2281 #define set__bt_set_team(xteam, xtid, xval)                                    \
2282   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2283 
2284 #define set__dynamic(xthread, xval)                                            \
2285   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2286 #define get__dynamic(xthread)                                                  \
2287   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2288 
2289 #define set__nproc(xthread, xval)                                              \
2290   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2291 
2292 #define set__thread_limit(xthread, xval)                                       \
2293   (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2294 
2295 #define set__max_active_levels(xthread, xval)                                  \
2296   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2297 
2298 #define get__max_active_levels(xthread)                                        \
2299   ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2300 
2301 #define set__sched(xthread, xval)                                              \
2302   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2303 
2304 #define set__proc_bind(xthread, xval)                                          \
2305   (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2306 #define get__proc_bind(xthread)                                                \
2307   ((xthread)->th.th_current_task->td_icvs.proc_bind)
2308 
2309 // OpenMP tasking data structures
2310 
2311 typedef enum kmp_tasking_mode {
2312   tskm_immediate_exec = 0,
2313   tskm_extra_barrier = 1,
2314   tskm_task_teams = 2,
2315   tskm_max = 2
2316 } kmp_tasking_mode_t;
2317 
2318 extern kmp_tasking_mode_t
2319     __kmp_tasking_mode; /* determines how/when to execute tasks */
2320 extern int __kmp_task_stealing_constraint;
2321 extern int __kmp_enable_task_throttling;
2322 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2323 // specified, defaults to 0 otherwise
2324 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2325 extern kmp_int32 __kmp_max_task_priority;
2326 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2327 extern kmp_uint64 __kmp_taskloop_min_tasks;
2328 
2329 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2330    taskdata first */
2331 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2332 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2333 
2334 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2335 // were spawned and queued since the previous barrier release.
2336 #define KMP_TASKING_ENABLED(task_team)                                         \
2337   (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2338 /*!
2339 @ingroup BASIC_TYPES
2340 @{
2341 */
2342 
2343 /*!
2344  */
2345 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2346 
2347 typedef union kmp_cmplrdata {
2348   kmp_int32 priority; /**< priority specified by user for the task */
2349   kmp_routine_entry_t
2350       destructors; /* pointer to function to invoke deconstructors of
2351                       firstprivate C++ objects */
2352   /* future data */
2353 } kmp_cmplrdata_t;
2354 
2355 /*  sizeof_kmp_task_t passed as arg to kmpc_omp_task call  */
2356 /*!
2357  */
2358 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2359   void *shareds; /**< pointer to block of pointers to shared vars   */
2360   kmp_routine_entry_t
2361       routine; /**< pointer to routine to call for executing task */
2362   kmp_int32 part_id; /**< part id for the task                          */
2363   kmp_cmplrdata_t
2364       data1; /* Two known optional additions: destructors and priority */
2365   kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2366   /* future data */
2367   /*  private vars  */
2368 } kmp_task_t;
2369 
2370 /*!
2371 @}
2372 */
2373 
2374 typedef struct kmp_taskgroup {
2375   std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2376   std::atomic<kmp_int32>
2377       cancel_request; // request for cancellation of this taskgroup
2378   struct kmp_taskgroup *parent; // parent taskgroup
2379   // Block of data to perform task reduction
2380   void *reduce_data; // reduction related info
2381   kmp_int32 reduce_num_data; // number of data items to reduce
2382   uintptr_t *gomp_data; // gomp reduction data
2383 } kmp_taskgroup_t;
2384 
2385 // forward declarations
2386 typedef union kmp_depnode kmp_depnode_t;
2387 typedef struct kmp_depnode_list kmp_depnode_list_t;
2388 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2389 
2390 // macros for checking dep flag as an integer
2391 #define KMP_DEP_IN 0x1
2392 #define KMP_DEP_OUT 0x2
2393 #define KMP_DEP_INOUT 0x3
2394 #define KMP_DEP_MTX 0x4
2395 #define KMP_DEP_SET 0x8
2396 #define KMP_DEP_ALL 0x80
2397 // Compiler sends us this info:
2398 typedef struct kmp_depend_info {
2399   kmp_intptr_t base_addr;
2400   size_t len;
2401   union {
2402     kmp_uint8 flag; // flag as an unsigned char
2403     struct { // flag as a set of 8 bits
2404       unsigned in : 1;
2405       unsigned out : 1;
2406       unsigned mtx : 1;
2407       unsigned set : 1;
2408       unsigned unused : 3;
2409       unsigned all : 1;
2410     } flags;
2411   };
2412 } kmp_depend_info_t;
2413 
2414 // Internal structures to work with task dependencies:
2415 struct kmp_depnode_list {
2416   kmp_depnode_t *node;
2417   kmp_depnode_list_t *next;
2418 };
2419 
2420 // Max number of mutexinoutset dependencies per node
2421 #define MAX_MTX_DEPS 4
2422 
2423 typedef struct kmp_base_depnode {
2424   kmp_depnode_list_t *successors; /* used under lock */
2425   kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2426   kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2427   kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2428   kmp_lock_t lock; /* guards shared fields: task, successors */
2429 #if KMP_SUPPORT_GRAPH_OUTPUT
2430   kmp_uint32 id;
2431 #endif
2432   std::atomic<kmp_int32> npredecessors;
2433   std::atomic<kmp_int32> nrefs;
2434 } kmp_base_depnode_t;
2435 
2436 union KMP_ALIGN_CACHE kmp_depnode {
2437   double dn_align; /* use worst case alignment */
2438   char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2439   kmp_base_depnode_t dn;
2440 };
2441 
2442 struct kmp_dephash_entry {
2443   kmp_intptr_t addr;
2444   kmp_depnode_t *last_out;
2445   kmp_depnode_list_t *last_set;
2446   kmp_depnode_list_t *prev_set;
2447   kmp_uint8 last_flag;
2448   kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2449   kmp_dephash_entry_t *next_in_bucket;
2450 };
2451 
2452 typedef struct kmp_dephash {
2453   kmp_dephash_entry_t **buckets;
2454   size_t size;
2455   kmp_depnode_t *last_all;
2456   size_t generation;
2457   kmp_uint32 nelements;
2458   kmp_uint32 nconflicts;
2459 } kmp_dephash_t;
2460 
2461 typedef struct kmp_task_affinity_info {
2462   kmp_intptr_t base_addr;
2463   size_t len;
2464   struct {
2465     bool flag1 : 1;
2466     bool flag2 : 1;
2467     kmp_int32 reserved : 30;
2468   } flags;
2469 } kmp_task_affinity_info_t;
2470 
2471 typedef enum kmp_event_type_t {
2472   KMP_EVENT_UNINITIALIZED = 0,
2473   KMP_EVENT_ALLOW_COMPLETION = 1
2474 } kmp_event_type_t;
2475 
2476 typedef struct {
2477   kmp_event_type_t type;
2478   kmp_tas_lock_t lock;
2479   union {
2480     kmp_task_t *task;
2481   } ed;
2482 } kmp_event_t;
2483 
2484 #ifdef BUILD_TIED_TASK_STACK
2485 
2486 /* Tied Task stack definitions */
2487 typedef struct kmp_stack_block {
2488   kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2489   struct kmp_stack_block *sb_next;
2490   struct kmp_stack_block *sb_prev;
2491 } kmp_stack_block_t;
2492 
2493 typedef struct kmp_task_stack {
2494   kmp_stack_block_t ts_first_block; // first block of stack entries
2495   kmp_taskdata_t **ts_top; // pointer to the top of stack
2496   kmp_int32 ts_entries; // number of entries on the stack
2497 } kmp_task_stack_t;
2498 
2499 #endif // BUILD_TIED_TASK_STACK
2500 
2501 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2502   /* Compiler flags */ /* Total compiler flags must be 16 bits */
2503   unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2504   unsigned final : 1; /* task is final(1) so execute immediately */
2505   unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2506                               code path */
2507   unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2508                                      invoke destructors from the runtime */
2509   unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2510                          context of the RTL) */
2511   unsigned priority_specified : 1; /* set if the compiler provides priority
2512                                       setting for the task */
2513   unsigned detachable : 1; /* 1 == can detach */
2514   unsigned hidden_helper : 1; /* 1 == hidden helper task */
2515   unsigned reserved : 8; /* reserved for compiler use */
2516 
2517   /* Library flags */ /* Total library flags must be 16 bits */
2518   unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2519   unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2520   unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2521   // (1) or may be deferred (0)
2522   unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2523   // (0) [>= 2 threads]
2524   /* If either team_serial or tasking_ser is set, task team may be NULL */
2525   /* Task State Flags: */
2526   unsigned started : 1; /* 1==started, 0==not started     */
2527   unsigned executing : 1; /* 1==executing, 0==not executing */
2528   unsigned complete : 1; /* 1==complete, 0==not complete   */
2529   unsigned freed : 1; /* 1==freed, 0==allocated        */
2530   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2531   unsigned reserved31 : 7; /* reserved for library use */
2532 
2533 } kmp_tasking_flags_t;
2534 
2535 typedef struct kmp_target_data {
2536   void *async_handle; // libomptarget async handle for task completion query
2537 } kmp_target_data_t;
2538 
2539 struct kmp_taskdata { /* aligned during dynamic allocation       */
2540   kmp_int32 td_task_id; /* id, assigned by debugger                */
2541   kmp_tasking_flags_t td_flags; /* task flags                              */
2542   kmp_team_t *td_team; /* team for this task                      */
2543   kmp_info_p *td_alloc_thread; /* thread that allocated data structures   */
2544   /* Currently not used except for perhaps IDB */
2545   kmp_taskdata_t *td_parent; /* parent task                             */
2546   kmp_int32 td_level; /* task nesting level                      */
2547   std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2548   ident_t *td_ident; /* task identifier                         */
2549   // Taskwait data.
