xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp.h (revision 5036d9652a5701d00e9e40ea942c278e9f77d33d)
1 /*! \file */
2 /*
3  * kmp.h -- KPTS runtime header file.
4  */
5 
6 //===----------------------------------------------------------------------===//
7 //
8 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9 // See https://llvm.org/LICENSE.txt for license information.
10 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef KMP_H
15 #define KMP_H
16 
17 #include "kmp_config.h"
18 
19 /* #define BUILD_PARALLEL_ORDERED 1 */
20 
21 /* This fix replaces gettimeofday with clock_gettime for better scalability on
22    the Altix.  Requires user code to be linked with -lrt. */
23 //#define FIX_SGI_CLOCK
24 
25 /* Defines for OpenMP 3.0 tasking and auto scheduling */
26 
27 #ifndef KMP_STATIC_STEAL_ENABLED
28 #define KMP_STATIC_STEAL_ENABLED 1
29 #endif
30 #define KMP_WEIGHTED_ITERATIONS_SUPPORTED                                      \
31   (KMP_AFFINITY_SUPPORTED && KMP_STATIC_STEAL_ENABLED &&                       \
32    (KMP_ARCH_X86 || KMP_ARCH_X86_64))
33 
34 #define TASK_CURRENT_NOT_QUEUED 0
35 #define TASK_CURRENT_QUEUED 1
36 
37 #ifdef BUILD_TIED_TASK_STACK
38 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
39 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
40 // Number of entries in each task stack array
41 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
42 // Mask for determining index into stack block
43 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
44 #endif // BUILD_TIED_TASK_STACK
45 
46 #define TASK_NOT_PUSHED 1
47 #define TASK_SUCCESSFULLY_PUSHED 0
48 #define TASK_TIED 1
49 #define TASK_UNTIED 0
50 #define TASK_EXPLICIT 1
51 #define TASK_IMPLICIT 0
52 #define TASK_PROXY 1
53 #define TASK_FULL 0
54 #define TASK_DETACHABLE 1
55 #define TASK_UNDETACHABLE 0
56 
57 #define KMP_CANCEL_THREADS
58 #define KMP_THREAD_ATTR
59 
60 // Android does not have pthread_cancel.  Undefine KMP_CANCEL_THREADS if being
61 // built on Android
62 #if defined(__ANDROID__)
63 #undef KMP_CANCEL_THREADS
64 #endif
65 
66 // Some WASI targets (e.g., wasm32-wasi-threads) do not support thread
67 // cancellation.
68 #if KMP_OS_WASI
69 #undef KMP_CANCEL_THREADS
70 #endif
71 
72 #if !KMP_OS_WASI
73 #include <signal.h>
74 #endif
75 #include <stdarg.h>
76 #include <stddef.h>
77 #include <stdio.h>
78 #include <stdlib.h>
79 #include <string.h>
80 #include <limits>
81 #include <type_traits>
82 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
83    Microsoft library. Some macros provided below to replace these functions  */
84 #ifndef __ABSOFT_WIN
85 #include <sys/types.h>
86 #endif
87 #include <limits.h>
88 #include <time.h>
89 
90 #include <errno.h>
91 
92 #include "kmp_os.h"
93 
94 #include "kmp_safe_c_api.h"
95 
96 #if KMP_STATS_ENABLED
97 class kmp_stats_list;
98 #endif
99 
100 #if KMP_USE_HIER_SCHED
101 // Only include hierarchical scheduling if affinity is supported
102 #undef KMP_USE_HIER_SCHED
103 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
104 #endif
105 
106 // OMPD_SKIP_HWLOC used in libompd/omp-icv.cpp to avoid OMPD depending on hwloc
107 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED && !defined(OMPD_SKIP_HWLOC)
108 #include "hwloc.h"
109 #ifndef HWLOC_OBJ_NUMANODE
110 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
111 #endif
112 #ifndef HWLOC_OBJ_PACKAGE
113 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
114 #endif
115 #endif
116 
117 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
118 #include <xmmintrin.h>
119 #endif
120 
121 // The below has to be defined before including "kmp_barrier.h".
122 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
123 #define KMP_INTERNAL_FREE(p) free(p)
124 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
125 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
126 
127 #include "kmp_debug.h"
128 #include "kmp_lock.h"
129 #include "kmp_version.h"
130 #include "kmp_barrier.h"
131 #if USE_DEBUGGER
132 #include "kmp_debugger.h"
133 #endif
134 #include "kmp_i18n.h"
135 
136 #define KMP_HANDLE_SIGNALS ((KMP_OS_UNIX && !KMP_OS_WASI) || KMP_OS_WINDOWS)
137 
138 #include "kmp_wrapper_malloc.h"
139 #if KMP_OS_UNIX
140 #include <unistd.h>
141 #if !defined NSIG && defined _NSIG
142 #define NSIG _NSIG
143 #endif
144 #endif
145 
146 #if KMP_OS_LINUX
147 #pragma weak clock_gettime
148 #endif
149 
150 #if OMPT_SUPPORT
151 #include "ompt-internal.h"
152 #endif
153 
154 #if OMPD_SUPPORT
155 #include "ompd-specific.h"
156 #endif
157 
158 #ifndef UNLIKELY
159 #define UNLIKELY(x) (x)
160 #endif
161 
162 // Affinity format function
163 #include "kmp_str.h"
164 
165 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
166 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
167 // free lists of limited size.
168 #ifndef USE_FAST_MEMORY
169 #define USE_FAST_MEMORY 3
170 #endif
171 
172 #ifndef KMP_NESTED_HOT_TEAMS
173 #define KMP_NESTED_HOT_TEAMS 0
174 #define USE_NESTED_HOT_ARG(x)
175 #else
176 #if KMP_NESTED_HOT_TEAMS
177 #define USE_NESTED_HOT_ARG(x) , x
178 #else
179 #define USE_NESTED_HOT_ARG(x)
180 #endif
181 #endif
182 
183 // Assume using BGET compare_exchange instruction instead of lock by default.
184 #ifndef USE_CMP_XCHG_FOR_BGET
185 #define USE_CMP_XCHG_FOR_BGET 1
186 #endif
187 
188 // Test to see if queuing lock is better than bootstrap lock for bget
189 // #ifndef USE_QUEUING_LOCK_FOR_BGET
190 // #define USE_QUEUING_LOCK_FOR_BGET
191 // #endif
192 
193 #define KMP_NSEC_PER_SEC 1000000000L
194 #define KMP_USEC_PER_SEC 1000000L
195 #define KMP_NSEC_PER_USEC 1000L
196 
197 /*!
198 @ingroup BASIC_TYPES
199 @{
200 */
201 
202 /*!
203 Values for bit flags used in the ident_t to describe the fields.
204 */
205 enum {
206   /*! Use trampoline for internal microtasks */
207   KMP_IDENT_IMB = 0x01,
208   /*! Use c-style ident structure */
209   KMP_IDENT_KMPC = 0x02,
210   /* 0x04 is no longer used */
211   /*! Entry point generated by auto-parallelization */
212   KMP_IDENT_AUTOPAR = 0x08,
213   /*! Compiler generates atomic reduction option for kmpc_reduce* */
214   KMP_IDENT_ATOMIC_REDUCE = 0x10,
215   /*! To mark a 'barrier' directive in user code */
216   KMP_IDENT_BARRIER_EXPL = 0x20,
217   /*! To Mark implicit barriers. */
218   KMP_IDENT_BARRIER_IMPL = 0x0040,
219   KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
220   KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
221   KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
222 
223   KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
224   KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
225 
226   /*! To mark a static loop in OMPT callbacks */
227   KMP_IDENT_WORK_LOOP = 0x200,
228   /*! To mark a sections directive in OMPT callbacks */
229   KMP_IDENT_WORK_SECTIONS = 0x400,
230   /*! To mark a distribute construct in OMPT callbacks */
231   KMP_IDENT_WORK_DISTRIBUTE = 0x800,
232   /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
233       not currently used. If one day we need more bits, then we can use
234       an invalid combination of hints to mean that another, larger field
235       should be used in a different flag. */
236   KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
237   KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
238   KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
239   KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
240   KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
241   KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
242 };
243 
244 /*!
245  * The ident structure that describes a source location.
246  */
247 typedef struct ident {
248   kmp_int32 reserved_1; /**<  might be used in Fortran; see above  */
249   kmp_int32 flags; /**<  also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
250                       identifies this union member  */
251   kmp_int32 reserved_2; /**<  not really used in Fortran any more; see above */
252 #if USE_ITT_BUILD
253 /*  but currently used for storing region-specific ITT */
254 /*  contextual information. */
255 #endif /* USE_ITT_BUILD */
256   kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++  */
257   char const *psource; /**< String describing the source location.
258                        The string is composed of semi-colon separated fields
259                        which describe the source file, the function and a pair
260                        of line numbers that delimit the construct. */
261   // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
262   kmp_int32 get_openmp_version() {
263     return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
264   }
265 } ident_t;
266 /*!
267 @}
268 */
269 
270 // Some forward declarations.
271 typedef union kmp_team kmp_team_t;
272 typedef struct kmp_taskdata kmp_taskdata_t;
273 typedef union kmp_task_team kmp_task_team_t;
274 typedef union kmp_team kmp_team_p;
275 typedef union kmp_info kmp_info_p;
276 typedef union kmp_root kmp_root_p;
277 
278 template <bool C = false, bool S = true> class kmp_flag_32;
279 template <bool C = false, bool S = true> class kmp_flag_64;
280 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
281 class kmp_flag_oncore;
282 
283 #ifdef __cplusplus
284 extern "C" {
285 #endif
286 
287 /* ------------------------------------------------------------------------ */
288 
289 /* Pack two 32-bit signed integers into a 64-bit signed integer */
290 /* ToDo: Fix word ordering for big-endian machines. */
291 #define KMP_PACK_64(HIGH_32, LOW_32)                                           \
292   ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
293 
294 // Generic string manipulation macros. Assume that _x is of type char *
295 #define SKIP_WS(_x)                                                            \
296   {                                                                            \
297     while (*(_x) == ' ' || *(_x) == '\t')                                      \
298       (_x)++;                                                                  \
299   }
300 #define SKIP_DIGITS(_x)                                                        \
301   {                                                                            \
302     while (*(_x) >= '0' && *(_x) <= '9')                                       \
303       (_x)++;                                                                  \
304   }
305 #define SKIP_TOKEN(_x)                                                         \
306   {                                                                            \
307     while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
308            (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_')                     \
309       (_x)++;                                                                  \
310   }
311 #define SKIP_TO(_x, _c)                                                        \
312   {                                                                            \
313     while (*(_x) != '\0' && *(_x) != (_c))                                     \
314       (_x)++;                                                                  \
315   }
316 
317 /* ------------------------------------------------------------------------ */
318 
319 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
320 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
321 
322 /* ------------------------------------------------------------------------ */
323 /* Enumeration types */
324 
325 enum kmp_state_timer {
326   ts_stop,
327   ts_start,
328   ts_pause,
329 
330   ts_last_state
331 };
332 
333 enum dynamic_mode {
334   dynamic_default,
335 #ifdef USE_LOAD_BALANCE
336   dynamic_load_balance,
337 #endif /* USE_LOAD_BALANCE */
338   dynamic_random,
339   dynamic_thread_limit,
340   dynamic_max
341 };
342 
343 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
344  * not include it here */
345 #ifndef KMP_SCHED_TYPE_DEFINED
346 #define KMP_SCHED_TYPE_DEFINED
347 typedef enum kmp_sched {
348   kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
349   // Note: need to adjust __kmp_sch_map global array in case enum is changed
350   kmp_sched_static = 1, // mapped to kmp_sch_static_chunked           (33)
351   kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked          (35)
352   kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked           (36)
353   kmp_sched_auto = 4, // mapped to kmp_sch_auto                     (38)
354   kmp_sched_upper_std = 5, // upper bound for standard schedules
355   kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
356   kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
357 #if KMP_STATIC_STEAL_ENABLED
358   kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
359 #endif
360   kmp_sched_upper,
361   kmp_sched_default = kmp_sched_static, // default scheduling
362   kmp_sched_monotonic = 0x80000000
363 } kmp_sched_t;
364 #endif
365 
366 /*!
367  @ingroup WORK_SHARING
368  * Describes the loop schedule to be used for a parallel for loop.
369  */
370 enum sched_type : kmp_int32 {
371   kmp_sch_lower = 32, /**< lower bound for unordered values */
372   kmp_sch_static_chunked = 33,
373   kmp_sch_static = 34, /**< static unspecialized */
374   kmp_sch_dynamic_chunked = 35,
375   kmp_sch_guided_chunked = 36, /**< guided unspecialized */
376   kmp_sch_runtime = 37,
377   kmp_sch_auto = 38, /**< auto */
378   kmp_sch_trapezoidal = 39,
379 
380   /* accessible only through KMP_SCHEDULE environment variable */
381   kmp_sch_static_greedy = 40,
382   kmp_sch_static_balanced = 41,
383   /* accessible only through KMP_SCHEDULE environment variable */
384   kmp_sch_guided_iterative_chunked = 42,
385   kmp_sch_guided_analytical_chunked = 43,
386   /* accessible only through KMP_SCHEDULE environment variable */
387   kmp_sch_static_steal = 44,
388 
389   /* static with chunk adjustment (e.g., simd) */
390   kmp_sch_static_balanced_chunked = 45,
391   kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
392   kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
393 
394   /* accessible only through KMP_SCHEDULE environment variable */
395   kmp_sch_upper, /**< upper bound for unordered values */
396 
397   kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
398   kmp_ord_static_chunked = 65,
399   kmp_ord_static = 66, /**< ordered static unspecialized */
400   kmp_ord_dynamic_chunked = 67,
401   kmp_ord_guided_chunked = 68,
402   kmp_ord_runtime = 69,
403   kmp_ord_auto = 70, /**< ordered auto */
404   kmp_ord_trapezoidal = 71,
405   kmp_ord_upper, /**< upper bound for ordered values */
406 
407   /* Schedules for Distribute construct */
408   kmp_distribute_static_chunked = 91, /**< distribute static chunked */
409   kmp_distribute_static = 92, /**< distribute static unspecialized */
410 
411   /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
412      single iteration/chunk, even if the loop is serialized. For the schedule
413      types listed above, the entire iteration vector is returned if the loop is
414      serialized. This doesn't work for gcc/gcomp sections. */
415   kmp_nm_lower = 160, /**< lower bound for nomerge values */
416 
417   kmp_nm_static_chunked =
418       (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
419   kmp_nm_static = 162, /**< static unspecialized */
420   kmp_nm_dynamic_chunked = 163,
421   kmp_nm_guided_chunked = 164, /**< guided unspecialized */
422   kmp_nm_runtime = 165,
423   kmp_nm_auto = 166, /**< auto */
424   kmp_nm_trapezoidal = 167,
425 
426   /* accessible only through KMP_SCHEDULE environment variable */
427   kmp_nm_static_greedy = 168,
428   kmp_nm_static_balanced = 169,
429   /* accessible only through KMP_SCHEDULE environment variable */
430   kmp_nm_guided_iterative_chunked = 170,
431   kmp_nm_guided_analytical_chunked = 171,
432   kmp_nm_static_steal =
433       172, /* accessible only through OMP_SCHEDULE environment variable */
434 
435   kmp_nm_ord_static_chunked = 193,
436   kmp_nm_ord_static = 194, /**< ordered static unspecialized */
437   kmp_nm_ord_dynamic_chunked = 195,
438   kmp_nm_ord_guided_chunked = 196,
439   kmp_nm_ord_runtime = 197,
440   kmp_nm_ord_auto = 198, /**< auto */
441   kmp_nm_ord_trapezoidal = 199,
442   kmp_nm_upper, /**< upper bound for nomerge values */
443 
444   /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
445      we need to distinguish the three possible cases (no modifier, monotonic
446      modifier, nonmonotonic modifier), we need separate bits for each modifier.
447      The absence of monotonic does not imply nonmonotonic, especially since 4.5
448      says that the behaviour of the "no modifier" case is implementation defined
449      in 4.5, but will become "nonmonotonic" in 5.0.
450 
451      Since we're passing a full 32 bit value, we can use a couple of high bits
452      for these flags; out of paranoia we avoid the sign bit.
453 
454      These modifiers can be or-ed into non-static schedules by the compiler to
455      pass the additional information. They will be stripped early in the
456      processing in __kmp_dispatch_init when setting up schedules, so most of the
457      code won't ever see schedules with these bits set.  */
458   kmp_sch_modifier_monotonic =
459       (1 << 29), /**< Set if the monotonic schedule modifier was present */
460   kmp_sch_modifier_nonmonotonic =
461       (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
462 
463 #define SCHEDULE_WITHOUT_MODIFIERS(s)                                          \
464   (enum sched_type)(                                                           \
465       (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
466 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
467 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
468 #define SCHEDULE_HAS_NO_MODIFIERS(s)                                           \
469   (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
470 #define SCHEDULE_GET_MODIFIERS(s)                                              \
471   ((enum sched_type)(                                                          \
472       (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
473 #define SCHEDULE_SET_MODIFIERS(s, m)                                           \
474   (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
475 #define SCHEDULE_NONMONOTONIC 0
476 #define SCHEDULE_MONOTONIC 1
477 
478   kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
479 };
480 
481 // Apply modifiers on internal kind to standard kind
482 static inline void
483 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
484                                enum sched_type internal_kind) {
485   if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
486     *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
487   }
488 }
489 
490 // Apply modifiers on standard kind to internal kind
491 static inline void
492 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
493                                enum sched_type *internal_kind) {
494   if ((int)kind & (int)kmp_sched_monotonic) {
495     *internal_kind = (enum sched_type)((int)*internal_kind |
496                                        (int)kmp_sch_modifier_monotonic);
497   }
498 }
499 
500 // Get standard schedule without modifiers
501 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
502   return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
503 }
504 
505 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
506 typedef union kmp_r_sched {
507   struct {
508     enum sched_type r_sched_type;
509     int chunk;
510   };
511   kmp_int64 sched;
512 } kmp_r_sched_t;
513 
514 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
515 // internal schedule types
516 
517 enum library_type {
518   library_none,
519   library_serial,
520   library_turnaround,
521   library_throughput
522 };
523 
524 #if KMP_OS_LINUX
525 enum clock_function_type {
526   clock_function_gettimeofday,
527   clock_function_clock_gettime
528 };
529 #endif /* KMP_OS_LINUX */
530 
531 #if KMP_MIC_SUPPORTED
532 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
533 #endif
534 
535 // OpenMP 3.1 - Nested num threads array
536 typedef struct kmp_nested_nthreads_t {
537   int *nth;
538   int size;
539   int used;
540 } kmp_nested_nthreads_t;
541 
542 extern kmp_nested_nthreads_t __kmp_nested_nth;
543 
544 /* -- fast reduction stuff ------------------------------------------------ */
545 
546 #undef KMP_FAST_REDUCTION_BARRIER
547 #define KMP_FAST_REDUCTION_BARRIER 1
548 
549 #undef KMP_FAST_REDUCTION_CORE_DUO
550 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
551 #define KMP_FAST_REDUCTION_CORE_DUO 1
552 #endif
553 
554 enum _reduction_method {
555   reduction_method_not_defined = 0,
556   critical_reduce_block = (1 << 8),
557   atomic_reduce_block = (2 << 8),
558   tree_reduce_block = (3 << 8),
559   empty_reduce_block = (4 << 8)
560 };
561 
562 // Description of the packed_reduction_method variable:
563 // The packed_reduction_method variable consists of two enum types variables
564 // that are packed together into 0-th byte and 1-st byte:
565 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
566 // barrier that will be used in fast reduction: bs_plain_barrier or
567 // bs_reduction_barrier
568 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
569 // be used in fast reduction;
570 // Reduction method is of 'enum _reduction_method' type and it's defined the way
571 // so that the bits of 0-th byte are empty, so no need to execute a shift
572 // instruction while packing/unpacking
573 
574 #if KMP_FAST_REDUCTION_BARRIER
575 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
576   ((reduction_method) | (barrier_type))
577 
578 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
579   ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
580 
581 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method)                      \
582   ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
583 #else
584 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
585   (reduction_method)
586 
587 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
588   (packed_reduction_method)
589 
590 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
591 #endif
592 
593 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block)  \
594   ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) ==                       \
595    (which_reduction_block))
596 
597 #if KMP_FAST_REDUCTION_BARRIER
598 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER                               \
599   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
600 
601 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER                                   \
602   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
603 #endif
604 
605 typedef int PACKED_REDUCTION_METHOD_T;
606 
607 /* -- end of fast reduction stuff ----------------------------------------- */
608 
609 #if KMP_OS_WINDOWS
610 #define USE_CBLKDATA
611 #if KMP_MSVC_COMPAT
612 #pragma warning(push)
613 #pragma warning(disable : 271 310)
614 #endif
615 #include <windows.h>
616 #if KMP_MSVC_COMPAT
617 #pragma warning(pop)
618 #endif
619 #endif
620 
621 #if KMP_OS_UNIX
622 #if !KMP_OS_WASI
623 #include <dlfcn.h>
624 #endif
625 #include <pthread.h>
626 #endif
627 
628 enum kmp_hw_t : int {
629   KMP_HW_UNKNOWN = -1,
630   KMP_HW_SOCKET = 0,
631   KMP_HW_PROC_GROUP,
632   KMP_HW_NUMA,
633   KMP_HW_DIE,
634   KMP_HW_LLC,
635   KMP_HW_L3,
636   KMP_HW_TILE,
637   KMP_HW_MODULE,
638   KMP_HW_L2,
639   KMP_HW_L1,
640   KMP_HW_CORE,
641   KMP_HW_THREAD,
642   KMP_HW_LAST
643 };
644 
645 typedef enum kmp_hw_core_type_t {
646   KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
647 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
648   KMP_HW_CORE_TYPE_ATOM = 0x20,
649   KMP_HW_CORE_TYPE_CORE = 0x40,
650   KMP_HW_MAX_NUM_CORE_TYPES = 3,
651 #else
652   KMP_HW_MAX_NUM_CORE_TYPES = 1,
653 #endif
654 } kmp_hw_core_type_t;
655 
656 #define KMP_HW_MAX_NUM_CORE_EFFS 8
657 
658 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type)                                   \
659   KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
660 #define KMP_ASSERT_VALID_HW_TYPE(type)                                         \
661   KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
662 
663 #define KMP_FOREACH_HW_TYPE(type)                                              \
664   for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST;                        \
665        type = (kmp_hw_t)((int)type + 1))
666 
667 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
668 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
669 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
670 
671 /* Only Linux* OS and Windows* OS support thread affinity. */
672 #if KMP_AFFINITY_SUPPORTED
673 
674 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
675 #if KMP_OS_WINDOWS
676 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
677 typedef struct GROUP_AFFINITY {
678   KAFFINITY Mask;
679   WORD Group;
680   WORD Reserved[3];
681 } GROUP_AFFINITY;
682 #endif /* _MSC_VER < 1600 */
683 #if KMP_GROUP_AFFINITY
684 extern int __kmp_num_proc_groups;
685 #else
686 static const int __kmp_num_proc_groups = 1;
687 #endif /* KMP_GROUP_AFFINITY */
688 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
689 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
690 
691 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
692 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
693 
694 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
695 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
696 
697 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
698                                              GROUP_AFFINITY *);
699 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
700 #endif /* KMP_OS_WINDOWS */
701 
702 #if KMP_USE_HWLOC && !defined(OMPD_SKIP_HWLOC)
703 extern hwloc_topology_t __kmp_hwloc_topology;
704 extern int __kmp_hwloc_error;
705 #endif
706 
707 extern size_t __kmp_affin_mask_size;
708 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
709 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
710 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
711 #define KMP_CPU_SET_ITERATE(i, mask)                                           \
712   for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
713 #define KMP_CPU_SET(i, mask) (mask)->set(i)
714 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
715 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
716 #define KMP_CPU_ZERO(mask) (mask)->zero()
717 #define KMP_CPU_ISEMPTY(mask) (mask)->empty()
718 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
719 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
720 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
721 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
722 #define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
723 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
724 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
725 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
726 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
727 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
728 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
729 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
730 #define KMP_CPU_ALLOC_ARRAY(arr, n)                                            \
731   (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
732 #define KMP_CPU_FREE_ARRAY(arr, n)                                             \
733   __kmp_affinity_dispatch->deallocate_mask_array(arr)
734 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
735 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
736 #define __kmp_get_system_affinity(mask, abort_bool)                            \
737   (mask)->get_system_affinity(abort_bool)
738 #define __kmp_set_system_affinity(mask, abort_bool)                            \
739   (mask)->set_system_affinity(abort_bool)
740 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
741 
742 class KMPAffinity {
743 public:
744   class Mask {
745   public:
746     void *operator new(size_t n);
747     void operator delete(void *p);
748     void *operator new[](size_t n);
749     void operator delete[](void *p);
750     virtual ~Mask() {}
751     // Set bit i to 1
752     virtual void set(int i) {}
753     // Return bit i
754     virtual bool is_set(int i) const { return false; }
755     // Set bit i to 0
756     virtual void clear(int i) {}
757     // Zero out entire mask
758     virtual void zero() {}
759     // Check whether mask is empty
760     virtual bool empty() const { return true; }
761     // Copy src into this mask
762     virtual void copy(const Mask *src) {}
763     // this &= rhs
764     virtual void bitwise_and(const Mask *rhs) {}
765     // this |= rhs
766     virtual void bitwise_or(const Mask *rhs) {}
767     // this = ~this
768     virtual void bitwise_not() {}
769     // this == rhs
770     virtual bool is_equal(const Mask *rhs) const { return false; }
771     // API for iterating over an affinity mask
772     // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
773     virtual int begin() const { return 0; }
774     virtual int end() const { return 0; }
775     virtual int next(int previous) const { return 0; }
776 #if KMP_OS_WINDOWS
777     virtual int set_process_affinity(bool abort_on_error) const { return -1; }
778 #endif
779     // Set the system's affinity to this affinity mask's value
780     virtual int set_system_affinity(bool abort_on_error) const { return -1; }
781     // Set this affinity mask to the current system affinity
782     virtual int get_system_affinity(bool abort_on_error) { return -1; }
783     // Only 1 DWORD in the mask should have any procs set.
784     // Return the appropriate index, or -1 for an invalid mask.
785     virtual int get_proc_group() const { return -1; }
786     int get_max_cpu() const {
787       int cpu;
788       int max_cpu = -1;
789       KMP_CPU_SET_ITERATE(cpu, this) {
790         if (cpu > max_cpu)
791           max_cpu = cpu;
792       }
793       return max_cpu;
794     }
795   };
796   void *operator new(size_t n);
797   void operator delete(void *p);
798   // Need virtual destructor
799   virtual ~KMPAffinity() = default;
800   // Determine if affinity is capable
801   virtual void determine_capable(const char *env_var) {}
802   // Bind the current thread to os proc
803   virtual void bind_thread(int proc) {}
804   // Factory functions to allocate/deallocate a mask
805   virtual Mask *allocate_mask() { return nullptr; }
806   virtual void deallocate_mask(Mask *m) {}
807   virtual Mask *allocate_mask_array(int num) { return nullptr; }
808   virtual void deallocate_mask_array(Mask *m) {}
809   virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
810   static void pick_api();
811   static void destroy_api();
812   enum api_type {
813     NATIVE_OS
814 #if KMP_USE_HWLOC
815     ,
816     HWLOC
817 #endif
818   };
819   virtual api_type get_api_type() const {
820     KMP_ASSERT(0);
821     return NATIVE_OS;
822   }
823 
824 private:
825   static bool picked_api;
826 };
827 
828 typedef KMPAffinity::Mask kmp_affin_mask_t;
829 extern KMPAffinity *__kmp_affinity_dispatch;
830 
831 #ifndef KMP_OS_AIX
832 class kmp_affinity_raii_t {
833   kmp_affin_mask_t *mask;
834   bool restored;
835 
836 public:
837   kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
838       : mask(nullptr), restored(false) {
839     if (KMP_AFFINITY_CAPABLE()) {
840       KMP_CPU_ALLOC(mask);
841       KMP_ASSERT(mask != NULL);
842       __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
843       if (new_mask)
844         __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
845     }
846   }
847   void restore() {
848     if (mask && KMP_AFFINITY_CAPABLE() && !restored) {
849       __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
850       KMP_CPU_FREE(mask);
851     }
852     restored = true;
853   }
854   ~kmp_affinity_raii_t() { restore(); }
855 };
856 #endif // !KMP_OS_AIX
857 
858 // Declare local char buffers with this size for printing debug and info
859 // messages, using __kmp_affinity_print_mask().