2550   ident_t *td_taskwait_ident;
2551   kmp_uint32 td_taskwait_counter;
2552   kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2553   KMP_ALIGN_CACHE kmp_internal_control_t
2554       td_icvs; /* Internal control variables for the task */
2555   KMP_ALIGN_CACHE std::atomic<kmp_int32>
2556       td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2557                                    deallocated */
2558   std::atomic<kmp_int32>
2559       td_incomplete_child_tasks; /* Child tasks not yet complete */
2560   kmp_taskgroup_t
2561       *td_taskgroup; // Each task keeps pointer to its current taskgroup
2562   kmp_dephash_t
2563       *td_dephash; // Dependencies for children tasks are tracked from here
2564   kmp_depnode_t
2565       *td_depnode; // Pointer to graph node if this task has dependencies
2566   kmp_task_team_t *td_task_team;
2567   size_t td_size_alloc; // Size of task structure, including shareds etc.
2568 #if defined(KMP_GOMP_COMPAT)
2569   // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2570   kmp_int32 td_size_loop_bounds;
2571 #endif
2572   kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2573 #if defined(KMP_GOMP_COMPAT)
2574   // GOMP sends in a copy function for copy constructors
2575   void (*td_copy_func)(void *, void *);
2576 #endif
2577   kmp_event_t td_allow_completion_event;
2578 #if OMPT_SUPPORT
2579   ompt_task_info_t ompt_task_info;
2580 #endif
2581   kmp_target_data_t td_target_data;
2582 }; // struct kmp_taskdata
2583 
2584 // Make sure padding above worked
2585 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2586 
2587 // Data for task team but per thread
2588 typedef struct kmp_base_thread_data {
2589   kmp_info_p *td_thr; // Pointer back to thread info
2590   // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2591   // queued?
2592   kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2593   kmp_taskdata_t *
2594       *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2595   kmp_int32 td_deque_size; // Size of deck
2596   kmp_uint32 td_deque_head; // Head of deque (will wrap)
2597   kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2598   kmp_int32 td_deque_ntasks; // Number of tasks in deque
2599   // GEH: shouldn't this be volatile since used in while-spin?
2600   kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2601 #ifdef BUILD_TIED_TASK_STACK
2602   kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2603 // scheduling constraint
2604 #endif // BUILD_TIED_TASK_STACK
2605 } kmp_base_thread_data_t;
2606 
2607 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2608 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2609 
2610 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2611 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2612 
2613 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2614   kmp_base_thread_data_t td;
2615   double td_align; /* use worst case alignment */
2616   char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2617 } kmp_thread_data_t;
2618 
2619 typedef struct kmp_task_pri {
2620   kmp_thread_data_t td;
2621   kmp_int32 priority;
2622   kmp_task_pri *next;
2623 } kmp_task_pri_t;
2624 
2625 // Data for task teams which are used when tasking is enabled for the team
2626 typedef struct kmp_base_task_team {
2627   kmp_bootstrap_lock_t
2628       tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2629   /* must be bootstrap lock since used at library shutdown*/
2630 
2631   // TODO: check performance vs kmp_tas_lock_t
2632   kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2633   kmp_task_pri_t *tt_task_pri_list;
2634 
2635   kmp_task_team_t *tt_next; /* For linking the task team free list */
2636   kmp_thread_data_t
2637       *tt_threads_data; /* Array of per-thread structures for task team */
2638   /* Data survives task team deallocation */
2639   kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2640                                executing this team? */
2641   /* TRUE means tt_threads_data is set up and initialized */
2642   kmp_int32 tt_nproc; /* #threads in team           */
2643   kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2644   kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2645   kmp_int32 tt_untied_task_encountered;
2646   std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2647   // There is hidden helper thread encountered in this task team so that we must
2648   // wait when waiting on task team
2649   kmp_int32 tt_hidden_helper_task_encountered;
2650 
2651   KMP_ALIGN_CACHE
2652   std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2653 
2654   KMP_ALIGN_CACHE
2655   volatile kmp_uint32
2656       tt_active; /* is the team still actively executing tasks */
2657 } kmp_base_task_team_t;
2658 
2659 union KMP_ALIGN_CACHE kmp_task_team {
2660   kmp_base_task_team_t tt;
2661   double tt_align; /* use worst case alignment */
2662   char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2663 };
2664 
2665 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2666 // Free lists keep same-size free memory slots for fast memory allocation
2667 // routines
2668 typedef struct kmp_free_list {
2669   void *th_free_list_self; // Self-allocated tasks free list
2670   void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2671   // threads
2672   void *th_free_list_other; // Non-self free list (to be returned to owner's
2673   // sync list)
2674 } kmp_free_list_t;
2675 #endif
2676 #if KMP_NESTED_HOT_TEAMS
2677 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2678 // are not put in teams pool, and they don't put threads in threads pool.
2679 typedef struct kmp_hot_team_ptr {
2680   kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2681   kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2682 } kmp_hot_team_ptr_t;
2683 #endif
2684 typedef struct kmp_teams_size {
2685   kmp_int32 nteams; // number of teams in a league
2686   kmp_int32 nth; // number of threads in each team of the league
2687 } kmp_teams_size_t;
2688 
2689 // This struct stores a thread that acts as a "root" for a contention
2690 // group. Contention groups are rooted at kmp_root threads, but also at
2691 // each primary thread of each team created in the teams construct.
2692 // This struct therefore also stores a thread_limit associated with
2693 // that contention group, and a counter to track the number of threads
2694 // active in that contention group. Each thread has a list of these: CG
2695 // root threads have an entry in their list in which cg_root refers to
2696 // the thread itself, whereas other workers in the CG will have a
2697 // single entry where cg_root is same as the entry containing their CG
2698 // root. When a thread encounters a teams construct, it will add a new
2699 // entry to the front of its list, because it now roots a new CG.
2700 typedef struct kmp_cg_root {
2701   kmp_info_p *cg_root; // "root" thread for a contention group
2702   // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2703   // thread_limit clause for teams primary threads
2704   kmp_int32 cg_thread_limit;
2705   kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2706   struct kmp_cg_root *up; // pointer to higher level CG root in list
2707 } kmp_cg_root_t;
2708 
2709 // OpenMP thread data structures
2710 
2711 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2712   /* Start with the readonly data which is cache aligned and padded. This is
2713      written before the thread starts working by the primary thread. Uber
2714      masters may update themselves later. Usage does not consider serialized
2715      regions.  */
2716   kmp_desc_t th_info;
2717   kmp_team_p *th_team; /* team we belong to */
2718   kmp_root_p *th_root; /* pointer to root of task hierarchy */
2719   kmp_info_p *th_next_pool; /* next available thread in the pool */
2720   kmp_disp_t *th_dispatch; /* thread's dispatch data */
2721   int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2722 
2723   /* The following are cached from the team info structure */
2724   /* TODO use these in more places as determined to be needed via profiling */
2725   int th_team_nproc; /* number of threads in a team */
2726   kmp_info_p *th_team_master; /* the team's primary thread */
2727   int th_team_serialized; /* team is serialized */
2728   microtask_t th_teams_microtask; /* save entry address for teams construct */
2729   int th_teams_level; /* save initial level of teams construct */
2730 /* it is 0 on device but may be any on host */
2731 
2732 /* The blocktime info is copied from the team struct to the thread struct */
2733 /* at the start of a barrier, and the values stored in the team are used  */
2734 /* at points in the code where the team struct is no longer guaranteed    */
2735 /* to exist (from the POV of worker threads).                             */
2736 #if KMP_USE_MONITOR
2737   int th_team_bt_intervals;
2738   int th_team_bt_set;
2739 #else
2740   kmp_uint64 th_team_bt_intervals;
2741 #endif
2742 
2743 #if KMP_AFFINITY_SUPPORTED
2744   kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2745   kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2746   kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2747 #endif
2748   omp_allocator_handle_t th_def_allocator; /* default allocator */
2749   /* The data set by the primary thread at reinit, then R/W by the worker */
2750   KMP_ALIGN_CACHE int
2751       th_set_nproc; /* if > 0, then only use this request for the next fork */
2752 #if KMP_NESTED_HOT_TEAMS
2753   kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2754 #endif
2755   kmp_proc_bind_t
2756       th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2757   kmp_teams_size_t
2758       th_teams_size; /* number of teams/threads in teams construct */
2759 #if KMP_AFFINITY_SUPPORTED
2760   int th_current_place; /* place currently bound to */
2761   int th_new_place; /* place to bind to in par reg */
2762   int th_first_place; /* first place in partition */
2763   int th_last_place; /* last place in partition */
2764 #endif
2765   int th_prev_level; /* previous level for affinity format */
2766   int th_prev_num_threads; /* previous num_threads for affinity format */
2767 #if USE_ITT_BUILD
2768   kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2769   kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2770   kmp_uint64 th_frame_time; /* frame timestamp */
2771 #endif /* USE_ITT_BUILD */
2772   kmp_local_t th_local;
2773   struct private_common *th_pri_head;
2774 
2775   /* Now the data only used by the worker (after initial allocation) */
2776   /* TODO the first serial team should actually be stored in the info_t
2777      structure.  this will help reduce initial allocation overhead */
2778   KMP_ALIGN_CACHE kmp_team_p
2779       *th_serial_team; /*serialized team held in reserve*/
2780 
2781 #if OMPT_SUPPORT
2782   ompt_thread_info_t ompt_thread_info;
2783 #endif
2784 
2785   /* The following are also read by the primary thread during reinit */
2786   struct common_table *th_pri_common;
2787 
2788   volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2789   /* while awaiting queuing lock acquire */
2790 
2791   volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2792   flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2793 
2794   ident_t *th_ident;
2795   unsigned th_x; // Random number generator data
2796   unsigned th_a; // Random number generator data
2797 
2798   /* Tasking-related data for the thread */
2799   kmp_task_team_t *th_task_team; // Task team struct
2800   kmp_taskdata_t *th_current_task; // Innermost Task being executed
2801   kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2802   kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2803   // at nested levels
2804   kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2805   kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2806   kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2807   // tasking, thus safe to reap
2808 
2809   /* More stuff for keeping track of active/sleeping threads (this part is
2810      written by the worker thread) */
2811   kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2812   int th_active; // ! sleeping; 32 bits for TCR/TCW
2813   std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2814   // 0 = not used in team; 1 = used in team;
2815   // 2 = transitioning to not used in team; 3 = transitioning to used in team
2816   struct cons_header *th_cons; // used for consistency check
2817 #if KMP_USE_HIER_SCHED
2818   // used for hierarchical scheduling
2819   kmp_hier_private_bdata_t *th_hier_bar_data;
2820 #endif
2821 
2822   /* Add the syncronizing data which is cache aligned and padded. */
2823   KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2824 
2825   KMP_ALIGN_CACHE volatile kmp_int32
2826       th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2827 
2828 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2829 #define NUM_LISTS 4
2830   kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2831 // allocation routines
2832 #endif
2833 
2834 #if KMP_OS_WINDOWS
2835   kmp_win32_cond_t th_suspend_cv;
2836   kmp_win32_mutex_t th_suspend_mx;
2837   std::atomic<int> th_suspend_init;
2838 #endif
2839 #if KMP_OS_UNIX
2840   kmp_cond_align_t th_suspend_cv;
2841   kmp_mutex_align_t th_suspend_mx;
2842   std::atomic<int> th_suspend_init_count;
2843 #endif
2844 
2845 #if USE_ITT_BUILD
2846   kmp_itt_mark_t th_itt_mark_single;
2847 // alignment ???