860 #define KMP_AFFIN_MASK_PRINT_LEN 1024
861 
862 enum affinity_type {
863   affinity_none = 0,
864   affinity_physical,
865   affinity_logical,
866   affinity_compact,
867   affinity_scatter,
868   affinity_explicit,
869   affinity_balanced,
870   affinity_disabled, // not used outsize the env var parser
871   affinity_default
872 };
873 
874 enum affinity_top_method {
875   affinity_top_method_all = 0, // try all (supported) methods, in order
876 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
877   affinity_top_method_apicid,
878   affinity_top_method_x2apicid,
879   affinity_top_method_x2apicid_1f,
880 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
881   affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
882 #if KMP_GROUP_AFFINITY
883   affinity_top_method_group,
884 #endif /* KMP_GROUP_AFFINITY */
885   affinity_top_method_flat,
886 #if KMP_USE_HWLOC
887   affinity_top_method_hwloc,
888 #endif
889   affinity_top_method_default
890 };
891 
892 #define affinity_respect_mask_default (2)
893 
894 typedef struct kmp_affinity_flags_t {
895   unsigned dups : 1;
896   unsigned verbose : 1;
897   unsigned warnings : 1;
898   unsigned respect : 2;
899   unsigned reset : 1;
900   unsigned initialized : 1;
901   unsigned core_types_gran : 1;
902   unsigned core_effs_gran : 1;
903   unsigned omp_places : 1;
904   unsigned reserved : 22;
905 } kmp_affinity_flags_t;
906 KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
907 
908 typedef struct kmp_affinity_ids_t {
909   int os_id;
910   int ids[KMP_HW_LAST];
911 } kmp_affinity_ids_t;
912 
913 typedef struct kmp_affinity_attrs_t {
914   int core_type : 8;
915   int core_eff : 8;
916   unsigned valid : 1;
917   unsigned reserved : 15;
918 } kmp_affinity_attrs_t;
919 #define KMP_AFFINITY_ATTRS_UNKNOWN                                             \
920   { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
921 
922 typedef struct kmp_affinity_t {
923   char *proclist;
924   enum affinity_type type;
925   kmp_hw_t gran;
926   int gran_levels;
927   kmp_affinity_attrs_t core_attr_gran;
928   int compact;
929   int offset;
930   kmp_affinity_flags_t flags;
931   unsigned num_masks;
932   kmp_affin_mask_t *masks;
933   kmp_affinity_ids_t *ids;
934   kmp_affinity_attrs_t *attrs;
935   unsigned num_os_id_masks;
936   kmp_affin_mask_t *os_id_masks;
937   const char *env_var;
938 } kmp_affinity_t;
939 
940 #define KMP_AFFINITY_INIT(env)                                                 \
941   {                                                                            \
942     nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
943         0, 0,                                                                  \
944         {TRUE,  FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE,      \
945          FALSE, FALSE, FALSE},                                                 \
946         0, nullptr, nullptr, nullptr, 0, nullptr, env                          \
947   }
948 
949 extern enum affinity_top_method __kmp_affinity_top_method;
950 extern kmp_affinity_t __kmp_affinity;
951 extern kmp_affinity_t __kmp_hh_affinity;
952 extern kmp_affinity_t *__kmp_affinities[2];
953 
954 extern void __kmp_affinity_bind_thread(int which);
955 
956 extern kmp_affin_mask_t *__kmp_affin_fullMask;
957 extern kmp_affin_mask_t *__kmp_affin_origMask;
958 extern char *__kmp_cpuinfo_file;
959 
960 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
961 extern int __kmp_first_osid_with_ecore;
962 #endif
963 
964 #endif /* KMP_AFFINITY_SUPPORTED */
965 
966 // This needs to be kept in sync with the values in omp.h !!!
967 typedef enum kmp_proc_bind_t {
968   proc_bind_false = 0,
969   proc_bind_true,
970   proc_bind_primary,
971   proc_bind_close,
972   proc_bind_spread,
973   proc_bind_intel, // use KMP_AFFINITY interface
974   proc_bind_default
975 } kmp_proc_bind_t;
976 
977 typedef struct kmp_nested_proc_bind_t {
978   kmp_proc_bind_t *bind_types;
979   int size;
980   int used;
981 } kmp_nested_proc_bind_t;
982 
983 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
984 extern kmp_proc_bind_t __kmp_teams_proc_bind;
985 
986 extern int __kmp_display_affinity;
987 extern char *__kmp_affinity_format;
988 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
989 #if OMPT_SUPPORT
990 extern int __kmp_tool;
991 extern char *__kmp_tool_libraries;
992 #endif // OMPT_SUPPORT
993 
994 #if KMP_AFFINITY_SUPPORTED
995 #define KMP_PLACE_ALL (-1)
996 #define KMP_PLACE_UNDEFINED (-2)
997 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
998 #define KMP_AFFINITY_NON_PROC_BIND                                             \
999   ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false ||                 \
1000     __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) &&                \
1001    (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
1002 #endif /* KMP_AFFINITY_SUPPORTED */
1003 
1004 extern int __kmp_affinity_num_places;
1005 
1006 typedef enum kmp_cancel_kind_t {
1007   cancel_noreq = 0,
1008   cancel_parallel = 1,
1009   cancel_loop = 2,
1010   cancel_sections = 3,
1011   cancel_taskgroup = 4
1012 } kmp_cancel_kind_t;
1013 
1014 // KMP_HW_SUBSET support:
1015 typedef struct kmp_hws_item {
1016   int num;
1017   int offset;
1018 } kmp_hws_item_t;
1019 
1020 extern kmp_hws_item_t __kmp_hws_socket;
1021 extern kmp_hws_item_t __kmp_hws_die;
1022 extern kmp_hws_item_t __kmp_hws_node;
1023 extern kmp_hws_item_t __kmp_hws_tile;
1024 extern kmp_hws_item_t __kmp_hws_core;
1025 extern kmp_hws_item_t __kmp_hws_proc;
1026 extern int __kmp_hws_requested;
1027 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1028 
1029 /* ------------------------------------------------------------------------ */
1030 
1031 #define KMP_PAD(type, sz)                                                      \
1032   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1033 
1034 // We need to avoid using -1 as a GTID as +1 is added to the gtid
1035 // when storing it in a lock, and the value 0 is reserved.
1036 #define KMP_GTID_DNE (-2) /* Does not exist */
1037 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1038 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1039 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
1040 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1041 
1042 /* OpenMP 5.0 Memory Management support */
1043 
1044 #ifndef __OMP_H
1045 // Duplicate type definitions from omp.h
1046 typedef uintptr_t omp_uintptr_t;
1047 
1048 typedef enum {
1049   omp_atk_sync_hint = 1,
1050   omp_atk_alignment = 2,
1051   omp_atk_access = 3,
1052   omp_atk_pool_size = 4,
1053   omp_atk_fallback = 5,
1054   omp_atk_fb_data = 6,
1055   omp_atk_pinned = 7,
1056   omp_atk_partition = 8
1057 } omp_alloctrait_key_t;
1058 
1059 typedef enum {
1060   omp_atv_false = 0,
1061   omp_atv_true = 1,
1062   omp_atv_contended = 3,
1063   omp_atv_uncontended = 4,
1064   omp_atv_serialized = 5,
1065   omp_atv_sequential = omp_atv_serialized, // (deprecated)
1066   omp_atv_private = 6,
1067   omp_atv_all = 7,
1068   omp_atv_thread = 8,
1069   omp_atv_pteam = 9,
1070   omp_atv_cgroup = 10,
1071   omp_atv_default_mem_fb = 11,
1072   omp_atv_null_fb = 12,
1073   omp_atv_abort_fb = 13,
1074   omp_atv_allocator_fb = 14,
1075   omp_atv_environment = 15,
1076   omp_atv_nearest = 16,
1077   omp_atv_blocked = 17,
1078   omp_atv_interleaved = 18
1079 } omp_alloctrait_value_t;
1080 #define omp_atv_default ((omp_uintptr_t)-1)
1081 
1082 typedef void *omp_memspace_handle_t;
1083 extern omp_memspace_handle_t const omp_default_mem_space;
1084 extern omp_memspace_handle_t const omp_large_cap_mem_space;
1085 extern omp_memspace_handle_t const omp_const_mem_space;
1086 extern omp_memspace_handle_t const omp_high_bw_mem_space;
1087 extern omp_memspace_handle_t const omp_low_lat_mem_space;
1088 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1089 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1090 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1091 
1092 typedef struct {
1093   omp_alloctrait_key_t key;
1094   omp_uintptr_t value;
1095 } omp_alloctrait_t;
1096 
1097 typedef void *omp_allocator_handle_t;
1098 extern omp_allocator_handle_t const omp_null_allocator;
1099 extern omp_allocator_handle_t const omp_default_mem_alloc;
1100 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1101 extern omp_allocator_handle_t const omp_const_mem_alloc;
1102 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1103 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1104 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1105 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1106 extern omp_allocator_handle_t const omp_thread_mem_alloc;
1107 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1108 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1109 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1110 extern omp_allocator_handle_t const kmp_max_mem_alloc;
1111 extern omp_allocator_handle_t __kmp_def_allocator;
1112 
1113 // end of duplicate type definitions from omp.h
1114 #endif
1115 
1116 extern int __kmp_memkind_available;
1117 
1118 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1119 
1120 typedef struct kmp_allocator_t {
1121   omp_memspace_handle_t memspace;
1122   void **memkind; // pointer to memkind
1123   size_t alignment;
1124   omp_alloctrait_value_t fb;
1125   kmp_allocator_t *fb_data;
1126   kmp_uint64 pool_size;
1127   kmp_uint64 pool_used;
1128   bool pinned;
1129 } kmp_allocator_t;
1130 
1131 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1132                                                     omp_memspace_handle_t,
1133                                                     int ntraits,
1134                                                     omp_alloctrait_t traits[]);
1135 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1136 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1137 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1138 // external interfaces, may be used by compiler
1139 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1140 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1141                                   omp_allocator_handle_t al);
1142 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1143                            omp_allocator_handle_t al);
1144 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1145                             omp_allocator_handle_t al,
1146                             omp_allocator_handle_t free_al);
1147 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1148 // internal interfaces, contain real implementation
1149 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1150                          omp_allocator_handle_t al);
1151 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1152                           omp_allocator_handle_t al);
1153 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1154                            omp_allocator_handle_t al,
1155                            omp_allocator_handle_t free_al);
1156 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1157 
1158 extern void __kmp_init_memkind();
1159 extern void __kmp_fini_memkind();
1160 extern void __kmp_init_target_mem();
1161 
1162 /* ------------------------------------------------------------------------ */
1163 
1164 #if ENABLE_LIBOMPTARGET
1165 extern void __kmp_init_target_task();
1166 #endif
1167 
1168 /* ------------------------------------------------------------------------ */
1169 
1170 #define KMP_UINT64_MAX                                                         \
1171   (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1172 
1173 #define KMP_MIN_NTH 1
1174 
1175 #ifndef KMP_MAX_NTH
1176 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1177 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1178 #else
1179 #ifdef __ve__
1180 // VE's pthread supports only up to 64 threads per a VE process.
1181 // Please check p. 14 of following documentation for more details.
1182 // https://sxauroratsubasa.sakura.ne.jp/documents/veos/en/VEOS_high_level_design.pdf
1183 #define KMP_MAX_NTH 64
1184 #else
1185 #define KMP_MAX_NTH INT_MAX
1186 #endif
1187 #endif
1188 #endif /* KMP_MAX_NTH */
1189 
1190 #ifdef PTHREAD_STACK_MIN
1191 #define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1192 #else
1193 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1194 #endif
1195 
1196 #if KMP_OS_AIX && KMP_ARCH_PPC
1197 #define KMP_MAX_STKSIZE 0x10000000 /* 256Mb max size on 32-bit AIX */
1198 #else
1199 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1200 #endif
1201 
1202 #if KMP_ARCH_X86
1203 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1204 #elif KMP_ARCH_X86_64
1205 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1206 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1207 #elif KMP_ARCH_VE
1208 // Minimum stack size for pthread for VE is 4MB.
1209 //   https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1210 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1211 #elif KMP_OS_AIX
1212 // The default stack size for worker threads on AIX is 4MB.
1213 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1214 #else
1215 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1216 #endif
1217 
1218 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1219 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1220 #define KMP_MAX_MALLOC_POOL_INCR                                               \
1221   (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1222 
1223 #define KMP_MIN_STKOFFSET (0)
1224 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1225 #if KMP_OS_DARWIN
1226 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1227 #else
1228 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1229 #endif
1230 
1231 #define KMP_MIN_STKPADDING (0)
1232 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1233 
1234 #define KMP_BLOCKTIME_MULTIPLIER                                               \
1235   (1000000) /* number of blocktime units per second */
1236 #define KMP_MIN_BLOCKTIME (0)
1237 #define KMP_MAX_BLOCKTIME                                                      \
1238   (INT_MAX) /* Must be this for "infinite" setting the work */
1239 
1240 /* __kmp_blocktime is in microseconds */
1241 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1242 
1243 #if KMP_USE_MONITOR
1244 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1245 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1246 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1247 
1248 /* Calculate new number of monitor wakeups for a specific block time based on
1249    previous monitor_wakeups. Only allow increasing number of wakeups */
1250 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups)                 \
1251   (((blocktime) == KMP_MAX_BLOCKTIME)   ? (monitor_wakeups)                    \
1252    : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS              \
1253    : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime)))            \
1254        ? (monitor_wakeups)                                                     \
1255        : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1256 
1257 /* Calculate number of intervals for a specific block time based on
1258    monitor_wakeups */
1259 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups)               \
1260   (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) /        \
1261    (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1262 #else
1263 #define KMP_BLOCKTIME(team, tid)                                               \
1264   (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1265 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1266 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1267 extern kmp_uint64 __kmp_ticks_per_msec;
1268 extern kmp_uint64 __kmp_ticks_per_usec;
1269 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1270 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1271 #else
1272 #define KMP_NOW() __kmp_hardware_timestamp()
1273 #endif
1274 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1275   ((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1276 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1277 #else
1278 // System time is retrieved sporadically while blocking.
1279 extern kmp_uint64 __kmp_now_nsec();
1280 #define KMP_NOW() __kmp_now_nsec()
1281 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1282   ((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1283 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1284 #endif
1285 #endif // KMP_USE_MONITOR
1286 
1287 #define KMP_MIN_STATSCOLS 40
1288 #define KMP_MAX_STATSCOLS 4096
1289 #define KMP_DEFAULT_STATSCOLS 80
1290 
1291 #define KMP_MIN_INTERVAL 0
1292 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1293 #define KMP_DEFAULT_INTERVAL 0
1294 
1295 #define KMP_MIN_CHUNK 1
1296 #define KMP_MAX_CHUNK (INT_MAX - 1)
1297 #define KMP_DEFAULT_CHUNK 1
1298 
1299 #define KMP_MIN_DISP_NUM_BUFF 1
1300 #define KMP_DFLT_DISP_NUM_BUFF 7
1301 #define KMP_MAX_DISP_NUM_BUFF 4096
1302 
1303 #define KMP_MAX_ORDERED 8
1304 
1305 #define KMP_MAX_FIELDS 32
1306 
1307 #define KMP_MAX_BRANCH_BITS 31
1308 
1309 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1310 
1311 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1312 
1313 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1314 
1315 /* Minimum number of threads before switch to TLS gtid (experimentally
1316    determined) */
1317 /* josh TODO: what about OS X* tuning? */
1318 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1319 #define KMP_TLS_GTID_MIN 5
1320 #else
1321 #define KMP_TLS_GTID_MIN INT_MAX
1322 #endif
1323 
1324 #define KMP_MASTER_TID(tid) (0 == (tid))
1325 #define KMP_WORKER_TID(tid) (0 != (tid))
1326 
1327 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1328 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1329 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1330 
1331 #ifndef TRUE
1332 #define FALSE 0
1333 #define TRUE (!FALSE)
1334 #endif
1335 
1336 /* NOTE: all of the following constants must be even */
1337 
1338 #if KMP_OS_WINDOWS
1339 #define KMP_INIT_WAIT 64U /* initial number of spin-tests   */
1340 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1341 #elif KMP_OS_LINUX
1342 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1343 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1344 #elif KMP_OS_DARWIN
1345 /* TODO: tune for KMP_OS_DARWIN */
1346 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1347 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1348 #elif KMP_OS_DRAGONFLY
1349 /* TODO: tune for KMP_OS_DRAGONFLY */
1350 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1351 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1352 #elif KMP_OS_FREEBSD
1353 /* TODO: tune for KMP_OS_FREEBSD */
1354 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1355 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1356 #elif KMP_OS_NETBSD
1357 /* TODO: tune for KMP_OS_NETBSD */
1358 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1359 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1360 #elif KMP_OS_OPENBSD
1361 /* TODO: tune for KMP_OS_OPENBSD */
1362 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1363 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1364 #elif KMP_OS_HURD
1365 /* TODO: tune for KMP_OS_HURD */
1366 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1367 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1368 #elif KMP_OS_SOLARIS
1369 /* TODO: tune for KMP_OS_SOLARIS */
1370 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1371 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1372 #elif KMP_OS_WASI
1373 /* TODO: tune for KMP_OS_WASI */
1374 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1375 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1376 #elif KMP_OS_AIX
1377 /* TODO: tune for KMP_OS_AIX */
1378 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1379 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1380 #endif
1381 
1382 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1383 typedef struct kmp_cpuid {
1384   kmp_uint32 eax;
1385   kmp_uint32 ebx;
1386   kmp_uint32 ecx;
1387   kmp_uint32 edx;
1388 } kmp_cpuid_t;
1389 
1390 typedef struct kmp_cpuinfo_flags_t {
1391   unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1392   unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1393   unsigned hybrid : 1;
1394   unsigned reserved : 29; // Ensure size of 32 bits
1395 } kmp_cpuinfo_flags_t;
1396 
1397 typedef struct kmp_cpuinfo {
1398   int initialized; // If 0, other fields are not initialized.
1399   int signature; // CPUID(1).EAX
1400   int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1401   int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1402   // Model << 4 ) + Model)
1403   int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1404   kmp_cpuinfo_flags_t flags;
1405   int apic_id;
1406   kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1407   char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1408 } kmp_cpuinfo_t;
1409 
1410 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1411 
1412 #if KMP_OS_UNIX
1413 // subleaf is only needed for cache and topology discovery and can be set to
1414 // zero in most cases
1415 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1416   __asm__ __volatile__("cpuid"
1417                        : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1418                        : "a"(leaf), "c"(subleaf));
1419 }
1420 // Load p into FPU control word
1421 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1422   __asm__ __volatile__("fldcw %0" : : "m"(*p));
1423 }
1424 // Store FPU control word into p
1425 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1426   __asm__ __volatile__("fstcw %0" : "=m"(*p));
1427 }
1428 static inline void __kmp_clear_x87_fpu_status_word() {
1429 #if KMP_MIC
1430   // 32-bit protected mode x87 FPU state
1431   struct x87_fpu_state {
1432     unsigned cw;
1433     unsigned sw;
1434     unsigned tw;
1435     unsigned fip;
1436     unsigned fips;
1437     unsigned fdp;
1438     unsigned fds;
1439   };
1440   struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1441   __asm__ __volatile__("fstenv %0\n\t" // store FP env
1442                        "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1443                        "fldenv %0\n\t" // load FP env back
1444                        : "+m"(fpu_state), "+m"(fpu_state.sw));
1445 #else
1446   __asm__ __volatile__("fnclex");
1447 #endif // KMP_MIC
1448 }
1449 #if __SSE__
1450 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1451 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1452 #else
1453 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1454 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1455 #endif
1456 #else
1457 // Windows still has these as external functions in assembly file
1458 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1459 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1460 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1461 extern void __kmp_clear_x87_fpu_status_word();
1462 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1463 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1464 #endif // KMP_OS_UNIX
1465 
1466 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1467 
1468 // User-level Monitor/Mwait
1469 #if KMP_HAVE_UMWAIT
1470 // We always try for UMWAIT first
1471 #if KMP_HAVE_WAITPKG_INTRINSICS
1472 #if KMP_HAVE_IMMINTRIN_H
1473 #include <immintrin.h>
1474 #elif KMP_HAVE_INTRIN_H
1475 #include <intrin.h>
1476 #endif
1477 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1478 
1479 KMP_ATTRIBUTE_TARGET_WAITPKG
1480 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1481 #if !KMP_HAVE_WAITPKG_INTRINSICS
1482   uint32_t timeHi = uint32_t(counter >> 32);
1483   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1484   char flag;
1485   __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1486                    "setb   %0"
1487                    // The "=q" restraint means any register accessible as rl
1488                    //   in 32-bit mode: a, b, c, and d;
1489                    //   in 64-bit mode: any integer register
1490                    : "=q"(flag)
1491                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1492                    :);
1493   return flag;
1494 #else
1495   return _tpause(hint, counter);
1496 #endif
1497 }
1498 KMP_ATTRIBUTE_TARGET_WAITPKG
1499 static inline void __kmp_umonitor(void *cacheline) {
1500 #if !KMP_HAVE_WAITPKG_INTRINSICS
1501   __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1502                    :
1503                    : "a"(cacheline)
1504                    :);
1505 #else
1506   _umonitor(cacheline);
1507 #endif
1508 }
1509 KMP_ATTRIBUTE_TARGET_WAITPKG
1510 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1511 #if !KMP_HAVE_WAITPKG_INTRINSICS
1512   uint32_t timeHi = uint32_t(counter >> 32);
1513   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1514   char flag;
1515   __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1516                    "setb   %0"
1517                    // The "=q" restraint means any register accessible as rl
1518                    //   in 32-bit mode: a, b, c, and d;
1519                    //   in 64-bit mode: any integer register
1520                    : "=q"(flag)
1521                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1522                    :);
1523   return flag;
1524 #else
1525   return _umwait(hint, counter);
1526 #endif
1527 }
1528 #elif KMP_HAVE_MWAIT
1529 #if KMP_OS_UNIX
1530 #include <pmmintrin.h>
1531 #else
1532 #include <intrin.h>
1533 #endif
1534 #if KMP_OS_UNIX
1535 __attribute__((target("sse3")))
1536 #endif
1537 static inline void
1538 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1539   _mm_monitor(cacheline, extensions, hints);
1540 }
1541 #if KMP_OS_UNIX
1542 __attribute__((target("sse3")))
1543 #endif
1544 static inline void
1545 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1546   _mm_mwait(extensions, hints);
1547 }
1548 #endif // KMP_HAVE_UMWAIT
1549 
1550 #if KMP_ARCH_X86
1551 extern void __kmp_x86_pause(void);
1552 #elif KMP_MIC
1553 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1554 // regression after removal of extra PAUSE from spin loops. Changing
1555 // the delay from 100 to 300 showed even better performance than double PAUSE
1556 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1557 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1558 #else
1559 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1560 #endif
1561 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1562 #elif KMP_ARCH_PPC64
1563 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1564 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1565 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1566 #define KMP_CPU_PAUSE()                                                        \
1567   do {                                                                         \
1568     KMP_PPC64_PRI_LOW();                                                       \
1569     KMP_PPC64_PRI_MED();                                                       \
1570     KMP_PPC64_PRI_LOC_MB();                                                    \
1571   } while (0)
1572 #else
1573 #define KMP_CPU_PAUSE() /* nothing to do */
1574 #endif
1575 
1576 #define KMP_INIT_YIELD(count)                                                  \
1577   { (count) = __kmp_yield_init; }
1578 
1579 #define KMP_INIT_BACKOFF(time)                                                 \
1580   { (time) = __kmp_pause_init; }
1581 
1582 #define KMP_OVERSUBSCRIBED                                                     \
1583   (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1584 
1585 #define KMP_TRY_YIELD                                                          \
1586   ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1587 
1588 #define KMP_TRY_YIELD_OVERSUB                                                  \
1589   ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1590 
1591 #define KMP_YIELD(cond)                                                        \
1592   {                                                                            \
1593     KMP_CPU_PAUSE();                                                           \
1594     if ((cond) && (KMP_TRY_YIELD))                                             \
1595       __kmp_yield();                                                           \
1596   }
1597 
1598 #define KMP_YIELD_OVERSUB()                                                    \
1599   {                                                                            \
1600     KMP_CPU_PAUSE();                                                           \
1601     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1602       __kmp_yield();                                                           \
1603   }
1604 
1605 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1606 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1607 #define KMP_YIELD_SPIN(count)                                                  \
1608   {                                                                            \
1609     KMP_CPU_PAUSE();                                                           \
1610     if (KMP_TRY_YIELD) {                                                       \
1611       (count) -= 2;                                                            \
1612       if (!(count)) {                                                          \
1613         __kmp_yield();                                                         \
1614         (count) = __kmp_yield_next;                                            \
1615       }                                                                        \
1616     }                                                                          \
1617   }
1618 
1619 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1620 // (C0.2) state, which improves performance of other SMT threads on the same
1621 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1622 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1623 // available, fall back to the regular CPU pause and yield combination.
1624 #if KMP_HAVE_UMWAIT
1625 #define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1626 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1627   {                                                                            \
1628     if (__kmp_tpause_enabled) {                                                \
1629       if (KMP_OVERSUBSCRIBED) {                                                \
1630         __kmp_tpause(0, (time));                                               \
1631       } else {                                                                 \
1632         __kmp_tpause(__kmp_tpause_hint, (time));                               \
1633       }                                                                        \
1634       (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK;                          \
1635     } else {                                                                   \
1636       KMP_CPU_PAUSE();                                                         \
1637       if ((KMP_TRY_YIELD_OVERSUB)) {                                           \
1638         __kmp_yield();                                                         \
1639       } else if (__kmp_use_yield == 1) {                                       \
1640         (count) -= 2;                                                          \
1641         if (!(count)) {                                                        \
1642           __kmp_yield();                                                       \
1643           (count) = __kmp_yield_next;                                          \
1644         }                                                                      \
1645       }                                                                        \
1646     }                                                                          \
1647   }
1648 #else
1649 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1650   {                                                                            \
1651     KMP_CPU_PAUSE();                                                           \
1652     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1653       __kmp_yield();                                                           \
1654     else if (__kmp_use_yield == 1) {                                           \
1655       (count) -= 2;                                                            \
1656       if (!(count)) {                                                          \
1657         __kmp_yield();                                                         \
1658         (count) = __kmp_yield_next;                                            \
1659       }                                                                        \
1660     }                                                                          \
1661   }
1662 #endif // KMP_HAVE_UMWAIT
1663 
1664 /* ------------------------------------------------------------------------ */
1665 /* Support datatypes for the orphaned construct nesting checks.             */
1666 /* ------------------------------------------------------------------------ */
1667 
1668 /* When adding to this enum, add its corresponding string in cons_text_c[]
1669  * array in kmp_error.cpp */
1670 enum cons_type {
1671   ct_none,
1672   ct_parallel,
1673   ct_pdo,
1674   ct_pdo_ordered,
1675   ct_psections,
1676   ct_psingle,
1677   ct_critical,
1678   ct_ordered_in_parallel,
1679   ct_ordered_in_pdo,
1680   ct_master,
1681   ct_reduce,
1682   ct_barrier,
1683   ct_masked
1684 };
1685 
1686 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1687 
1688 struct cons_data {
1689   ident_t const *ident;
1690   enum cons_type type;
1691   int prev;
1692   kmp_user_lock_p
1693       name; /* address exclusively for critical section name comparison */
1694 };
1695 
1696 struct cons_header {
1697   int p_top, w_top, s_top;
1698   int stack_size, stack_top;
1699   struct cons_data *stack_data;
1700 };
1701 
1702 struct kmp_region_info {
1703   char *text;
1704   int offset[KMP_MAX_FIELDS];
1705   int length[KMP_MAX_FIELDS];
1706 };
1707 
1708 /* ---------------------------------------------------------------------- */
1709 /* ---------------------------------------------------------------------- */
1710 
1711 #if KMP_OS_WINDOWS
1712 typedef HANDLE kmp_thread_t;
1713 typedef DWORD kmp_key_t;
1714 #endif /* KMP_OS_WINDOWS */
1715 
1716 #if KMP_OS_UNIX
1717 typedef pthread_t kmp_thread_t;
1718 typedef pthread_key_t kmp_key_t;
1719 #endif
1720 
1721 extern kmp_key_t __kmp_gtid_threadprivate_key;
1722 
1723 typedef struct kmp_sys_info {
1724   long maxrss; /* the maximum resident set size utilized (in kilobytes)     */
1725   long minflt; /* the number of page faults serviced without any I/O        */
1726   long majflt; /* the number of page faults serviced that required I/O      */
1727   long nswap; /* the number of times a process was "swapped" out of memory */
1728   long inblock; /* the number of times the file system had to perform input  */
1729   long oublock; /* the number of times the file system had to perform output */
1730   long nvcsw; /* the number of times a context switch was voluntarily      */
1731   long nivcsw; /* the number of times a context switch was forced           */
1732 } kmp_sys_info_t;
1733 
1734 #if USE_ITT_BUILD
1735 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1736 // required type here. Later we will check the type meets requirements.
1737 typedef int kmp_itt_mark_t;
1738 #define KMP_ITT_DEBUG 0
1739 #endif /* USE_ITT_BUILD */
1740 
1741 typedef kmp_int32 kmp_critical_name[8];
1742 
1743 /*!
1744 @ingroup PARALLEL
1745 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1746 The arguments to the outlined function are
1747 @param global_tid the global thread identity of the thread executing the
1748 function.
1749 @param bound_tid  the local identity of the thread executing the function
1750 @param ... pointers to shared variables accessed by the function.
1751 */
1752 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1753 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1754                                  ...);
1755 
1756 /*!
1757 @ingroup THREADPRIVATE
1758 @{
1759 */
1760 /* ---------------------------------------------------------------------------
1761  */
1762 /* Threadprivate initialization/finalization function declarations */
1763 
1764 /*  for non-array objects:  __kmpc_threadprivate_register()  */
1765 
1766 /*!
1767  Pointer to the constructor function.
1768  The first argument is the <tt>this</tt> pointer
1769 */
1770 typedef void *(*kmpc_ctor)(void *);
1771 
1772 /*!
1773  Pointer to the destructor function.
1774  The first argument is the <tt>this</tt> pointer
1775 */
1776 typedef void (*kmpc_dtor)(
1777     void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1778                               compiler */
1779 /*!
1780  Pointer to an alternate constructor.
1781  The first argument is the <tt>this</tt> pointer.
1782 */
1783 typedef void *(*kmpc_cctor)(void *, void *);
1784 
1785 /* for array objects: __kmpc_threadprivate_register_vec() */
1786 /* First arg: "this" pointer */
1787 /* Last arg: number of array elements */
1788 /*!
1789  Array constructor.
1790  First argument is the <tt>this</tt> pointer
1791  Second argument the number of array elements.
1792 */
1793 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1794 /*!
1795  Pointer to the array destructor function.
1796  The first argument is the <tt>this</tt> pointer
1797  Second argument the number of array elements.
1798 */
1799 typedef void (*kmpc_dtor_vec)(void *, size_t);
1800 /*!
1801  Array constructor.
1802  First argument is the <tt>this</tt> pointer
1803  Third argument the number of array elements.
1804 */
1805 typedef void *(*kmpc_cctor_vec)(void *, void *,
1806                                 size_t); /* function unused by compiler */
1807 
1808 /*!