2848 #endif /* USE_ITT_BUILD */
2849 #if KMP_STATS_ENABLED
2850   kmp_stats_list *th_stats;
2851 #endif
2852 #if KMP_OS_UNIX
2853   std::atomic<bool> th_blocking;
2854 #endif
2855   kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2856 } kmp_base_info_t;
2857 
2858 typedef union KMP_ALIGN_CACHE kmp_info {
2859   double th_align; /* use worst case alignment */
2860   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2861   kmp_base_info_t th;
2862 } kmp_info_t;
2863 
2864 // OpenMP thread team data structures
2865 
2866 typedef struct kmp_base_data {
2867   volatile kmp_uint32 t_value;
2868 } kmp_base_data_t;
2869 
2870 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2871   double dt_align; /* use worst case alignment */
2872   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2873   kmp_base_data_t dt;
2874 } kmp_sleep_team_t;
2875 
2876 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2877   double dt_align; /* use worst case alignment */
2878   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2879   kmp_base_data_t dt;
2880 } kmp_ordered_team_t;
2881 
2882 typedef int (*launch_t)(int gtid);
2883 
2884 /* Minimum number of ARGV entries to malloc if necessary */
2885 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2886 
2887 // Set up how many argv pointers will fit in cache lines containing
2888 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2889 // larger value for more space between the primary write/worker read section and
2890 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2891 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2892 #define KMP_INLINE_ARGV_BYTES                                                  \
2893   (4 * CACHE_LINE -                                                            \
2894    ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) +               \
2895      sizeof(kmp_int16) + sizeof(kmp_uint32)) %                                 \
2896     CACHE_LINE))
2897 #else
2898 #define KMP_INLINE_ARGV_BYTES                                                  \
2899   (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2900 #endif
2901 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2902 
2903 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2904   // Synchronization Data
2905   // ---------------------------------------------------------------------------
2906   KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2907   kmp_balign_team_t t_bar[bs_last_barrier];
2908   std::atomic<int> t_construct; // count of single directive encountered by team
2909   char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2910 
2911   // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2912   std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2913   std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2914 
2915   // Primary thread only
2916   // ---------------------------------------------------------------------------
2917   KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2918   int t_master_this_cons; // "this_construct" single counter of primary thread
2919   // in parent team
2920   ident_t *t_ident; // if volatile, have to change too much other crud to
2921   // volatile too
2922   kmp_team_p *t_parent; // parent team
2923   kmp_team_p *t_next_pool; // next free team in the team pool
2924   kmp_disp_t *t_dispatch; // thread's dispatch data
2925   kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2926   kmp_proc_bind_t t_proc_bind; // bind type for par region
2927 #if USE_ITT_BUILD
2928   kmp_uint64 t_region_time; // region begin timestamp
2929 #endif /* USE_ITT_BUILD */
2930 
2931   // Primary thread write, workers read
2932   // --------------------------------------------------------------------------
2933   KMP_ALIGN_CACHE void **t_argv;
2934   int t_argc;
2935   int t_nproc; // number of threads in team
2936   microtask_t t_pkfn;
2937   launch_t t_invoke; // procedure to launch the microtask
2938 
2939 #if OMPT_SUPPORT
2940   ompt_team_info_t ompt_team_info;
2941   ompt_lw_taskteam_t *ompt_serialized_team_info;
2942 #endif
2943 
2944 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2945   kmp_int8 t_fp_control_saved;
2946   kmp_int8 t_pad2b;
2947   kmp_int16 t_x87_fpu_control_word; // FP control regs
2948   kmp_uint32 t_mxcsr;
2949 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2950 
2951   void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2952 
2953   KMP_ALIGN_CACHE kmp_info_t **t_threads;
2954   kmp_taskdata_t
2955       *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2956   int t_level; // nested parallel level
2957 
2958   KMP_ALIGN_CACHE int t_max_argc;
2959   int t_max_nproc; // max threads this team can handle (dynamically expandable)
2960   int t_serialized; // levels deep of serialized teams
2961   dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2962   int t_id; // team's id, assigned by debugger.
2963   int t_active_level; // nested active parallel level
2964   kmp_r_sched_t t_sched; // run-time schedule for the team
2965 #if KMP_AFFINITY_SUPPORTED
2966   int t_first_place; // first & last place in parent thread's partition.
2967   int t_last_place; // Restore these values to primary thread after par region.
2968 #endif // KMP_AFFINITY_SUPPORTED
2969   int t_display_affinity;
2970   int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2971   // omp_set_num_threads() call
2972   omp_allocator_handle_t t_def_allocator; /* default allocator */
2973 
2974 // Read/write by workers as well
2975 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2976   // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2977   // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2978   // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2979   // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2980   char dummy_padding[1024];
2981 #endif
2982   // Internal control stack for additional nested teams.
2983   KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2984   // for SERIALIZED teams nested 2 or more levels deep
2985   // typed flag to store request state of cancellation
2986   std::atomic<kmp_int32> t_cancel_request;
2987   int t_master_active; // save on fork, restore on join
2988   void *t_copypriv_data; // team specific pointer to copyprivate data array
2989 #if KMP_OS_WINDOWS
2990   std::atomic<kmp_uint32> t_copyin_counter;
2991 #endif
2992 #if USE_ITT_BUILD
2993   void *t_stack_id; // team specific stack stitching id (for ittnotify)
2994 #endif /* USE_ITT_BUILD */
2995   distributedBarrier *b; // Distributed barrier data associated with team
2996 } kmp_base_team_t;
2997 
2998 union KMP_ALIGN_CACHE kmp_team {
2999   kmp_base_team_t t;
3000   double t_align; /* use worst case alignment */
3001   char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3002 };
3003 
3004 typedef union KMP_ALIGN_CACHE kmp_time_global {
3005   double dt_align; /* use worst case alignment */
3006   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3007   kmp_base_data_t dt;
3008 } kmp_time_global_t;
3009 
3010 typedef struct kmp_base_global {
3011   /* cache-aligned */
3012   kmp_time_global_t g_time;
3013 
3014   /* non cache-aligned */
3015   volatile int g_abort;
3016   volatile int g_done;
3017 
3018   int g_dynamic;
3019   enum dynamic_mode g_dynamic_mode;
3020 } kmp_base_global_t;
3021 
3022 typedef union KMP_ALIGN_CACHE kmp_global {
3023   kmp_base_global_t g;
3024   double g_align; /* use worst case alignment */
3025   char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3026 } kmp_global_t;
3027 
3028 typedef struct kmp_base_root {
3029   // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3030   // (r_in_parallel>= 0)
3031   // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3032   // the synch overhead or keeping r_active
3033   volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3034   // keeps a count of active parallel regions per root
3035   std::atomic<int> r_in_parallel;
3036   // GEH: This is misnamed, should be r_active_levels
3037   kmp_team_t *r_root_team;
3038   kmp_team_t *r_hot_team;
3039   kmp_info_t *r_uber_thread;
3040   kmp_lock_t r_begin_lock;
3041   volatile int r_begin;
3042   int r_blocktime; /* blocktime for this root and descendants */
3043 #if KMP_AFFINITY_SUPPORTED
3044   int r_affinity_assigned;
3045 #endif // KMP_AFFINITY_SUPPORTED
3046 } kmp_base_root_t;
3047 
3048 typedef union KMP_ALIGN_CACHE kmp_root {
3049   kmp_base_root_t r;
3050   double r_align; /* use worst case alignment */
3051   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3052 } kmp_root_t;
3053 
3054 struct fortran_inx_info {
3055   kmp_int32 data;
3056 };
3057 
3058 // This list type exists to hold old __kmp_threads arrays so that
3059 // old references to them may complete while reallocation takes place when
3060 // expanding the array. The items in this list are kept alive until library
3061 // shutdown.