1809 @}
1810 */
1811 
1812 /* keeps tracked of threadprivate cache allocations for cleanup later */
1813 typedef struct kmp_cached_addr {
1814   void **addr; /* address of allocated cache */
1815   void ***compiler_cache; /* pointer to compiler's cache */
1816   void *data; /* pointer to global data */
1817   struct kmp_cached_addr *next; /* pointer to next cached address */
1818 } kmp_cached_addr_t;
1819 
1820 struct private_data {
1821   struct private_data *next; /* The next descriptor in the list      */
1822   void *data; /* The data buffer for this descriptor  */
1823   int more; /* The repeat count for this descriptor */
1824   size_t size; /* The data size for this descriptor    */
1825 };
1826 
1827 struct private_common {
1828   struct private_common *next;
1829   struct private_common *link;
1830   void *gbl_addr;
1831   void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1832   size_t cmn_size;
1833 };
1834 
1835 struct shared_common {
1836   struct shared_common *next;
1837   struct private_data *pod_init;
1838   void *obj_init;
1839   void *gbl_addr;
1840   union {
1841     kmpc_ctor ctor;
1842     kmpc_ctor_vec ctorv;
1843   } ct;
1844   union {
1845     kmpc_cctor cctor;
1846     kmpc_cctor_vec cctorv;
1847   } cct;
1848   union {
1849     kmpc_dtor dtor;
1850     kmpc_dtor_vec dtorv;
1851   } dt;
1852   size_t vec_len;
1853   int is_vec;
1854   size_t cmn_size;
1855 };
1856 
1857 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1858 #define KMP_HASH_TABLE_SIZE                                                    \
1859   (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1860 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1861 #define KMP_HASH(x)                                                            \
1862   ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1863 
1864 struct common_table {
1865   struct private_common *data[KMP_HASH_TABLE_SIZE];
1866 };
1867 
1868 struct shared_table {
1869   struct shared_common *data[KMP_HASH_TABLE_SIZE];
1870 };
1871 
1872 /* ------------------------------------------------------------------------ */
1873 
1874 #if KMP_USE_HIER_SCHED
1875 // Shared barrier data that exists inside a single unit of the scheduling
1876 // hierarchy
1877 typedef struct kmp_hier_private_bdata_t {
1878   kmp_int32 num_active;
1879   kmp_uint64 index;
1880   kmp_uint64 wait_val[2];
1881 } kmp_hier_private_bdata_t;
1882 #endif
1883 
1884 typedef struct kmp_sched_flags {
1885   unsigned ordered : 1;
1886   unsigned nomerge : 1;
1887   unsigned contains_last : 1;
1888   unsigned use_hier : 1; // Used in KMP_USE_HIER_SCHED code
1889   unsigned use_hybrid : 1; // Used in KMP_WEIGHTED_ITERATIONS_SUPPORTED code
1890   unsigned unused : 27;
1891 } kmp_sched_flags_t;
1892 
1893 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1894 
1895 #if KMP_STATIC_STEAL_ENABLED
1896 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1897   kmp_int32 count;
1898   kmp_int32 ub;
1899   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1900   kmp_int32 lb;
1901   kmp_int32 st;
1902   kmp_int32 tc;
1903   kmp_lock_t *steal_lock; // lock used for chunk stealing
1904 
1905   kmp_uint32 ordered_lower;
1906   kmp_uint32 ordered_upper;
1907 
1908   // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1909   //    a) parm3 is properly aligned and
1910   //    b) all parm1-4 are on the same cache line.
1911   // Because of parm1-4 are used together, performance seems to be better
1912   // if they are on the same cache line (not measured though).
1913 
1914   struct KMP_ALIGN(32) {
1915     kmp_int32 parm1;
1916     kmp_int32 parm2;
1917     kmp_int32 parm3;
1918     kmp_int32 parm4;
1919   };
1920 
1921 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1922   kmp_uint32 pchunks;
1923   kmp_uint32 num_procs_with_pcore;
1924   kmp_int32 first_thread_with_ecore;
1925 #endif
1926 #if KMP_OS_WINDOWS
1927   kmp_int32 last_upper;
1928 #endif /* KMP_OS_WINDOWS */
1929 } dispatch_private_info32_t;
1930 
1931 #if CACHE_LINE <= 128
1932 KMP_BUILD_ASSERT(sizeof(dispatch_private_info32_t) <= 128);
1933 #endif
1934 
1935 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1936   kmp_int64 count; // current chunk number for static & static-steal scheduling
1937   kmp_int64 ub; /* upper-bound */
1938   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1939   kmp_int64 lb; /* lower-bound */
1940   kmp_int64 st; /* stride */
1941   kmp_int64 tc; /* trip count (number of iterations) */
1942   kmp_lock_t *steal_lock; // lock used for chunk stealing
1943 
1944   kmp_uint64 ordered_lower;
1945   kmp_uint64 ordered_upper;
1946   /* parm[1-4] are used in different ways by different scheduling algorithms */
1947 
1948   // KMP_ALIGN(32) ensures ( if the KMP_ALIGN macro is turned on )
1949   //    a) parm3 is properly aligned and
1950   //    b) all parm1-4 are in the same cache line.
1951   // Because of parm1-4 are used together, performance seems to be better
1952   // if they are in the same line (not measured though).
1953   struct KMP_ALIGN(32) {
1954     kmp_int64 parm1;
1955     kmp_int64 parm2;
1956     kmp_int64 parm3;
1957     kmp_int64 parm4;
1958   };
1959 
1960 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1961   kmp_uint64 pchunks;
1962   kmp_uint64 num_procs_with_pcore;
1963   kmp_int64 first_thread_with_ecore;
1964 #endif
1965 
1966 #if KMP_OS_WINDOWS
1967   kmp_int64 last_upper;
1968 #endif /* KMP_OS_WINDOWS */
1969 } dispatch_private_info64_t;
1970 
1971 #if CACHE_LINE <= 128
1972 KMP_BUILD_ASSERT(sizeof(dispatch_private_info64_t) <= 128);
1973 #endif
1974 
1975 #else /* KMP_STATIC_STEAL_ENABLED */
1976 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1977   kmp_int32 lb;
1978   kmp_int32 ub;
1979   kmp_int32 st;
1980   kmp_int32 tc;
1981 
1982   kmp_int32 parm1;
1983   kmp_int32 parm2;
1984   kmp_int32 parm3;
1985   kmp_int32 parm4;
1986 
1987   kmp_int32 count;
1988 
1989   kmp_uint32 ordered_lower;
1990   kmp_uint32 ordered_upper;
1991 #if KMP_OS_WINDOWS
1992   kmp_int32 last_upper;
1993 #endif /* KMP_OS_WINDOWS */
1994 } dispatch_private_info32_t;
1995 
1996 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1997   kmp_int64 lb; /* lower-bound */
1998   kmp_int64 ub; /* upper-bound */
1999   kmp_int64 st; /* stride */
2000   kmp_int64 tc; /* trip count (number of iterations) */
2001 
2002   /* parm[1-4] are used in different ways by different scheduling algorithms */
2003   kmp_int64 parm1;
2004   kmp_int64 parm2;
2005   kmp_int64 parm3;
2006   kmp_int64 parm4;
2007 
2008   kmp_int64 count; /* current chunk number for static scheduling */
2009 
2010   kmp_uint64 ordered_lower;
2011   kmp_uint64 ordered_upper;
2012 #if KMP_OS_WINDOWS
2013   kmp_int64 last_upper;
2014 #endif /* KMP_OS_WINDOWS */
2015 } dispatch_private_info64_t;
2016 #endif /* KMP_STATIC_STEAL_ENABLED */
2017 
2018 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
2019   union private_info {
2020     dispatch_private_info32_t p32;
2021     dispatch_private_info64_t p64;
2022   } u;
2023   enum sched_type schedule; /* scheduling algorithm */
2024   kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
2025   std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
2026   kmp_int32 ordered_bumped;
2027   // Stack of buffers for nest of serial regions
2028   struct dispatch_private_info *next;
2029   kmp_int32 type_size; /* the size of types in private_info */
2030 #if KMP_USE_HIER_SCHED
2031   kmp_int32 hier_id;
2032   void *parent; /* hierarchical scheduling parent pointer */
2033 #endif
2034   enum cons_type pushed_ws;
2035 } dispatch_private_info_t;
2036 
2037 typedef struct dispatch_shared_info32 {
2038   /* chunk index under dynamic, number of idle threads under static-steal;
2039      iteration index otherwise */
2040   volatile kmp_uint32 iteration;
2041   volatile kmp_int32 num_done;
2042   volatile kmp_uint32 ordered_iteration;
2043   // Dummy to retain the structure size after making ordered_iteration scalar
2044   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
2045 } dispatch_shared_info32_t;
2046 
2047 typedef struct dispatch_shared_info64 {
2048   /* chunk index under dynamic, number of idle threads under static-steal;
2049      iteration index otherwise */
2050   volatile kmp_uint64 iteration;
2051   volatile kmp_int64 num_done;
2052   volatile kmp_uint64 ordered_iteration;
2053   // Dummy to retain the structure size after making ordered_iteration scalar
2054   kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
2055 } dispatch_shared_info64_t;
2056 
2057 typedef struct dispatch_shared_info {
2058   union shared_info {
2059     dispatch_shared_info32_t s32;
2060     dispatch_shared_info64_t s64;
2061   } u;
2062   volatile kmp_uint32 buffer_index;
2063   volatile kmp_int32 doacross_buf_idx; // teamwise index
2064   volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
2065   kmp_int32 doacross_num_done; // count finished threads
2066 #if KMP_USE_HIER_SCHED
2067   void *hier;
2068 #endif
2069 #if KMP_USE_HWLOC
2070   // When linking with libhwloc, the ORDERED EPCC test slows down on big
2071   // machines (> 48 cores). Performance analysis showed that a cache thrash
2072   // was occurring and this padding helps alleviate the problem.
2073   char padding[64];
2074 #endif
2075 } dispatch_shared_info_t;
2076 
2077 typedef struct kmp_disp {
2078   /* Vector for ORDERED SECTION */
2079   void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2080   /* Vector for END ORDERED SECTION */
2081   void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2082 
2083   dispatch_shared_info_t *th_dispatch_sh_current;
2084   dispatch_private_info_t *th_dispatch_pr_current;
2085 
2086   dispatch_private_info_t *th_disp_buffer;
2087   kmp_uint32 th_disp_index;
2088   kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2089   volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2090   kmp_int64 *th_doacross_info; // info on loop bounds
2091 #if KMP_USE_INTERNODE_ALIGNMENT
2092   char more_padding[INTERNODE_CACHE_LINE];
2093 #endif
2094 } kmp_disp_t;
2095 
2096 /* ------------------------------------------------------------------------ */
2097 /* Barrier stuff */
2098 
2099 /* constants for barrier state update */
2100 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2101 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2102 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2103 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2104 
2105 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2106 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2107 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2108 
2109 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2110 #error "Barrier sleep bit must be smaller than barrier bump bit"
2111 #endif
2112 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2113 #error "Barrier unused bit must be smaller than barrier bump bit"
2114 #endif
2115 
2116 // Constants for release barrier wait state: currently, hierarchical only
2117 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2118 #define KMP_BARRIER_OWN_FLAG                                                   \
2119   1 // Normal state; worker waiting on own b_go flag in release
2120 #define KMP_BARRIER_PARENT_FLAG                                                \
2121   2 // Special state; worker waiting on parent's b_go flag in release
2122 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG                                         \
2123   3 // Special state; tells worker to shift from parent to own b_go
2124 #define KMP_BARRIER_SWITCHING                                                  \
2125   4 // Special state; worker resets appropriate flag on wake-up
2126 
2127 #define KMP_NOT_SAFE_TO_REAP                                                   \
2128   0 // Thread th_reap_state: not safe to reap (tasking)
2129 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2130 
2131 // The flag_type describes the storage used for the flag.
2132 enum flag_type {
2133   flag32, /**< atomic 32 bit flags */
2134   flag64, /**< 64 bit flags */
2135   atomic_flag64, /**< atomic 64 bit flags */
2136   flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */
2137   flag_unset
2138 };
2139 
2140 enum barrier_type {
2141   bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2142                            barriers if enabled) */
2143   bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2144 #if KMP_FAST_REDUCTION_BARRIER
2145   bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2146 #endif // KMP_FAST_REDUCTION_BARRIER
2147   bs_last_barrier /* Just a placeholder to mark the end */
2148 };
2149 
2150 // to work with reduction barriers just like with plain barriers
2151 #if !KMP_FAST_REDUCTION_BARRIER
2152 #define bs_reduction_barrier bs_plain_barrier
2153 #endif // KMP_FAST_REDUCTION_BARRIER
2154 
2155 typedef enum kmp_bar_pat { /* Barrier communication patterns */
2156                            bp_linear_bar =
2157                                0, /* Single level (degenerate) tree */
2158                            bp_tree_bar =
2159                                1, /* Balanced tree with branching factor 2^n */
2160                            bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2161                                                 branching factor 2^n */
2162                            bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2163                            bp_dist_bar = 4, /* Distributed barrier */
2164                            bp_last_bar /* Placeholder to mark the end */
2165 } kmp_bar_pat_e;
2166 
2167 #define KMP_BARRIER_ICV_PUSH 1
2168 
2169 /* Record for holding the values of the internal controls stack records */
2170 typedef struct kmp_internal_control {
2171   int serial_nesting_level; /* corresponds to the value of the
2172                                th_team_serialized field */
2173   kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2174                        thread) */
2175   kmp_int8
2176       bt_set; /* internal control for whether blocktime is explicitly set */
2177   int blocktime; /* internal control for blocktime */
2178 #if KMP_USE_MONITOR
2179   int bt_intervals; /* internal control for blocktime intervals */
2180 #endif
2181   int nproc; /* internal control for #threads for next parallel region (per
2182                 thread) */
2183   int thread_limit; /* internal control for thread-limit-var */
2184   int task_thread_limit; /* internal control for thread-limit-var of a task*/
2185   int max_active_levels; /* internal control for max_active_levels */
2186   kmp_r_sched_t
2187       sched; /* internal control for runtime schedule {sched,chunk} pair */
2188   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
2189   kmp_int32 default_device; /* internal control for default device */
2190   struct kmp_internal_control *next;
2191 } kmp_internal_control_t;
2192 
2193 static inline void copy_icvs(kmp_internal_control_t *dst,
2194                              kmp_internal_control_t *src) {
2195   *dst = *src;
2196 }
2197 
2198 /* Thread barrier needs volatile barrier fields */
2199 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2200   // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2201   // uses of it). It is not explicitly aligned below, because we *don't* want
2202   // it to be padded -- instead, we fit b_go into the same cache line with
2203   // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2204   kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2205   // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2206   // same NGO store
2207   volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2208   KMP_ALIGN_CACHE volatile kmp_uint64
2209       b_arrived; // STATE => task reached synch point.
2210   kmp_uint32 *skip_per_level;
2211   kmp_uint32 my_level;
2212   kmp_int32 parent_tid;
2213   kmp_int32 old_tid;
2214   kmp_uint32 depth;
2215   struct kmp_bstate *parent_bar;
2216   kmp_team_t *team;
2217   kmp_uint64 leaf_state;
2218   kmp_uint32 nproc;
2219   kmp_uint8 base_leaf_kids;
2220   kmp_uint8 leaf_kids;
2221   kmp_uint8 offset;
2222   kmp_uint8 wait_flag;
2223   kmp_uint8 use_oncore_barrier;
2224 #if USE_DEBUGGER
2225   // The following field is intended for the debugger solely. Only the worker
2226   // thread itself accesses this field: the worker increases it by 1 when it
2227   // arrives to a barrier.
2228   KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2229 #endif /* USE_DEBUGGER */
2230 } kmp_bstate_t;
2231 
2232 union KMP_ALIGN_CACHE kmp_barrier_union {
2233   double b_align; /* use worst case alignment */
2234   char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2235   kmp_bstate_t bb;
2236 };
2237 
2238 typedef union kmp_barrier_union kmp_balign_t;
2239 
2240 /* Team barrier needs only non-volatile arrived counter */
2241 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2242   double b_align; /* use worst case alignment */
2243   char b_pad[CACHE_LINE];
2244   struct {
2245     kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2246 #if USE_DEBUGGER
2247     // The following two fields are indended for the debugger solely. Only
2248     // primary thread of the team accesses these fields: the first one is
2249     // increased by 1 when the primary thread arrives to a barrier, the second
2250     // one is increased by one when all the threads arrived.
2251     kmp_uint b_master_arrived;
2252     kmp_uint b_team_arrived;
2253 #endif
2254   };
2255 };
2256 
2257 typedef union kmp_barrier_team_union kmp_balign_team_t;
2258 
2259 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2260    threads when a condition changes.  This is to workaround an NPTL bug where
2261    padding was added to pthread_cond_t which caused the initialization routine
2262    to write outside of the structure if compiled on pre-NPTL threads.  */
2263 #if KMP_OS_WINDOWS
2264 typedef struct kmp_win32_mutex {
2265   /* The Lock */
2266   CRITICAL_SECTION cs;
2267 } kmp_win32_mutex_t;
2268 
2269 typedef struct kmp_win32_cond {
2270   /* Count of the number of waiters. */
2271   int waiters_count_;
2272 
2273   /* Serialize access to <waiters_count_> */
2274   kmp_win32_mutex_t waiters_count_lock_;
2275 
2276   /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2277   int release_count_;
2278 
2279   /* Keeps track of the current "generation" so that we don't allow */
2280   /* one thread to steal all the "releases" from the broadcast. */
2281   int wait_generation_count_;
2282 
2283   /* A manual-reset event that's used to block and release waiting threads. */
2284   HANDLE event_;
2285 } kmp_win32_cond_t;
2286 #endif
2287 
2288 #if KMP_OS_UNIX
2289 
2290 union KMP_ALIGN_CACHE kmp_cond_union {
2291   double c_align;
2292   char c_pad[CACHE_LINE];
2293   pthread_cond_t c_cond;
2294 };
2295 
2296 typedef union kmp_cond_union kmp_cond_align_t;
2297 
2298 union KMP_ALIGN_CACHE kmp_mutex_union {
2299   double m_align;
2300   char m_pad[CACHE_LINE];
2301   pthread_mutex_t m_mutex;
2302 };
2303 
2304 typedef union kmp_mutex_union kmp_mutex_align_t;
2305 
2306 #endif /* KMP_OS_UNIX */
2307 
2308 typedef struct kmp_desc_base {
2309   void *ds_stackbase;
2310   size_t ds_stacksize;
2311   int ds_stackgrow;
2312   kmp_thread_t ds_thread;
2313   volatile int ds_tid;
2314   int ds_gtid;
2315 #if KMP_OS_WINDOWS
2316   volatile int ds_alive;
2317   DWORD ds_thread_id;
2318 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2319    However, debugger support (libomp_db) cannot work with handles, because they
2320    uncomparable. For example, debugger requests info about thread with handle h.
2321    h is valid within debugger process, and meaningless within debugee process.
2322    Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2323    within debugee process, but it is a *new* handle which does *not* equal to
2324    any other handle in debugee... The only way to compare handles is convert
2325    them to system-wide ids. GetThreadId() function is available only in
2326    Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2327    on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2328    thread id by call to GetCurrentThreadId() from within the thread and save it
2329    to let libomp_db identify threads.  */
2330 #endif /* KMP_OS_WINDOWS */
2331 } kmp_desc_base_t;
2332 
2333 typedef union KMP_ALIGN_CACHE kmp_desc {
2334   double ds_align; /* use worst case alignment */
2335   char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2336   kmp_desc_base_t ds;
2337 } kmp_desc_t;
2338 
2339 typedef struct kmp_local {
2340   volatile int this_construct; /* count of single's encountered by thread */
2341   void *reduce_data;
2342 #if KMP_USE_BGET
2343   void *bget_data;
2344   void *bget_list;
2345 #if !USE_CMP_XCHG_FOR_BGET
2346 #ifdef USE_QUEUING_LOCK_FOR_BGET
2347   kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2348 #else
2349   kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2350 // bootstrap lock so we can use it at library
2351 // shutdown.
2352 #endif /* USE_LOCK_FOR_BGET */
2353 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2354 #endif /* KMP_USE_BGET */
2355 
2356   PACKED_REDUCTION_METHOD_T
2357   packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2358                               __kmpc_end_reduce*() */
2359 
2360 } kmp_local_t;
2361 
2362 #define KMP_CHECK_UPDATE(a, b)                                                 \
2363   if ((a) != (b))                                                              \
2364   (a) = (b)
2365 #define KMP_CHECK_UPDATE_SYNC(a, b)                                            \
2366   if ((a) != (b))                                                              \
2367   TCW_SYNC_PTR((a), (b))
2368 
2369 #define get__blocktime(xteam, xtid)                                            \
2370   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2371 #define get__bt_set(xteam, xtid)                                               \
2372   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2373 #if KMP_USE_MONITOR
2374 #define get__bt_intervals(xteam, xtid)                                         \
2375   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2376 #endif
2377 
2378 #define get__dynamic_2(xteam, xtid)                                            \
2379   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2380 #define get__nproc_2(xteam, xtid)                                              \
2381   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2382 #define get__sched_2(xteam, xtid)                                              \
2383   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2384 
2385 #define set__blocktime_team(xteam, xtid, xval)                                 \
2386   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) =     \
2387        (xval))
2388 
2389 #if KMP_USE_MONITOR
2390 #define set__bt_intervals_team(xteam, xtid, xval)                              \
2391   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) =  \
2392        (xval))
2393 #endif
2394 
2395 #define set__bt_set_team(xteam, xtid, xval)                                    \
2396   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2397 
2398 #define set__dynamic(xthread, xval)                                            \
2399   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2400 #define get__dynamic(xthread)                                                  \
2401   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2402 
2403 #define set__nproc(xthread, xval)                                              \
2404   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2405 
2406 #define set__thread_limit(xthread, xval)                                       \
2407   (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2408 
2409 #define set__max_active_levels(xthread, xval)                                  \
2410   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2411 
2412 #define get__max_active_levels(xthread)                                        \
2413   ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2414 
2415 #define set__sched(xthread, xval)                                              \
2416   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2417 
2418 #define set__proc_bind(xthread, xval)                                          \
2419   (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2420 #define get__proc_bind(xthread)                                                \
2421   ((xthread)->th.th_current_task->td_icvs.proc_bind)
2422 
2423 // OpenMP tasking data structures
2424 
2425 typedef enum kmp_tasking_mode {
2426   tskm_immediate_exec = 0,
2427   tskm_extra_barrier = 1,
2428   tskm_task_teams = 2,
2429   tskm_max = 2
2430 } kmp_tasking_mode_t;
2431 
2432 extern kmp_tasking_mode_t
2433     __kmp_tasking_mode; /* determines how/when to execute tasks */
2434 extern int __kmp_task_stealing_constraint;
2435 extern int __kmp_enable_task_throttling;
2436 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2437 // specified, defaults to 0 otherwise
2438 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2439 extern kmp_int32 __kmp_max_task_priority;
2440 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2441 extern kmp_uint64 __kmp_taskloop_min_tasks;
2442 
2443 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2444    taskdata first */
2445 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2446 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2447 
2448 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2449 // were spawned and queued since the previous barrier release.
2450 #define KMP_TASKING_ENABLED(task_team)                                         \
2451   (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2452 /*!
2453 @ingroup BASIC_TYPES
2454 @{
2455 */
2456 
2457 /*!
2458  */
2459 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2460 
2461 typedef union kmp_cmplrdata {
2462   kmp_int32 priority; /**< priority specified by user for the task */
2463   kmp_routine_entry_t
2464       destructors; /* pointer to function to invoke deconstructors of
2465                       firstprivate C++ objects */
2466   /* future data */
2467 } kmp_cmplrdata_t;
2468 
2469 /*  sizeof_kmp_task_t passed as arg to kmpc_omp_task call  */
2470 /*!
2471  */
2472 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2473   void *shareds; /**< pointer to block of pointers to shared vars   */
2474   kmp_routine_entry_t
2475       routine; /**< pointer to routine to call for executing task */
2476   kmp_int32 part_id; /**< part id for the task                          */
2477   kmp_cmplrdata_t
2478       data1; /* Two known optional additions: destructors and priority */
2479   kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2480   /* future data */
2481   /*  private vars  */
2482 } kmp_task_t;
2483 
2484 /*!
2485 @}
2486 */
2487 
2488 typedef struct kmp_taskgroup {
2489   std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2490   std::atomic<kmp_int32>
2491       cancel_request; // request for cancellation of this taskgroup
2492   struct kmp_taskgroup *parent; // parent taskgroup
2493   // Block of data to perform task reduction
2494   void *reduce_data; // reduction related info
2495   kmp_int32 reduce_num_data; // number of data items to reduce
2496   uintptr_t *gomp_data; // gomp reduction data
2497 } kmp_taskgroup_t;
2498 
2499 // forward declarations
2500 typedef union kmp_depnode kmp_depnode_t;
2501 typedef struct kmp_depnode_list kmp_depnode_list_t;
2502 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2503 
2504 // macros for checking dep flag as an integer
2505 #define KMP_DEP_IN 0x1
2506 #define KMP_DEP_OUT 0x2
2507 #define KMP_DEP_INOUT 0x3
2508 #define KMP_DEP_MTX 0x4
2509 #define KMP_DEP_SET 0x8
2510 #define KMP_DEP_ALL 0x80
2511 // Compiler sends us this info. Note: some test cases contain an explicit copy
2512 // of this struct and should be in sync with any changes here.
2513 typedef struct kmp_depend_info {
2514   kmp_intptr_t base_addr;
2515   size_t len;
2516   union {
2517     kmp_uint8 flag; // flag as an unsigned char
2518     struct { // flag as a set of 8 bits
2519 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2520       /* Same fields as in the #else branch, but in reverse order */
2521       unsigned all : 1;
2522       unsigned unused : 3;
2523       unsigned set : 1;
2524       unsigned mtx : 1;
2525       unsigned out : 1;
2526       unsigned in : 1;
2527 #else
2528       unsigned in : 1;
2529       unsigned out : 1;
2530       unsigned mtx : 1;
2531       unsigned set : 1;
2532       unsigned unused : 3;
2533       unsigned all : 1;
2534 #endif
2535     } flags;
2536   };
2537 } kmp_depend_info_t;
2538 
2539 // Internal structures to work with task dependencies:
2540 struct kmp_depnode_list {
2541   kmp_depnode_t *node;
2542   kmp_depnode_list_t *next;
2543 };
2544 
2545 // Max number of mutexinoutset dependencies per node
2546 #define MAX_MTX_DEPS 4
2547 
2548 typedef struct kmp_base_depnode {
2549   kmp_depnode_list_t *successors; /* used under lock */
2550   kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2551   kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2552   kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2553   kmp_lock_t lock; /* guards shared fields: task, successors */
2554 #if KMP_SUPPORT_GRAPH_OUTPUT
2555   kmp_uint32 id;
2556 #endif
2557   std::atomic<kmp_int32> npredecessors;
2558   std::atomic<kmp_int32> nrefs;
2559 } kmp_base_depnode_t;
2560 
2561 union KMP_ALIGN_CACHE kmp_depnode {
2562   double dn_align; /* use worst case alignment */
2563   char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2564   kmp_base_depnode_t dn;
2565 };
2566 
2567 struct kmp_dephash_entry {
2568   kmp_intptr_t addr;
2569   kmp_depnode_t *last_out;
2570   kmp_depnode_list_t *last_set;
2571   kmp_depnode_list_t *prev_set;
2572   kmp_uint8 last_flag;
2573   kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2574   kmp_dephash_entry_t *next_in_bucket;
2575 };
2576 
2577 typedef struct kmp_dephash {
2578   kmp_dephash_entry_t **buckets;
2579   size_t size;
2580   kmp_depnode_t *last_all;
2581   size_t generation;
2582   kmp_uint32 nelements;
2583   kmp_uint32 nconflicts;
2584 } kmp_dephash_t;
2585 
2586 typedef struct kmp_task_affinity_info {
2587   kmp_intptr_t base_addr;
2588   size_t len;
2589   struct {
2590     bool flag1 : 1;
2591     bool flag2 : 1;
2592     kmp_int32 reserved : 30;
2593   } flags;
2594 } kmp_task_affinity_info_t;
2595 
2596 typedef enum kmp_event_type_t {
2597   KMP_EVENT_UNINITIALIZED = 0,
2598   KMP_EVENT_ALLOW_COMPLETION = 1
2599 } kmp_event_type_t;
2600 
2601 typedef struct {
2602   kmp_event_type_t type;
2603   kmp_tas_lock_t lock;
2604   union {
2605     kmp_task_t *task;
2606   } ed;
2607 } kmp_event_t;
2608 
2609 #if OMPX_TASKGRAPH
2610 // Initial number of allocated nodes while recording
2611 #define INIT_MAPSIZE 50
2612 
2613 typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2614   unsigned nowait : 1;
2615   unsigned re_record : 1;
2616   unsigned reserved : 30;
2617 } kmp_taskgraph_flags_t;
2618 
2619 /// Represents a TDG node
2620 typedef struct kmp_node_info {
2621   kmp_task_t *task; // Pointer to the actual task
2622   kmp_int32 *successors; // Array of the succesors ids
2623   kmp_int32 nsuccessors; // Number of succesors of the node
2624   std::atomic<kmp_int32>
2625       npredecessors_counter; // Number of predessors on the fly
2626   kmp_int32 npredecessors; // Total number of predecessors
2627   kmp_int32 successors_size; // Number of allocated succesors ids
2628   kmp_taskdata_t *parent_task; // Parent implicit task
2629 } kmp_node_info_t;
2630 
2631 /// Represent a TDG's current status
2632 typedef enum kmp_tdg_status {
2633   KMP_TDG_NONE = 0,
2634   KMP_TDG_RECORDING = 1,
2635   KMP_TDG_READY = 2
2636 } kmp_tdg_status_t;
2637 
2638 /// Structure that contains a TDG
2639 typedef struct kmp_tdg_info {
2640   kmp_int32 tdg_id; // Unique idenfifier of the TDG
2641   kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2642   kmp_int32 map_size; // Number of allocated TDG nodes
2643   kmp_int32 num_roots; // Number of roots tasks int the TDG
2644   kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2645   kmp_node_info_t *record_map; // Array of TDG nodes
2646   kmp_tdg_status_t tdg_status =
2647       KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2648   std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2649   kmp_bootstrap_lock_t
2650       graph_lock; // Protect graph attributes when updated via taskloop_recur
2651   // Taskloop reduction related
2652   void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2653                           // __kmpc_taskred_init
2654   kmp_int32 rec_num_taskred;
2655 } kmp_tdg_info_t;
2656 
2657 extern int __kmp_tdg_dot;
2658 extern kmp_int32 __kmp_max_tdgs;
2659 extern kmp_tdg_info_t **__kmp_global_tdgs;
2660 extern kmp_int32 __kmp_curr_tdg_idx;
2661 extern kmp_int32 __kmp_successors_size;
2662 extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2663 extern kmp_int32 __kmp_num_tdg;
2664 #endif
2665 
2666 #ifdef BUILD_TIED_TASK_STACK
2667 
2668 /* Tied Task stack definitions */
2669 typedef struct kmp_stack_block {
2670   kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2671   struct kmp_stack_block *sb_next;
2672   struct kmp_stack_block *sb_prev;
2673 } kmp_stack_block_t;
2674 
2675 typedef struct kmp_task_stack {
2676   kmp_stack_block_t ts_first_block; // first block of stack entries
2677   kmp_taskdata_t **ts_top; // pointer to the top of stack
2678   kmp_int32 ts_entries; // number of entries on the stack
2679 } kmp_task_stack_t;
2680 
2681 #endif // BUILD_TIED_TASK_STACK
2682 
2683 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2684 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2685   /* Same fields as in the #else branch, but in reverse order */
2686 #if OMPX_TASKGRAPH
2687   unsigned reserved31 : 5;
2688   unsigned onced : 1;
2689 #else
2690   unsigned reserved31 : 6;
2691 #endif
2692   unsigned target : 1;
2693   unsigned native : 1;
2694   unsigned freed : 1;
2695   unsigned complete : 1;
2696   unsigned executing : 1;
2697   unsigned started : 1;
2698   unsigned team_serial : 1;
2699   unsigned tasking_ser : 1;
2700   unsigned task_serial : 1;
2701   unsigned tasktype : 1;
2702   unsigned reserved : 8;
2703   unsigned hidden_helper : 1;
2704   unsigned detachable : 1;
2705   unsigned priority_specified : 1;
2706   unsigned proxy : 1;
2707   unsigned destructors_thunk : 1;
2708   unsigned merged_if0 : 1;
2709   unsigned final : 1;
2710   unsigned tiedness : 1;
2711 #else
2712   /* Compiler flags */ /* Total compiler flags must be 16 bits */
2713   unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2714   unsigned final : 1; /* task is final(1) so execute immediately */
2715   unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2716                               code path */
2717   unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2718                                      invoke destructors from the runtime */
2719   unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2720                          context of the RTL) */
2721   unsigned priority_specified : 1; /* set if the compiler provides priority
2722                                       setting for the task */
2723   unsigned detachable : 1; /* 1 == can detach */
2724   unsigned hidden_helper : 1; /* 1 == hidden helper task */
2725   unsigned reserved : 8; /* reserved for compiler use */
2726 
2727   /* Library flags */ /* Total library flags must be 16 bits */
2728   unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2729   unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2730   unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2731   // (1) or may be deferred (0)
2732   unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2733   // (0) [>= 2 threads]
2734   /* If either team_serial or tasking_ser is set, task team may be NULL */
2735   /* Task State Flags: */
2736   unsigned started : 1; /* 1==started, 0==not started     */
2737   unsigned executing : 1; /* 1==executing, 0==not executing */
2738   unsigned complete : 1; /* 1==complete, 0==not complete   */
2739   unsigned freed : 1; /* 1==freed, 0==allocated        */
2740   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2741   unsigned target : 1;
2742 #if OMPX_TASKGRAPH
2743   unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2744   unsigned reserved31 : 5; /* reserved for library use */
2745 #else
2746   unsigned reserved31 : 6; /* reserved for library use */
2747 #endif
2748 #endif
2749 } kmp_tasking_flags_t;
2750 
2751 typedef struct kmp_target_data {
2752   void *async_handle; // libomptarget async handle for task completion query
2753 } kmp_target_data_t;
2754 
2755 struct kmp_taskdata { /* aligned during dynamic allocation       */
2756   kmp_int32 td_task_id; /* id, assigned by debugger                */
2757   kmp_tasking_flags_t td_flags; /* task flags                              */
2758   kmp_team_t *td_team; /* team for this task                      */
2759   kmp_info_p *td_alloc_thread; /* thread that allocated data structures   */
2760   /* Currently not used except for perhaps IDB */
2761   kmp_taskdata_t *td_parent; /* parent task                             */
2762   kmp_int32 td_level; /* task nesting level                      */
2763   std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2764   ident_t *td_ident; /* task identifier                         */
2765   // Taskwait data.