3062 typedef struct kmp_old_threads_list_t {
3063   kmp_info_t **threads;
3064   struct kmp_old_threads_list_t *next;
3065 } kmp_old_threads_list_t;
3066 
3067 /* ------------------------------------------------------------------------ */
3068 
3069 extern int __kmp_settings;
3070 extern int __kmp_duplicate_library_ok;
3071 #if USE_ITT_BUILD
3072 extern int __kmp_forkjoin_frames;
3073 extern int __kmp_forkjoin_frames_mode;
3074 #endif
3075 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3076 extern int __kmp_determ_red;
3077 
3078 #ifdef KMP_DEBUG
3079 extern int kmp_a_debug;
3080 extern int kmp_b_debug;
3081 extern int kmp_c_debug;
3082 extern int kmp_d_debug;
3083 extern int kmp_e_debug;
3084 extern int kmp_f_debug;
3085 #endif /* KMP_DEBUG */
3086 
3087 /* For debug information logging using rotating buffer */
3088 #define KMP_DEBUG_BUF_LINES_INIT 512
3089 #define KMP_DEBUG_BUF_LINES_MIN 1
3090 
3091 #define KMP_DEBUG_BUF_CHARS_INIT 128
3092 #define KMP_DEBUG_BUF_CHARS_MIN 2
3093 
3094 extern int
3095     __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3096 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3097 extern int
3098     __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3099 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3100                                       entry pointer */
3101 
3102 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3103 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3104                                               printed in buffer so far */
3105 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3106                                           recommended in warnings */
3107 /* end rotating debug buffer */
3108 
3109 #ifdef KMP_DEBUG
3110 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3111 
3112 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3113 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3114 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3115 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3116 extern int __kmp_par_range_lb;
3117 extern int __kmp_par_range_ub;
3118 #endif
3119 
3120 /* For printing out dynamic storage map for threads and teams */
3121 extern int
3122     __kmp_storage_map; /* True means print storage map for threads and teams */
3123 extern int __kmp_storage_map_verbose; /* True means storage map includes
3124                                          placement info */
3125 extern int __kmp_storage_map_verbose_specified;
3126 
3127 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3128 extern kmp_cpuinfo_t __kmp_cpuinfo;
3129 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3130 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3131 static inline bool __kmp_is_hybrid_cpu() { return true; }
3132 #else
3133 static inline bool __kmp_is_hybrid_cpu() { return false; }
3134 #endif
3135 
3136 extern volatile int __kmp_init_serial;
3137 extern volatile int __kmp_init_gtid;
3138 extern volatile int __kmp_init_common;
3139 extern volatile int __kmp_need_register_serial;
3140 extern volatile int __kmp_init_middle;
3141 extern volatile int __kmp_init_parallel;
3142 #if KMP_USE_MONITOR
3143 extern volatile int __kmp_init_monitor;
3144 #endif
3145 extern volatile int __kmp_init_user_locks;
3146 extern volatile int __kmp_init_hidden_helper_threads;
3147 extern int __kmp_init_counter;
3148 extern int __kmp_root_counter;
3149 extern int __kmp_version;
3150 
3151 /* list of address of allocated caches for commons */
3152 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3153 
3154 /* Barrier algorithm types and options */
3155 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3156 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3157 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3158 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3159 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3160 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3161 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3162 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3163 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3164 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3165 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3166 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3167 
3168 /* Global Locks */
3169 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3170 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3171 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3172 extern kmp_bootstrap_lock_t
3173     __kmp_exit_lock; /* exit() is not always thread-safe */
3174 #if KMP_USE_MONITOR
3175 extern kmp_bootstrap_lock_t
3176     __kmp_monitor_lock; /* control monitor thread creation */
3177 #endif
3178 extern kmp_bootstrap_lock_t
3179     __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3180                              __kmp_threads expansion to co-exist */
3181 
3182 extern kmp_lock_t __kmp_global_lock; /* control OS/global access  */
3183 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access  */
3184 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3185 
3186 extern enum library_type __kmp_library;
3187 
3188 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3189 extern enum sched_type __kmp_static; /* default static scheduling method */
3190 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3191 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3192 extern int __kmp_chunk; /* default runtime chunk size */
3193 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3194 
3195 extern size_t __kmp_stksize; /* stack size per thread         */
3196 #if KMP_USE_MONITOR
3197 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3198 #endif
3199 extern size_t __kmp_stkoffset; /* stack offset per thread       */
3200 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3201 
3202 extern size_t
3203     __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3204 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3205 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3206 extern int __kmp_env_checks; /* was KMP_CHECKS specified?    */
3207 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3208 extern int __kmp_generate_warnings; /* should we issue warnings? */
3209 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3210 
3211 #ifdef DEBUG_SUSPEND
3212 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3213 #endif
3214 
3215 extern kmp_int32 __kmp_use_yield;
3216 extern kmp_int32 __kmp_use_yield_exp_set;
3217 extern kmp_uint32 __kmp_yield_init;
3218 extern kmp_uint32 __kmp_yield_next;
3219 extern kmp_uint64 __kmp_pause_init;
3220 
3221 /* ------------------------------------------------------------------------- */
3222 extern int __kmp_allThreadsSpecified;
3223 
3224 extern size_t __kmp_align_alloc;
3225 /* following data protected by initialization routines */
3226 extern int __kmp_xproc; /* number of processors in the system */
3227 extern int __kmp_avail_proc; /* number of processors available to the process */
3228 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3229 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3230 // maximum total number of concurrently-existing threads on device
3231 extern int __kmp_max_nth;
3232 // maximum total number of concurrently-existing threads in a contention group
3233 extern int __kmp_cg_max_nth;
3234 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3235 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3236                                       __kmp_root */
3237 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3238                                    region a la OMP_NUM_THREADS */
3239 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3240                                       initialization */
3241 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3242                                  used (fixed) */
3243 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3244                                (__kmpc_threadprivate_cached()) */
3245 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3246                                     blocking (env setting) */
3247 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3248 #if KMP_USE_MONITOR
3249 extern int
3250     __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3251 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3252                                   blocking */
3253 #endif
3254 #ifdef KMP_ADJUST_BLOCKTIME
3255 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3256 #endif /* KMP_ADJUST_BLOCKTIME */
3257 #ifdef KMP_DFLT_NTH_CORES
3258 extern int __kmp_ncores; /* Total number of cores for threads placement */
3259 #endif
3260 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3261 extern int __kmp_abort_delay;
3262 
3263 extern int __kmp_need_register_atfork_specified;
3264 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3265                                           to install fork handler */
3266 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3267                                0 - not set, will be set at runtime
3268                                1 - using stack search
3269                                2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3270                                    X*) or TlsGetValue(Windows* OS))
3271                                3 - static TLS (__declspec(thread) __kmp_gtid),
3272                                    Linux* OS .so only.  */
3273 extern int
3274     __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3275 #ifdef KMP_TDATA_GTID
3276 extern KMP_THREAD_LOCAL int __kmp_gtid;
3277 #endif
3278 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3279 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3280 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3281 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3282 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3283 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3284 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3285 
3286 // max_active_levels for nested parallelism enabled by default via
3287 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3288 extern int __kmp_dflt_max_active_levels;
3289 // Indicates whether value of __kmp_dflt_max_active_levels was already
3290 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3291 extern bool __kmp_dflt_max_active_levels_set;
3292 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3293                                           concurrent execution per team */
3294 #if KMP_NESTED_HOT_TEAMS
3295 extern int __kmp_hot_teams_mode;
3296 extern int __kmp_hot_teams_max_level;
3297 #endif
3298 
3299 #if KMP_OS_LINUX
3300 extern enum clock_function_type __kmp_clock_function;
3301 extern int __kmp_clock_function_param;
3302 #endif /* KMP_OS_LINUX */
3303 
3304 #if KMP_MIC_SUPPORTED
3305 extern enum mic_type __kmp_mic_type;
3306 #endif
3307 
3308 #ifdef USE_LOAD_BALANCE
3309 extern double __kmp_load_balance_interval; // load balance algorithm interval
3310 #endif /* USE_LOAD_BALANCE */
3311 
3312 // OpenMP 3.1 - Nested num threads array
3313 typedef struct kmp_nested_nthreads_t {
3314   int *nth;
3315   int size;
3316   int used;
3317 } kmp_nested_nthreads_t;
3318 
3319 extern kmp_nested_nthreads_t __kmp_nested_nth;
3320 
3321 #if KMP_USE_ADAPTIVE_LOCKS
3322 
3323 // Parameters for the speculative lock backoff system.
3324 struct kmp_adaptive_backoff_params_t {
3325   // Number of soft retries before it counts as a hard retry.
3326   kmp_uint32 max_soft_retries;
3327   // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3328   // the right
3329   kmp_uint32 max_badness;
3330 };
3331 
3332 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3333 
3334 #if KMP_DEBUG_ADAPTIVE_LOCKS
3335 extern const char *__kmp_speculative_statsfile;
3336 #endif
3337 
3338 #endif // KMP_USE_ADAPTIVE_LOCKS
3339 
3340 extern int __kmp_display_env; /* TRUE or FALSE */
3341 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3342 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3343 extern int __kmp_nteams;
3344 extern int __kmp_teams_thread_limit;
3345 
3346 /* ------------------------------------------------------------------------- */
3347 
3348 /* the following are protected by the fork/join lock */
3349 /* write: lock  read: anytime */
3350 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3351 /* Holds old arrays of __kmp_threads until library shutdown */
3352 extern kmp_old_threads_list_t *__kmp_old_threads_list;
3353 /* read/write: lock */
3354 extern volatile kmp_team_t *__kmp_team_pool;
3355 extern volatile kmp_info_t *__kmp_thread_pool;
3356 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3357 
3358 // total num threads reachable from some root thread including all root threads
3359 extern volatile int __kmp_nth;
3360 /* total number of threads reachable from some root thread including all root
3361    threads, and those in the thread pool */
3362 extern volatile int __kmp_all_nth;
3363 extern std::atomic<int> __kmp_thread_pool_active_nth;
3364 
3365 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3366 /* end data protected by fork/join lock */
3367 /* ------------------------------------------------------------------------- */
3368 
3369 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3370 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3371 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3372 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3373 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3374 
3375 // AT: Which way is correct?
3376 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3377 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3378 #define __kmp_get_team_num_threads(gtid)                                       \
3379   (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3380 
3381 static inline bool KMP_UBER_GTID(int gtid) {
3382   KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3383   KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3384   return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3385           __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3386 }
3387 
3388 static inline int __kmp_tid_from_gtid(int gtid) {
3389   KMP_DEBUG_ASSERT(gtid >= 0);
3390   return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3391 }
3392 
3393 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3394   KMP_DEBUG_ASSERT(tid >= 0 && team);
3395   return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3396 }
3397 
3398 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3399   KMP_DEBUG_ASSERT(thr);
3400   return thr->th.th_info.ds.ds_gtid;
3401 }
3402 
3403 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3404   KMP_DEBUG_ASSERT(gtid >= 0);
3405   return __kmp_threads[gtid];
3406 }
3407 
3408 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3409   KMP_DEBUG_ASSERT(gtid >= 0);
3410   return __kmp_threads[gtid]->th.th_team;
3411 }
3412 
3413 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3414   if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3415     KMP_FATAL(ThreadIdentInvalid);
3416 }
3417 
3418 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3419 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3420 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3421 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3422 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3423 #endif
3424 
3425 #if KMP_HAVE_UMWAIT
3426 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3427 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3428 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3429 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3430 #endif
3431 
3432 /* ------------------------------------------------------------------------- */
3433 
3434 extern kmp_global_t __kmp_global; /* global status */
3435 
3436 extern kmp_info_t __kmp_monitor;
3437 // For Debugging Support Library
3438 extern std::atomic<kmp_int32> __kmp_team_counter;
3439 // For Debugging Support Library
3440 extern std::atomic<kmp_int32> __kmp_task_counter;
3441 
3442 #if USE_DEBUGGER
3443 #define _KMP_GEN_ID(counter)                                                   \
3444   (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3445 #else
3446 #define _KMP_GEN_ID(counter) (~0)
3447 #endif /* USE_DEBUGGER */
3448 
3449 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3450 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3451 
3452 /* ------------------------------------------------------------------------ */
3453 
3454 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3455                                          size_t size, char const *format, ...);
3456 
3457 extern void __kmp_serial_initialize(void);
3458 extern void __kmp_middle_initialize(void);
3459 extern void __kmp_parallel_initialize(void);
3460 
3461 extern void __kmp_internal_begin(void);
3462 extern void __kmp_internal_end_library(int gtid);
3463 extern void __kmp_internal_end_thread(int gtid);
3464 extern void __kmp_internal_end_atexit(void);
3465 extern void __kmp_internal_end_dtor(void);
3466 extern void __kmp_internal_end_dest(void *);
3467 
3468 extern int __kmp_register_root(int initial_thread);
3469 extern void __kmp_unregister_root(int gtid);
3470 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3471 
3472 extern int __kmp_ignore_mppbeg(void);
3473 extern int __kmp_ignore_mppend(void);
3474 
3475 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3476 extern void __kmp_exit_single(int gtid);
3477 
3478 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3479 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3480 
3481 #ifdef USE_LOAD_BALANCE
3482 extern int __kmp_get_load_balance(int);
3483 #endif
3484 
3485 extern int __kmp_get_global_thread_id(void);
3486 extern int __kmp_get_global_thread_id_reg(void);
3487 extern void __kmp_exit_thread(int exit_status);
3488 extern void __kmp_abort(char const *format, ...);
3489 extern void __kmp_abort_thread(void);
3490 KMP_NORETURN extern void __kmp_abort_process(void);
3491 extern void __kmp_warn(char const *format, ...);
3492 
3493 extern void __kmp_set_num_threads(int new_nth, int gtid);
3494 
3495 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3496 // registered.
3497 static inline kmp_info_t *__kmp_entry_thread() {
3498   int gtid = __kmp_entry_gtid();
3499 
3500   return __kmp_threads[gtid];
3501 }
3502 
3503 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3504 extern int __kmp_get_max_active_levels(int gtid);
3505 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3506 extern int __kmp_get_team_size(int gtid, int level);
3507 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3508 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3509 
3510 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3511 extern void __kmp_init_random(kmp_info_t *thread);
3512 
3513 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3514 extern void __kmp_adjust_num_threads(int new_nproc);
3515 extern void __kmp_check_stksize(size_t *val);
3516 
3517 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3518 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3519 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3520 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3521 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3522 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3523 
3524 #if USE_FAST_MEMORY
3525 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3526                                   size_t size KMP_SRC_LOC_DECL);
3527 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3528 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3529 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3530 #define __kmp_fast_allocate(this_thr, size)                                    \
3531   ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3532 #define __kmp_fast_free(this_thr, ptr)                                         \
3533   ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3534 #endif
3535 
3536 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3537 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3538                                   size_t elsize KMP_SRC_LOC_DECL);
3539 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3540                                    size_t size KMP_SRC_LOC_DECL);
3541 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3542 #define __kmp_thread_malloc(th, size)                                          \
3543   ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3544 #define __kmp_thread_calloc(th, nelem, elsize)                                 \
3545   ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3546 #define __kmp_thread_realloc(th, ptr, size)                                    \
3547   ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3548 #define __kmp_thread_free(th, ptr)                                             \
3549   ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3550 
3551 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3552 
3553 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3554                                  kmp_proc_bind_t proc_bind);
3555 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3556                                  int num_threads);
3557 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3558                                     int num_teams_ub, int num_threads);
3559 
3560 extern void __kmp_yield();
3561 
3562 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3563                                    enum sched_type schedule, kmp_int32 lb,
3564                                    kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3565 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3566                                     enum sched_type schedule, kmp_uint32 lb,
3567                                     kmp_uint32 ub, kmp_int32 st,
3568                                     kmp_int32 chunk);
3569 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3570                                    enum sched_type schedule, kmp_int64 lb,
3571                                    kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3572 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3573                                     enum sched_type schedule, kmp_uint64 lb,
3574                                     kmp_uint64 ub, kmp_int64 st,
3575                                     kmp_int64 chunk);
3576 
3577 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3578                                   kmp_int32 *p_last, kmp_int32 *p_lb,
3579                                   kmp_int32 *p_ub, kmp_int32 *p_st);
3580 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3581                                    kmp_int32 *p_last, kmp_uint32 *p_lb,
3582                                    kmp_uint32 *p_ub, kmp_int32 *p_st);
3583 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3584                                   kmp_int32 *p_last, kmp_int64 *p_lb,
3585                                   kmp_int64 *p_ub, kmp_int64 *p_st);
3586 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3587                                    kmp_int32 *p_last, kmp_uint64 *p_lb,
3588                                    kmp_uint64 *p_ub, kmp_int64 *p_st);
3589 
3590 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3591 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3592 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3593 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3594 
3595 #ifdef KMP_GOMP_COMPAT
3596 
3597 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3598                                       enum sched_type schedule, kmp_int32 lb,
3599                                       kmp_int32 ub, kmp_int32 st,
3600                                       kmp_int32 chunk, int push_ws);
3601 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3602                                        enum sched_type schedule, kmp_uint32 lb,
3603                                        kmp_uint32 ub, kmp_int32 st,
3604                                        kmp_int32 chunk, int push_ws);
3605 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3606                                       enum sched_type schedule, kmp_int64 lb,
3607                                       kmp_int64 ub, kmp_int64 st,
3608                                       kmp_int64 chunk, int push_ws);
3609 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3610                                        enum sched_type schedule, kmp_uint64 lb,
3611                                        kmp_uint64 ub, kmp_int64 st,
3612                                        kmp_int64 chunk, int push_ws);
3613 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3614 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3615 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3616 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3617 
3618 #endif /* KMP_GOMP_COMPAT */
3619 
3620 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3621 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3622 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3623 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3624 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3625 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3626                                kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3627                                void *obj);
3628 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3629                              kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3630 
3631 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3632                           int final_spin
3633 #if USE_ITT_BUILD
3634                           ,
3635                           void *itt_sync_obj
3636 #endif
3637 );
3638 extern void __kmp_release_64(kmp_flag_64<> *flag);
3639 
3640 extern void __kmp_infinite_loop(void);
3641 
3642 extern void __kmp_cleanup(void);
3643 
3644 #if KMP_HANDLE_SIGNALS
3645 extern int __kmp_handle_signals;
3646 extern void __kmp_install_signals(int parallel_init);
3647 extern void __kmp_remove_signals(void);
3648 #endif
3649 
3650 extern void __kmp_clear_system_time(void);
3651 extern void __kmp_read_system_time(double *delta);
3652 
3653 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3654 
3655 extern void __kmp_expand_host_name(char *buffer, size_t size);
3656 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3657 
3658 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3659 extern void
3660 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3661 #endif
3662 
3663 extern void
3664 __kmp_runtime_initialize(void); /* machine specific initialization */
3665 extern void __kmp_runtime_destroy(void);
3666 
3667 #if KMP_AFFINITY_SUPPORTED
3668 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3669                                        kmp_affin_mask_t *mask);
3670 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3671                                                   kmp_affin_mask_t *mask);
3672 extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3673 extern void __kmp_affinity_uninitialize(void);
3674 extern void __kmp_affinity_set_init_mask(
3675     int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3676 extern void __kmp_affinity_set_place(int gtid);
3677 extern void __kmp_affinity_determine_capable(const char *env_var);
3678 extern int __kmp_aux_set_affinity(void **mask);
3679 extern int __kmp_aux_get_affinity(void **mask);
3680 extern int __kmp_aux_get_affinity_max_proc();
3681 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3682 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3683 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3684 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3685 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3686 extern int kmp_set_thread_affinity_mask_initial(void);
3687 #endif
3688 static inline void __kmp_assign_root_init_mask() {
3689   int gtid = __kmp_entry_gtid();
3690   kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3691   if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3692     __kmp_affinity_set_init_mask(gtid, TRUE);
3693     r->r.r_affinity_assigned = TRUE;
3694   }
3695 }
3696 static inline void __kmp_reset_root_init_mask(int gtid) {
3697   if (!KMP_AFFINITY_CAPABLE())
3698     return;
3699   kmp_info_t *th = __kmp_threads[gtid];