2766   ident_t *td_taskwait_ident;
2767   kmp_uint32 td_taskwait_counter;
2768   kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2769   KMP_ALIGN_CACHE kmp_internal_control_t
2770       td_icvs; /* Internal control variables for the task */
2771   KMP_ALIGN_CACHE std::atomic<kmp_int32>
2772       td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2773                                    deallocated */
2774   std::atomic<kmp_int32>
2775       td_incomplete_child_tasks; /* Child tasks not yet complete */
2776   kmp_taskgroup_t
2777       *td_taskgroup; // Each task keeps pointer to its current taskgroup
2778   kmp_dephash_t
2779       *td_dephash; // Dependencies for children tasks are tracked from here
2780   kmp_depnode_t
2781       *td_depnode; // Pointer to graph node if this task has dependencies
2782   kmp_task_team_t *td_task_team;
2783   size_t td_size_alloc; // Size of task structure, including shareds etc.
2784 #if defined(KMP_GOMP_COMPAT)
2785   // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2786   kmp_int32 td_size_loop_bounds;
2787 #endif
2788   kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2789 #if defined(KMP_GOMP_COMPAT)
2790   // GOMP sends in a copy function for copy constructors
2791   void (*td_copy_func)(void *, void *);
2792 #endif
2793   kmp_event_t td_allow_completion_event;
2794 #if OMPT_SUPPORT
2795   ompt_task_info_t ompt_task_info;
2796 #endif
2797 #if OMPX_TASKGRAPH
2798   bool is_taskgraph = 0; // whether the task is within a TDG
2799   kmp_tdg_info_t *tdg; // used to associate task with a TDG
2800 #endif
2801   kmp_target_data_t td_target_data;
2802 }; // struct kmp_taskdata
2803 
2804 // Make sure padding above worked
2805 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2806 
2807 // Data for task team but per thread
2808 typedef struct kmp_base_thread_data {
2809   kmp_info_p *td_thr; // Pointer back to thread info
2810   // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2811   // queued?
2812   kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2813   kmp_taskdata_t *
2814       *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2815   kmp_int32 td_deque_size; // Size of deck
2816   kmp_uint32 td_deque_head; // Head of deque (will wrap)
2817   kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2818   kmp_int32 td_deque_ntasks; // Number of tasks in deque
2819   // GEH: shouldn't this be volatile since used in while-spin?
2820   kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2821 #ifdef BUILD_TIED_TASK_STACK
2822   kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2823 // scheduling constraint
2824 #endif // BUILD_TIED_TASK_STACK
2825 } kmp_base_thread_data_t;
2826 
2827 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2828 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2829 
2830 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2831 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2832 
2833 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2834   kmp_base_thread_data_t td;
2835   double td_align; /* use worst case alignment */
2836   char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2837 } kmp_thread_data_t;
2838 
2839 typedef struct kmp_task_pri {
2840   kmp_thread_data_t td;
2841   kmp_int32 priority;
2842   kmp_task_pri *next;
2843 } kmp_task_pri_t;
2844 
2845 // Data for task teams which are used when tasking is enabled for the team
2846 typedef struct kmp_base_task_team {
2847   kmp_bootstrap_lock_t
2848       tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2849   /* must be bootstrap lock since used at library shutdown*/
2850 
2851   // TODO: check performance vs kmp_tas_lock_t
2852   kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2853   kmp_task_pri_t *tt_task_pri_list;
2854 
2855   kmp_task_team_t *tt_next; /* For linking the task team free list */
2856   kmp_thread_data_t
2857       *tt_threads_data; /* Array of per-thread structures for task team */
2858   /* Data survives task team deallocation */
2859   kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2860                                executing this team? */
2861   /* TRUE means tt_threads_data is set up and initialized */
2862   kmp_int32 tt_nproc; /* #threads in team           */
2863   kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2864   kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2865   kmp_int32 tt_untied_task_encountered;
2866   std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2867   // There is hidden helper thread encountered in this task team so that we must
2868   // wait when waiting on task team
2869   kmp_int32 tt_hidden_helper_task_encountered;
2870 
2871   KMP_ALIGN_CACHE
2872   std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2873 
2874   KMP_ALIGN_CACHE
2875   volatile kmp_uint32
2876       tt_active; /* is the team still actively executing tasks */
2877 } kmp_base_task_team_t;
2878 
2879 union KMP_ALIGN_CACHE kmp_task_team {
2880   kmp_base_task_team_t tt;
2881   double tt_align; /* use worst case alignment */
2882   char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2883 };
2884 
2885 typedef struct kmp_task_team_list_t {
2886   kmp_task_team_t *task_team;
2887   kmp_task_team_list_t *next;
2888 } kmp_task_team_list_t;
2889 
2890 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2891 // Free lists keep same-size free memory slots for fast memory allocation
2892 // routines
2893 typedef struct kmp_free_list {
2894   void *th_free_list_self; // Self-allocated tasks free list
2895   void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2896   // threads
2897   void *th_free_list_other; // Non-self free list (to be returned to owner's
2898   // sync list)
2899 } kmp_free_list_t;
2900 #endif
2901 #if KMP_NESTED_HOT_TEAMS
2902 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2903 // are not put in teams pool, and they don't put threads in threads pool.
2904 typedef struct kmp_hot_team_ptr {
2905   kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2906   kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2907 } kmp_hot_team_ptr_t;
2908 #endif
2909 typedef struct kmp_teams_size {
2910   kmp_int32 nteams; // number of teams in a league
2911   kmp_int32 nth; // number of threads in each team of the league
2912 } kmp_teams_size_t;
2913 
2914 // This struct stores a thread that acts as a "root" for a contention
2915 // group. Contention groups are rooted at kmp_root threads, but also at
2916 // each primary thread of each team created in the teams construct.
2917 // This struct therefore also stores a thread_limit associated with
2918 // that contention group, and a counter to track the number of threads
2919 // active in that contention group. Each thread has a list of these: CG
2920 // root threads have an entry in their list in which cg_root refers to
2921 // the thread itself, whereas other workers in the CG will have a
2922 // single entry where cg_root is same as the entry containing their CG
2923 // root. When a thread encounters a teams construct, it will add a new
2924 // entry to the front of its list, because it now roots a new CG.
2925 typedef struct kmp_cg_root {
2926   kmp_info_p *cg_root; // "root" thread for a contention group
2927   // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2928   // thread_limit clause for teams primary threads
2929   kmp_int32 cg_thread_limit;
2930   kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2931   struct kmp_cg_root *up; // pointer to higher level CG root in list
2932 } kmp_cg_root_t;
2933 
2934 // OpenMP thread data structures
2935 
2936 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2937   /* Start with the readonly data which is cache aligned and padded. This is
2938      written before the thread starts working by the primary thread. Uber
2939      masters may update themselves later. Usage does not consider serialized
2940      regions.  */
2941   kmp_desc_t th_info;
2942   kmp_team_p *th_team; /* team we belong to */
2943   kmp_root_p *th_root; /* pointer to root of task hierarchy */
2944   kmp_info_p *th_next_pool; /* next available thread in the pool */
2945   kmp_disp_t *th_dispatch; /* thread's dispatch data */
2946   int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2947 
2948   /* The following are cached from the team info structure */
2949   /* TODO use these in more places as determined to be needed via profiling */
2950   int th_team_nproc; /* number of threads in a team */
2951   kmp_info_p *th_team_master; /* the team's primary thread */
2952   int th_team_serialized; /* team is serialized */
2953   microtask_t th_teams_microtask; /* save entry address for teams construct */
2954   int th_teams_level; /* save initial level of teams construct */
2955 /* it is 0 on device but may be any on host */
2956 
2957 /* The blocktime info is copied from the team struct to the thread struct */
2958 /* at the start of a barrier, and the values stored in the team are used  */
2959 /* at points in the code where the team struct is no longer guaranteed    */
2960 /* to exist (from the POV of worker threads).                             */
2961 #if KMP_USE_MONITOR
2962   int th_team_bt_intervals;
2963   int th_team_bt_set;
2964 #else
2965   kmp_uint64 th_team_bt_intervals;
2966 #endif
2967 
2968 #if KMP_AFFINITY_SUPPORTED
2969   kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2970   kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2971   kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2972 #endif
2973   omp_allocator_handle_t th_def_allocator; /* default allocator */
2974   /* The data set by the primary thread at reinit, then R/W by the worker */
2975   KMP_ALIGN_CACHE int
2976       th_set_nproc; /* if > 0, then only use this request for the next fork */
2977   int *th_set_nested_nth;
2978   bool th_nt_strict; // num_threads clause has strict modifier
2979   ident_t *th_nt_loc; // loc for strict modifier
2980   int th_nt_sev; // error severity for strict modifier
2981   const char *th_nt_msg; // error message for strict modifier
2982   int th_set_nested_nth_sz;
2983 #if KMP_NESTED_HOT_TEAMS
2984   kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2985 #endif
2986   kmp_proc_bind_t
2987       th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2988   kmp_teams_size_t
2989       th_teams_size; /* number of teams/threads in teams construct */
2990 #if KMP_AFFINITY_SUPPORTED
2991   int th_current_place; /* place currently bound to */
2992   int th_new_place; /* place to bind to in par reg */
2993   int th_first_place; /* first place in partition */
2994   int th_last_place; /* last place in partition */
2995 #endif
2996   int th_prev_level; /* previous level for affinity format */
2997   int th_prev_num_threads; /* previous num_threads for affinity format */
2998 #if USE_ITT_BUILD
2999   kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
3000   kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
3001   kmp_uint64 th_frame_time; /* frame timestamp */
3002 #endif /* USE_ITT_BUILD */
3003   kmp_local_t th_local;
3004   struct private_common *th_pri_head;
3005 
3006   /* Now the data only used by the worker (after initial allocation) */
3007   /* TODO the first serial team should actually be stored in the info_t
3008      structure.  this will help reduce initial allocation overhead */
3009   KMP_ALIGN_CACHE kmp_team_p
3010       *th_serial_team; /*serialized team held in reserve*/
3011 
3012 #if OMPT_SUPPORT
3013   ompt_thread_info_t ompt_thread_info;
3014 #endif
3015 
3016   /* The following are also read by the primary thread during reinit */
3017   struct common_table *th_pri_common;
3018 
3019   volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
3020   /* while awaiting queuing lock acquire */
3021 
3022   volatile void *th_sleep_loc; // this points at a kmp_flag<T>
3023   flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
3024 
3025   ident_t *th_ident;
3026   unsigned th_x; // Random number generator data
3027   unsigned th_a; // Random number generator data
3028 
3029   /* Tasking-related data for the thread */
3030   kmp_task_team_t *th_task_team; // Task team struct
3031   kmp_taskdata_t *th_current_task; // Innermost Task being executed
3032   kmp_uint8 th_task_state; // alternating 0/1 for task team identification
3033   kmp_uint32 th_reap_state; // Non-zero indicates thread is not
3034   // tasking, thus safe to reap
3035 
3036   /* More stuff for keeping track of active/sleeping threads (this part is
3037      written by the worker thread) */
3038   kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
3039   int th_active; // ! sleeping; 32 bits for TCR/TCW
3040   std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
3041   // 0 = not used in team; 1 = used in team;
3042   // 2 = transitioning to not used in team; 3 = transitioning to used in team
3043   struct cons_header *th_cons; // used for consistency check
3044 #if KMP_USE_HIER_SCHED
3045   // used for hierarchical scheduling
3046   kmp_hier_private_bdata_t *th_hier_bar_data;
3047 #endif
3048 
3049   /* Add the syncronizing data which is cache aligned and padded. */
3050   KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
3051 
3052   KMP_ALIGN_CACHE volatile kmp_int32
3053       th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
3054 
3055 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
3056 #define NUM_LISTS 4
3057   kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
3058 // allocation routines
3059 #endif
3060 
3061 #if KMP_OS_WINDOWS
3062   kmp_win32_cond_t th_suspend_cv;
3063   kmp_win32_mutex_t th_suspend_mx;
3064   std::atomic<int> th_suspend_init;
3065 #endif
3066 #if KMP_OS_UNIX
3067   kmp_cond_align_t th_suspend_cv;
3068   kmp_mutex_align_t th_suspend_mx;
3069   std::atomic<int> th_suspend_init_count;
3070 #endif
3071 
3072 #if USE_ITT_BUILD
3073   kmp_itt_mark_t th_itt_mark_single;
3074 // alignment ???