3700   kmp_root_t *r = th->th.th_root;
3701   if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3702     __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3703     KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3704     r->r.r_affinity_assigned = FALSE;
3705   }
3706 }
3707 #else /* KMP_AFFINITY_SUPPORTED */
3708 #define __kmp_assign_root_init_mask() /* Nothing */
3709 static inline void __kmp_reset_root_init_mask(int gtid) {}
3710 #endif /* KMP_AFFINITY_SUPPORTED */
3711 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3712 // format string is for affinity, so platforms that do not support
3713 // affinity can still use the other fields, e.g., %n for num_threads
3714 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3715                                          kmp_str_buf_t *buffer);
3716 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3717 
3718 extern void __kmp_cleanup_hierarchy();
3719 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3720 
3721 #if KMP_USE_FUTEX
3722 
3723 extern int __kmp_futex_determine_capable(void);
3724 
3725 #endif // KMP_USE_FUTEX
3726 
3727 extern void __kmp_gtid_set_specific(int gtid);
3728 extern int __kmp_gtid_get_specific(void);
3729 
3730 extern double __kmp_read_cpu_time(void);
3731 
3732 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3733 
3734 #if KMP_USE_MONITOR
3735 extern void __kmp_create_monitor(kmp_info_t *th);
3736 #endif
3737 
3738 extern void *__kmp_launch_thread(kmp_info_t *thr);
3739 
3740 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3741 
3742 #if KMP_OS_WINDOWS
3743 extern int __kmp_still_running(kmp_info_t *th);
3744 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3745 extern void __kmp_free_handle(kmp_thread_t tHandle);
3746 #endif
3747 
3748 #if KMP_USE_MONITOR
3749 extern void __kmp_reap_monitor(kmp_info_t *th);
3750 #endif
3751 extern void __kmp_reap_worker(kmp_info_t *th);
3752 extern void __kmp_terminate_thread(int gtid);
3753 
3754 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3755 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3756 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3757 
3758 extern void __kmp_elapsed(double *);
3759 extern void __kmp_elapsed_tick(double *);
3760 
3761 extern void __kmp_enable(int old_state);
3762 extern void __kmp_disable(int *old_state);
3763 
3764 extern void __kmp_thread_sleep(int millis);
3765 
3766 extern void __kmp_common_initialize(void);
3767 extern void __kmp_common_destroy(void);
3768 extern void __kmp_common_destroy_gtid(int gtid);
3769 
3770 #if KMP_OS_UNIX
3771 extern void __kmp_register_atfork(void);
3772 #endif
3773 extern void __kmp_suspend_initialize(void);
3774 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3775 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3776 
3777 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3778                                          int tid);
3779 extern kmp_team_t *
3780 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3781 #if OMPT_SUPPORT
3782                     ompt_data_t ompt_parallel_data,
3783 #endif
3784                     kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3785                     int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3786 extern void __kmp_free_thread(kmp_info_t *);
3787 extern void __kmp_free_team(kmp_root_t *,
3788                             kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3789 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3790 
3791 /* ------------------------------------------------------------------------ */
3792 
3793 extern void __kmp_initialize_bget(kmp_info_t *th);
3794 extern void __kmp_finalize_bget(kmp_info_t *th);
3795 
3796 KMP_EXPORT void *kmpc_malloc(size_t size);
3797 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3798 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3799 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3800 KMP_EXPORT void kmpc_free(void *ptr);
3801 
3802 /* declarations for internal use */
3803 
3804 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3805                          size_t reduce_size, void *reduce_data,
3806                          void (*reduce)(void *, void *));
3807 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3808 extern int __kmp_barrier_gomp_cancel(int gtid);
3809 
3810 /*!
3811  * Tell the fork call which compiler generated the fork call, and therefore how
3812  * to deal with the call.
3813  */
3814 enum fork_context_e {
3815   fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3816                        microtask internally. */
3817   fork_context_intel, /**< Called from Intel generated code.  */
3818   fork_context_last
3819 };
3820 extern int __kmp_fork_call(ident_t *loc, int gtid,
3821                            enum fork_context_e fork_context, kmp_int32 argc,
3822                            microtask_t microtask, launch_t invoker,
3823                            kmp_va_list ap);
3824 
3825 extern void __kmp_join_call(ident_t *loc, int gtid
3826 #if OMPT_SUPPORT
3827                             ,
3828                             enum fork_context_e fork_context
3829 #endif
3830                             ,
3831                             int exit_teams = 0);
3832 
3833 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3834 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3835 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3836 extern int __kmp_invoke_task_func(int gtid);
3837 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3838                                           kmp_info_t *this_thr,
3839                                           kmp_team_t *team);
3840 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3841                                          kmp_info_t *this_thr,
3842                                          kmp_team_t *team);
3843 
3844 // should never have been exported
3845 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3846 extern int __kmp_invoke_teams_master(int gtid);
3847 extern void __kmp_teams_master(int gtid);
3848 extern int __kmp_aux_get_team_num();
3849 extern int __kmp_aux_get_num_teams();
3850 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3851 extern void __kmp_user_set_library(enum library_type arg);
3852 extern void __kmp_aux_set_library(enum library_type arg);
3853 extern void __kmp_aux_set_stacksize(size_t arg);
3854 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3855 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3856 
3857 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3858 void kmpc_set_blocktime(int arg);
3859 void ompc_set_nested(int flag);
3860 void ompc_set_dynamic(int flag);
3861 void ompc_set_num_threads(int arg);
3862 
3863 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3864                                               kmp_team_t *team, int tid);
3865 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3866 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3867                                     kmp_tasking_flags_t *flags,
3868                                     size_t sizeof_kmp_task_t,
3869                                     size_t sizeof_shareds,
3870                                     kmp_routine_entry_t task_entry);
3871 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3872                                      kmp_team_t *team, int tid,
3873                                      int set_curr_task);
3874 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3875 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3876 
3877 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3878                                                        int gtid,
3879                                                        kmp_task_t *task);
3880 extern void __kmp_fulfill_event(kmp_event_t *event);
3881 
3882 extern void __kmp_free_task_team(kmp_info_t *thread,
3883                                  kmp_task_team_t *task_team);
3884 extern void __kmp_reap_task_teams(void);
3885 extern void __kmp_wait_to_unref_task_teams(void);
3886 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3887                                   int always);
3888 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3889 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3890 #if USE_ITT_BUILD
3891                                  ,
3892                                  void *itt_sync_obj
3893 #endif /* USE_ITT_BUILD */
3894                                  ,
3895                                  int wait = 1);
3896 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3897                                   int gtid);
3898 
3899 extern int __kmp_is_address_mapped(void *addr);
3900 extern kmp_uint64 __kmp_hardware_timestamp(void);
3901 
3902 #if KMP_OS_UNIX
3903 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3904 #endif
3905 
3906 /* ------------------------------------------------------------------------ */
3907 //
3908 // Assembly routines that have no compiler intrinsic replacement
3909 //
3910 
3911 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3912                                   void *argv[]
3913 #if OMPT_SUPPORT
3914                                   ,
3915                                   void **exit_frame_ptr
3916 #endif
3917 );
3918 
3919 /* ------------------------------------------------------------------------ */
3920 
3921 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3922 KMP_EXPORT void __kmpc_end(ident_t *);
3923 
3924 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3925                                                   kmpc_ctor_vec ctor,
3926                                                   kmpc_cctor_vec cctor,
3927                                                   kmpc_dtor_vec dtor,
3928                                                   size_t vector_length);
3929 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3930                                               kmpc_ctor ctor, kmpc_cctor cctor,
3931                                               kmpc_dtor dtor);
3932 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3933                                       void *data, size_t size);
3934 
3935 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3936 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3937 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3938 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3939 
3940 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3941 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3942                                  kmpc_micro microtask, ...);
3943 KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
3944                                     kmpc_micro microtask, kmp_int32 cond,
3945                                     void *args);
3946 
3947 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3948 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3949 
3950 KMP_EXPORT void __kmpc_flush(ident_t *);
3951 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3952 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3953 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3954 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3955                                    kmp_int32 filter);
3956 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3957 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3958 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3959 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3960                                 kmp_critical_name *);
3961 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3962                                     kmp_critical_name *);
3963 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3964                                           kmp_critical_name *, uint32_t hint);
3965 
3966 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3967 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3968 
3969 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3970                                                   kmp_int32 global_tid);
3971 
3972 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3973 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3974 
3975 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
3976 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
3977                                          kmp_int32 numberOfSections);
3978 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
3979 
3980 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3981                                      kmp_int32 schedtype, kmp_int32 *plastiter,
3982                                      kmp_int *plower, kmp_int *pupper,
3983                                      kmp_int *pstride, kmp_int incr,
3984                                      kmp_int chunk);
3985 
3986 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3987 
3988 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3989                                    size_t cpy_size, void *cpy_data,
3990                                    void (*cpy_func)(void *, void *),
3991                                    kmp_int32 didit);
3992 
3993 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
3994                                           void *cpy_data);
3995 
3996 extern void KMPC_SET_NUM_THREADS(int arg);
3997 extern void KMPC_SET_DYNAMIC(int flag);
3998 extern void KMPC_SET_NESTED(int flag);
3999 
4000 /* OMP 3.0 tasking interface routines */
4001 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4002                                      kmp_task_t *new_task);
4003 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4004                                              kmp_int32 flags,
4005                                              size_t sizeof_kmp_task_t,
4006                                              size_t sizeof_shareds,
4007                                              kmp_routine_entry_t task_entry);
4008 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4009     ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4010     size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4011 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4012                                           kmp_task_t *task);
4013 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4014                                              kmp_task_t *task);
4015 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4016                                            kmp_task_t *new_task);
4017 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4018 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4019                                           int end_part);
4020 
4021 #if TASK_UNUSED
4022 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4023 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4024                               kmp_task_t *task);
4025 #endif // TASK_UNUSED
4026 
4027 /* ------------------------------------------------------------------------ */
4028 
4029 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4030 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4031 
4032 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4033     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4034     kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4035     kmp_depend_info_t *noalias_dep_list);
4036 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4037                                      kmp_int32 ndeps,
4038                                      kmp_depend_info_t *dep_list,
4039                                      kmp_int32 ndeps_noalias,
4040                                      kmp_depend_info_t *noalias_dep_list);
4041 /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4042  *                               Placeholder for taskwait with nowait clause.*/
4043 KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4044                                             kmp_int32 ndeps,
4045                                             kmp_depend_info_t *dep_list,
4046                                             kmp_int32 ndeps_noalias,
4047                                             kmp_depend_info_t *noalias_dep_list,
4048                                             kmp_int32 has_no_wait);
4049 
4050 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4051                                 bool serialize_immediate);
4052 
4053 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4054                                    kmp_int32 cncl_kind);
4055 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4056                                               kmp_int32 cncl_kind);
4057 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4058 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4059 
4060 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4061 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4062 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4063                                 kmp_int32 if_val, kmp_uint64 *lb,
4064                                 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4065                                 kmp_int32 sched, kmp_uint64 grainsize,
4066                                 void *task_dup);
4067 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4068                                   kmp_task_t *task, kmp_int32 if_val,
4069                                   kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4070                                   kmp_int32 nogroup, kmp_int32 sched,
4071                                   kmp_uint64 grainsize, kmp_int32 modifier,
4072                                   void *task_dup);
4073 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4074 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4075 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4076 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4077                                                      int is_ws, int num,
4078                                                      void *data);
4079 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4080                                               int num, void *data);
4081 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4082                                                     int is_ws);
4083 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4084     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4085     kmp_task_affinity_info_t *affin_list);
4086 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4087 KMP_EXPORT int __kmp_get_max_teams(void);
4088 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4089 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4090 
4091 /* Interface target task integration */
4092 KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4093 KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4094 
4095 /* Lock interface routines (fast versions with gtid passed in) */
4096 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4097                                  void **user_lock);
4098 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4099                                       void **user_lock);
4100 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4101                                     void **user_lock);
4102 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4103                                          void **user_lock);
4104 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4105 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4106                                      void **user_lock);
4107 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4108                                   void **user_lock);
4109 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4110                                        void **user_lock);
4111 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4112 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4113                                      void **user_lock);
4114 
4115 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4116                                            void **user_lock, uintptr_t hint);
4117 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4118                                                 void **user_lock,
4119                                                 uintptr_t hint);
4120 
4121 /* Interface to fast scalable reduce methods routines */
4122 
4123 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4124     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4125     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4126     kmp_critical_name *lck);
4127 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4128                                          kmp_critical_name *lck);
4129 KMP_EXPORT kmp_int32 __kmpc_reduce(
4130     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4131     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4132     kmp_critical_name *lck);
4133 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4134                                   kmp_critical_name *lck);
4135 
4136 /* Internal fast reduction routines */
4137 
4138 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4139     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4140     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4141     kmp_critical_name *lck);
4142 
4143 // this function is for testing set/get/determine reduce method
4144 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4145 
4146 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4147 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4148 
4149 // C++ port
4150 // missing 'extern "C"' declarations
4151 
4152 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4153 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4154 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4155                                         kmp_int32 num_threads);
4156 
4157 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4158                                       int proc_bind);
4159 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4160                                       kmp_int32 num_teams,
4161                                       kmp_int32 num_threads);
4162 /* Function for OpenMP 5.1 num_teams clause */
4163 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4164                                          kmp_int32 num_teams_lb,
4165                                          kmp_int32 num_teams_ub,
4166                                          kmp_int32 num_threads);
4167 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4168                                   kmpc_micro microtask, ...);
4169 struct kmp_dim { // loop bounds info casted to kmp_int64
4170   kmp_int64 lo; // lower
4171   kmp_int64 up; // upper
4172   kmp_int64 st; // stride
4173 };
4174 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4175                                      kmp_int32 num_dims,
4176                                      const struct kmp_dim *dims);
4177 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4178                                      const kmp_int64 *vec);
4179 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4180                                      const kmp_int64 *vec);
4181 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4182 
4183 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4184                                              void *data, size_t size,
4185                                              void ***cache);
4186 
4187 // Symbols for MS mutual detection.
4188 extern int _You_must_link_with_exactly_one_OpenMP_library;
4189 extern int _You_must_link_with_Intel_OpenMP_library;
4190 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
4191 extern int _You_must_link_with_Microsoft_OpenMP_library;
4192 #endif
4193 
4194 // The routines below are not exported.
4195 // Consider making them 'static' in corresponding source files.
4196 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4197                                            void *data_addr, size_t pc_size);
4198 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4199                                                 void *data_addr,
4200                                                 size_t pc_size);
4201 void __kmp_threadprivate_resize_cache(int newCapacity);
4202 void __kmp_cleanup_threadprivate_caches();
4203 
4204 // ompc_, kmpc_ entries moved from omp.h.
4205 #if KMP_OS_WINDOWS
4206 #define KMPC_CONVENTION __cdecl
4207 #else
4208 #define KMPC_CONVENTION
4209 #endif
4210 
4211 #ifndef __OMP_H
4212 typedef enum omp_sched_t {
4213   omp_sched_static = 1,
4214   omp_sched_dynamic = 2,
4215   omp_sched_guided = 3,
4216   omp_sched_auto = 4
4217 } omp_sched_t;
4218 typedef void *kmp_affinity_mask_t;
4219 #endif
4220 
4221 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4222 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4223 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4224 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4225 KMP_EXPORT int KMPC_CONVENTION
4226 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4227 KMP_EXPORT int KMPC_CONVENTION
4228 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4229 KMP_EXPORT int KMPC_CONVENTION
4230 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4231 
4232 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4233 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4234 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4235 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4236 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4237 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4238 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4239 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4240 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4241                                               char const *format);
4242 
4243 enum kmp_target_offload_kind {
4244   tgt_disabled = 0,
4245   tgt_default = 1,
4246   tgt_mandatory = 2
4247 };
4248 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4249 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4250 extern kmp_target_offload_kind_t __kmp_target_offload;
4251 extern int __kmpc_get_target_offload();
4252 
4253 // Constants used in libomptarget
4254 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4255 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4256 
4257 // OMP Pause Resource
4258 
4259 // The following enum is used both to set the status in __kmp_pause_status, and
4260 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4261 typedef enum kmp_pause_status_t {
4262   kmp_not_paused = 0, // status is not paused, or, requesting resume
4263   kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4264   kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4265 } kmp_pause_status_t;
4266 
4267 // This stores the pause state of the runtime
4268 extern kmp_pause_status_t __kmp_pause_status;
4269 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4270 extern int __kmp_pause_resource(kmp_pause_status_t level);
4271 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4272 extern void __kmp_resume_if_soft_paused();
4273 // Hard resume simply resets the status to not paused. Library will appear to
4274 // be uninitialized after hard pause. Let OMP constructs trigger required
4275 // initializations.
4276 static inline void __kmp_resume_if_hard_paused() {
4277   if (__kmp_pause_status == kmp_hard_paused) {
4278     __kmp_pause_status = kmp_not_paused;
4279   }
4280 }
4281 
4282 extern void __kmp_omp_display_env(int verbose);
4283 
4284 // 1: it is initializing hidden helper team
4285 extern volatile int __kmp_init_hidden_helper;
4286 // 1: the hidden helper team is done
4287 extern volatile int __kmp_hidden_helper_team_done;
4288 // 1: enable hidden helper task
4289 extern kmp_int32 __kmp_enable_hidden_helper;
4290 // Main thread of hidden helper team
4291 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4292 // Descriptors for the hidden helper threads
4293 extern kmp_info_t **__kmp_hidden_helper_threads;
4294 // Number of hidden helper threads
4295 extern kmp_int32 __kmp_hidden_helper_threads_num;
4296 // Number of hidden helper tasks that have not been executed yet
4297 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4298 
4299 extern void __kmp_hidden_helper_initialize();
4300 extern void __kmp_hidden_helper_threads_initz_routine();
4301 extern void __kmp_do_initialize_hidden_helper_threads();
4302 extern void __kmp_hidden_helper_threads_initz_wait();
4303 extern void __kmp_hidden_helper_initz_release();
4304 extern void __kmp_hidden_helper_threads_deinitz_wait();
4305 extern void __kmp_hidden_helper_threads_deinitz_release();
4306 extern void __kmp_hidden_helper_main_thread_wait();
4307 extern void __kmp_hidden_helper_worker_thread_wait();
4308 extern void __kmp_hidden_helper_worker_thread_signal();
4309 extern void __kmp_hidden_helper_main_thread_release();
4310 
4311 // Check whether a given thread is a hidden helper thread
4312 #define KMP_HIDDEN_HELPER_THREAD(gtid)                                         \
4313   ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4314 
4315 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)                                  \
4316   ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4317 
4318 #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)                                    \
4319   ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4320 
4321 #define KMP_HIDDEN_HELPER_TEAM(team)                                           \
4322   (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4323 
4324 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4325 // main thread, is skipped.
4326 #define KMP_GTID_TO_SHADOW_GTID(gtid)                                          \
4327   ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4328 
4329 // Return the adjusted gtid value by subtracting from gtid the number
4330 // of hidden helper threads. This adjusted value is the gtid the thread would
4331 // have received if there were no hidden helper threads.
4332 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4333   int adjusted_gtid = gtid;
4334   if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4335       gtid - __kmp_hidden_helper_threads_num >= 0) {
4336     adjusted_gtid -= __kmp_hidden_helper_threads_num;
4337   }
4338   return adjusted_gtid;
4339 }
4340 
4341 // Support for error directive
4342 typedef enum kmp_severity_t {
4343   severity_warning = 1,
4344   severity_fatal = 2
4345 } kmp_severity_t;
4346 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4347 
4348 // Support for scope directive
4349 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4350 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4351 
4352 #ifdef __cplusplus
4353 }
4354 #endif
4355 
4356 template <bool C, bool S>
4357 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4358 template <bool C, bool S>
4359 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4360 template <bool C, bool S>
4361 extern void __kmp_atomic_suspend_64(int th_gtid,
4362                                     kmp_atomic_flag_64<C, S> *flag);
4363 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4364 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4365 template <bool C, bool S>
4366 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4367 template <bool C, bool S>
4368 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4369 template <bool C, bool S>
4370 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4371 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4372 #endif
4373 template <bool C, bool S>
4374 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4375 template <bool C, bool S>
4376 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4377 template <bool C, bool S>
4378 extern void __kmp_atomic_resume_64(int target_gtid,
4379                                    kmp_atomic_flag_64<C, S> *flag);
4380 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4381 
4382 template <bool C, bool S>
4383 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4384                            kmp_flag_32<C, S> *flag, int final_spin,
4385                            int *thread_finished,
4386 #if USE_ITT_BUILD
4387                            void *itt_sync_obj,
4388 #endif /* USE_ITT_BUILD */
4389                            kmp_int32 is_constrained);
4390 template <bool C, bool S>
4391 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4392                            kmp_flag_64<C, S> *flag, int final_spin,
4393                            int *thread_finished,
4394 #if USE_ITT_BUILD
4395                            void *itt_sync_obj,
4396 #endif /* USE_ITT_BUILD */
4397                            kmp_int32 is_constrained);
4398 template <bool C, bool S>
4399 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4400                                   kmp_atomic_flag_64<C, S> *flag,
4401                                   int final_spin, int *thread_finished,
4402 #if USE_ITT_BUILD
4403                                   void *itt_sync_obj,
4404 #endif /* USE_ITT_BUILD */
4405                                   kmp_int32 is_constrained);
4406 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4407                                kmp_flag_oncore *flag, int final_spin,
4408                                int *thread_finished,
4409 #if USE_ITT_BUILD
4410                                void *itt_sync_obj,
4411 #endif /* USE_ITT_BUILD */
4412                                kmp_int32 is_constrained);
4413 
4414 extern int __kmp_nesting_mode;
4415 extern int __kmp_nesting_mode_nlevels;
4416 extern int *__kmp_nesting_nth_level;
4417 extern void __kmp_init_nesting_mode();
4418 extern void __kmp_set_nesting_mode_threads();
4419 
4420 /// This class safely opens and closes a C-style FILE* object using RAII
4421 /// semantics. There are also methods which allow using stdout or stderr as
4422 /// the underlying FILE* object. With the implicit conversion operator to
4423 /// FILE*, an object with this type can be used in any function which takes
4424 /// a FILE* object e.g., fprintf().
4425 /// No close method is needed at use sites.
4426 class kmp_safe_raii_file_t {
4427   FILE *f;
4428 
4429   void close() {
4430     if (f && f != stdout && f != stderr) {
4431       fclose(f);
4432       f = nullptr;
4433     }
4434   }
4435 
4436 public:
4437   kmp_safe_raii_file_t() : f(nullptr) {}
4438   kmp_safe_raii_file_t(const char *filename, const char *mode,
4439                        const char *env_var = nullptr)
4440       : f(nullptr) {
4441     open(filename, mode, env_var);
4442   }
4443   ~kmp_safe_raii_file_t() { close(); }
4444 
4445   /// Open filename using mode. This is automatically closed in the destructor.
4446   /// The env_var parameter indicates the environment variable the filename
4447   /// came from if != nullptr.
4448   void open(const char *filename, const char *mode,
4449             const char *env_var = nullptr) {
4450     KMP_ASSERT(!f);
4451     f = fopen(filename, mode);
4452     if (!f) {
4453       int code = errno;
4454       if (env_var) {
4455         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4456                     KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4457       } else {
4458         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4459                     __kmp_msg_null);
4460       }
4461     }
4462   }
4463   /// Instead of erroring out, return non-zero when
4464   /// unsuccessful fopen() for any reason
4465   int try_open(const char *filename, const char *mode) {
4466     KMP_ASSERT(!f);
4467     f = fopen(filename, mode);
4468     if (!f)
4469       return errno;
4470     return 0;
4471   }
4472   /// Set the FILE* object to stdout and output there
4473   /// No open call should happen before this call.
4474   void set_stdout() {
4475     KMP_ASSERT(!f);
4476     f = stdout;
4477   }
4478   /// Set the FILE* object to stderr and output there
4479   /// No open call should happen before this call.
4480   void set_stderr() {
4481     KMP_ASSERT(!f);
4482     f = stderr;
4483   }
4484   operator bool() { return bool(f); }
4485   operator FILE *() { return f; }
4486 };
4487 
4488 template <typename SourceType, typename TargetType,
4489           bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4490           bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4491           bool isSourceSigned = std::is_signed<SourceType>::value,
4492           bool isTargetSigned = std::is_signed<TargetType>::value>
4493 struct kmp_convert {};
4494 
4495 // Both types are signed; Source smaller
4496 template <typename SourceType, typename TargetType>
4497 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4498   static TargetType to(SourceType src) { return (TargetType)src; }
4499 };
4500 // Source equal
4501 template <typename SourceType, typename TargetType>
4502 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4503   static TargetType to(SourceType src) { return src; }
4504 };
4505 // Source bigger
4506 template <typename SourceType, typename TargetType>
4507 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4508   static TargetType to(SourceType src) {
4509     KMP_ASSERT(src <= static_cast<SourceType>(
4510                           (std::numeric_limits<TargetType>::max)()));
4511     KMP_ASSERT(src >= static_cast<SourceType>(
4512                           (std::numeric_limits<TargetType>::min)()));
4513     return (TargetType)src;
4514   }
4515 };
4516 
4517 // Source signed, Target unsigned
4518 // Source smaller
4519 template <typename SourceType, typename TargetType>
4520 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4521   static TargetType to(SourceType src) {
4522     KMP_ASSERT(src >= 0);
4523     return (TargetType)src;
4524   }
4525 };
4526 // Source equal
4527 template <typename SourceType, typename TargetType>
4528 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4529   static TargetType to(SourceType src) {
4530     KMP_ASSERT(src >= 0);
4531     return (TargetType)src;
4532   }
4533 };
4534 // Source bigger
4535 template <typename SourceType, typename TargetType>
4536 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4537   static TargetType to(SourceType src) {
4538     KMP_ASSERT(src >= 0);
4539     KMP_ASSERT(src <= static_cast<SourceType>(
4540                           (std::numeric_limits<TargetType>::max)()));
4541     return (TargetType)src;
4542   }
4543 };
4544 
4545 // Source unsigned, Target signed
4546 // Source smaller
4547 template <typename SourceType, typename TargetType>
4548 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4549   static TargetType to(SourceType src) { return (TargetType)src; }
4550 };
4551 // Source equal
4552 template <typename SourceType, typename TargetType>
4553 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4554   static TargetType to(SourceType src) {
4555     KMP_ASSERT(src <= static_cast<SourceType>(
4556                           (std::numeric_limits<TargetType>::max)()));
4557     return (TargetType)src;
4558   }
4559 };
4560 // Source bigger
4561 template <typename SourceType, typename TargetType>
4562 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4563   static TargetType to(SourceType src) {
4564     KMP_ASSERT(src <= static_cast<SourceType>(
4565                           (std::numeric_limits<TargetType>::max)()));
4566     return (TargetType)src;
4567   }
4568 };
4569 
4570 // Source unsigned, Target unsigned
4571 // Source smaller
4572 template <typename SourceType, typename TargetType>
4573 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4574   static TargetType to(SourceType src) { return (TargetType)src; }
4575 };
4576 // Source equal
4577 template <typename SourceType, typename TargetType>
4578 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4579   static TargetType to(SourceType src) { return src; }
4580 };
4581 // Source bigger
4582 template <typename SourceType, typename TargetType>
4583 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4584   static TargetType to(SourceType src) {
4585     KMP_ASSERT(src <= static_cast<SourceType>(
4586                           (std::numeric_limits<TargetType>::max)()));
4587     return (TargetType)src;
4588   }
4589 };
4590 
4591 template <typename T1, typename T2>
4592 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4593   *dest = kmp_convert<T1, T2>::to(src);
4594 }
4595 
4596 #endif /* KMP_H */
4597