3075 #endif /* USE_ITT_BUILD */
3076 #if KMP_STATS_ENABLED
3077   kmp_stats_list *th_stats;
3078 #endif
3079 #if KMP_OS_UNIX
3080   std::atomic<bool> th_blocking;
3081 #endif
3082   kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
3083 } kmp_base_info_t;
3084 
3085 typedef union KMP_ALIGN_CACHE kmp_info {
3086   double th_align; /* use worst case alignment */
3087   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
3088   kmp_base_info_t th;
3089 } kmp_info_t;
3090 
3091 // OpenMP thread team data structures
3092 
3093 typedef struct kmp_base_data {
3094   volatile kmp_uint32 t_value;
3095 } kmp_base_data_t;
3096 
3097 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
3098   double dt_align; /* use worst case alignment */
3099   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3100   kmp_base_data_t dt;
3101 } kmp_sleep_team_t;
3102 
3103 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
3104   double dt_align; /* use worst case alignment */
3105   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3106   kmp_base_data_t dt;
3107 } kmp_ordered_team_t;
3108 
3109 typedef int (*launch_t)(int gtid);
3110 
3111 /* Minimum number of ARGV entries to malloc if necessary */
3112 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3113 
3114 // Set up how many argv pointers will fit in cache lines containing
3115 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3116 // larger value for more space between the primary write/worker read section and
3117 // read/write by all section seems to buy more performance on EPCC PARALLEL.
3118 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3119 #define KMP_INLINE_ARGV_BYTES                                                  \
3120   (4 * CACHE_LINE -                                                            \
3121    ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) +               \
3122      sizeof(kmp_int16) + sizeof(kmp_uint32)) %                                 \
3123     CACHE_LINE))
3124 #else
3125 #define KMP_INLINE_ARGV_BYTES                                                  \
3126   (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3127 #endif
3128 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3129 
3130 typedef struct KMP_ALIGN_CACHE kmp_base_team {
3131   // Synchronization Data
3132   // ---------------------------------------------------------------------------
3133   KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3134   kmp_balign_team_t t_bar[bs_last_barrier];
3135   std::atomic<int> t_construct; // count of single directive encountered by team
3136   char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3137 
3138   // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3139   std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3140   std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3141 
3142   // Primary thread only
3143   // ---------------------------------------------------------------------------
3144   KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3145   int t_master_this_cons; // "this_construct" single counter of primary thread
3146   // in parent team
3147   ident_t *t_ident; // if volatile, have to change too much other crud to
3148   // volatile too
3149   kmp_team_p *t_parent; // parent team
3150   kmp_team_p *t_next_pool; // next free team in the team pool
3151   kmp_disp_t *t_dispatch; // thread's dispatch data
3152   kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3153   kmp_proc_bind_t t_proc_bind; // bind type for par region
3154   int t_primary_task_state; // primary thread's task state saved
3155 #if USE_ITT_BUILD
3156   kmp_uint64 t_region_time; // region begin timestamp
3157 #endif /* USE_ITT_BUILD */
3158 
3159   // Primary thread write, workers read
3160   // --------------------------------------------------------------------------
3161   KMP_ALIGN_CACHE void **t_argv;
3162   int t_argc;
3163   int t_nproc; // number of threads in team
3164   microtask_t t_pkfn;
3165   launch_t t_invoke; // procedure to launch the microtask
3166 
3167 #if OMPT_SUPPORT
3168   ompt_team_info_t ompt_team_info;
3169   ompt_lw_taskteam_t *ompt_serialized_team_info;
3170 #endif
3171 
3172 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3173   kmp_int8 t_fp_control_saved;
3174   kmp_int8 t_pad2b;
3175   kmp_int16 t_x87_fpu_control_word; // FP control regs
3176   kmp_uint32 t_mxcsr;
3177 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3178 
3179   void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3180 
3181   KMP_ALIGN_CACHE kmp_info_t **t_threads;
3182   kmp_taskdata_t
3183       *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3184   int t_level; // nested parallel level
3185 
3186   KMP_ALIGN_CACHE int t_max_argc;
3187   int t_max_nproc; // max threads this team can handle (dynamically expandable)
3188   int t_serialized; // levels deep of serialized teams
3189   dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3190   int t_id; // team's id, assigned by debugger.
3191   int t_active_level; // nested active parallel level
3192   kmp_r_sched_t t_sched; // run-time schedule for the team
3193 #if KMP_AFFINITY_SUPPORTED
3194   int t_first_place; // first & last place in parent thread's partition.
3195   int t_last_place; // Restore these values to primary thread after par region.
3196 #endif // KMP_AFFINITY_SUPPORTED
3197   int t_display_affinity;
3198   int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3199   // omp_set_num_threads() call
3200   omp_allocator_handle_t t_def_allocator; /* default allocator */
3201 
3202 // Read/write by workers as well
3203 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3204   // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3205   // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3206   // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3207   // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3208   char dummy_padding[1024];
3209 #endif
3210   // Internal control stack for additional nested teams.
3211   KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3212   // for SERIALIZED teams nested 2 or more levels deep
3213   // typed flag to store request state of cancellation
3214   std::atomic<kmp_int32> t_cancel_request;
3215   int t_master_active; // save on fork, restore on join
3216   void *t_copypriv_data; // team specific pointer to copyprivate data array
3217 #if KMP_OS_WINDOWS
3218   std::atomic<kmp_uint32> t_copyin_counter;
3219 #endif
3220 #if USE_ITT_BUILD
3221   void *t_stack_id; // team specific stack stitching id (for ittnotify)
3222 #endif /* USE_ITT_BUILD */
3223   distributedBarrier *b; // Distributed barrier data associated with team
3224   kmp_nested_nthreads_t *t_nested_nth;
3225 } kmp_base_team_t;
3226 
3227 // Assert that the list structure fits and aligns within
3228 // the double task team pointer
3229 KMP_BUILD_ASSERT(sizeof(kmp_task_team_t *[2]) == sizeof(kmp_task_team_list_t));
3230 KMP_BUILD_ASSERT(alignof(kmp_task_team_t *[2]) ==
3231                  alignof(kmp_task_team_list_t));
3232 
3233 union KMP_ALIGN_CACHE kmp_team {
3234   kmp_base_team_t t;
3235   double t_align; /* use worst case alignment */
3236   char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3237 };
3238 
3239 typedef union KMP_ALIGN_CACHE kmp_time_global {
3240   double dt_align; /* use worst case alignment */
3241   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3242   kmp_base_data_t dt;
3243 } kmp_time_global_t;
3244 
3245 typedef struct kmp_base_global {
3246   /* cache-aligned */
3247   kmp_time_global_t g_time;
3248 
3249   /* non cache-aligned */
3250   volatile int g_abort;
3251   volatile int g_done;
3252 
3253   int g_dynamic;
3254   enum dynamic_mode g_dynamic_mode;
3255 } kmp_base_global_t;
3256 
3257 typedef union KMP_ALIGN_CACHE kmp_global {
3258   kmp_base_global_t g;
3259   double g_align; /* use worst case alignment */
3260   char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3261 } kmp_global_t;
3262 
3263 typedef struct kmp_base_root {
3264   // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3265   // (r_in_parallel>= 0)
3266   // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3267   // the synch overhead or keeping r_active
3268   volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3269   // keeps a count of active parallel regions per root
3270   std::atomic<int> r_in_parallel;
3271   // GEH: This is misnamed, should be r_active_levels
3272   kmp_team_t *r_root_team;
3273   kmp_team_t *r_hot_team;
3274   kmp_info_t *r_uber_thread;
3275   kmp_lock_t r_begin_lock;
3276   volatile int r_begin;
3277   int r_blocktime; /* blocktime for this root and descendants */
3278 #if KMP_AFFINITY_SUPPORTED
3279   int r_affinity_assigned;
3280 #endif // KMP_AFFINITY_SUPPORTED
3281 } kmp_base_root_t;
3282 
3283 typedef union KMP_ALIGN_CACHE kmp_root {
3284   kmp_base_root_t r;
3285   double r_align; /* use worst case alignment */
3286   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3287 } kmp_root_t;
3288 
3289 struct fortran_inx_info {
3290   kmp_int32 data;
3291 };
3292 
3293 // This list type exists to hold old __kmp_threads arrays so that
3294 // old references to them may complete while reallocation takes place when
3295 // expanding the array. The items in this list are kept alive until library
3296 // shutdown.
3297 typedef struct kmp_old_threads_list_t {
3298   kmp_info_t **threads;
3299   struct kmp_old_threads_list_t *next;
3300 } kmp_old_threads_list_t;
3301 
3302 /* ------------------------------------------------------------------------ */
3303 
3304 extern int __kmp_settings;
3305 extern int __kmp_duplicate_library_ok;
3306 #if USE_ITT_BUILD
3307 extern int __kmp_forkjoin_frames;
3308 extern int __kmp_forkjoin_frames_mode;
3309 #endif
3310 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3311 extern int __kmp_determ_red;
3312 
3313 #ifdef KMP_DEBUG
3314 extern int kmp_a_debug;
3315 extern int kmp_b_debug;
3316 extern int kmp_c_debug;
3317 extern int kmp_d_debug;
3318 extern int kmp_e_debug;
3319 extern int kmp_f_debug;
3320 #endif /* KMP_DEBUG */
3321 
3322 /* For debug information logging using rotating buffer */
3323 #define KMP_DEBUG_BUF_LINES_INIT 512
3324 #define KMP_DEBUG_BUF_LINES_MIN 1
3325 
3326 #define KMP_DEBUG_BUF_CHARS_INIT 128
3327 #define KMP_DEBUG_BUF_CHARS_MIN 2
3328 
3329 extern int
3330     __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3331 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3332 extern int
3333     __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3334 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3335                                       entry pointer */
3336 
3337 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3338 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3339                                               printed in buffer so far */
3340 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3341                                           recommended in warnings */
3342 /* end rotating debug buffer */
3343 
3344 #ifdef KMP_DEBUG
3345 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3346 
3347 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3348 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3349 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3350 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3351 extern int __kmp_par_range_lb;
3352 extern int __kmp_par_range_ub;
3353 #endif
3354 
3355 /* For printing out dynamic storage map for threads and teams */
3356 extern int
3357     __kmp_storage_map; /* True means print storage map for threads and teams */
3358 extern int __kmp_storage_map_verbose; /* True means storage map includes
3359                                          placement info */
3360 extern int __kmp_storage_map_verbose_specified;
3361 
3362 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3363 extern kmp_cpuinfo_t __kmp_cpuinfo;
3364 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3365 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3366 static inline bool __kmp_is_hybrid_cpu() { return true; }
3367 #else
3368 static inline bool __kmp_is_hybrid_cpu() { return false; }
3369 #endif
3370 
3371 extern volatile int __kmp_init_serial;
3372 extern volatile int __kmp_init_gtid;
3373 extern volatile int __kmp_init_common;
3374 extern volatile int __kmp_need_register_serial;
3375 extern volatile int __kmp_init_middle;
3376 extern volatile int __kmp_init_parallel;
3377 #if KMP_USE_MONITOR
3378 extern volatile int __kmp_init_monitor;
3379 #endif
3380 extern volatile int __kmp_init_user_locks;
3381 extern volatile int __kmp_init_hidden_helper_threads;
3382 extern int __kmp_init_counter;
3383 extern int __kmp_root_counter;
3384 extern int __kmp_version;
3385 
3386 /* list of address of allocated caches for commons */
3387 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3388 
3389 /* Barrier algorithm types and options */
3390 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3391 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3392 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3393 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3394 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3395 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3396 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3397 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3398 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3399 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3400 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3401 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3402 
3403 /* Global Locks */
3404 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3405 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3406 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3407 extern kmp_bootstrap_lock_t
3408     __kmp_exit_lock; /* exit() is not always thread-safe */
3409 #if KMP_USE_MONITOR
3410 extern kmp_bootstrap_lock_t
3411     __kmp_monitor_lock; /* control monitor thread creation */
3412 #endif
3413 extern kmp_bootstrap_lock_t
3414     __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3415                              __kmp_threads expansion to co-exist */
3416 
3417 extern kmp_lock_t __kmp_global_lock; /* control OS/global access  */
3418 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access  */
3419 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3420 
3421 extern enum library_type __kmp_library;
3422 
3423 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3424 extern enum sched_type __kmp_static; /* default static scheduling method */
3425 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3426 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3427 extern int __kmp_chunk; /* default runtime chunk size */
3428 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3429 
3430 extern size_t __kmp_stksize; /* stack size per thread         */
3431 #if KMP_USE_MONITOR
3432 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3433 #endif
3434 extern size_t __kmp_stkoffset; /* stack offset per thread       */
3435 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3436 
3437 extern size_t
3438     __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3439 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3440 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3441 extern int __kmp_env_checks; /* was KMP_CHECKS specified?    */
3442 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3443 extern int __kmp_generate_warnings; /* should we issue warnings? */
3444 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3445 
3446 #ifdef DEBUG_SUSPEND
3447 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3448 #endif
3449 
3450 extern kmp_int32 __kmp_use_yield;
3451 extern kmp_int32 __kmp_use_yield_exp_set;
3452 extern kmp_uint32 __kmp_yield_init;
3453 extern kmp_uint32 __kmp_yield_next;
3454 extern kmp_uint64 __kmp_pause_init;
3455 
3456 /* ------------------------------------------------------------------------- */
3457 extern int __kmp_allThreadsSpecified;
3458 
3459 extern size_t __kmp_align_alloc;
3460 /* following data protected by initialization routines */
3461 extern int __kmp_xproc; /* number of processors in the system */
3462 extern int __kmp_avail_proc; /* number of processors available to the process */
3463 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3464 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3465 // maximum total number of concurrently-existing threads on device
3466 extern int __kmp_max_nth;
3467 // maximum total number of concurrently-existing threads in a contention group
3468 extern int __kmp_cg_max_nth;
3469 extern int __kmp_task_max_nth; // max threads used in a task
3470 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3471 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3472                                       __kmp_root */
3473 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3474                                    region a la OMP_NUM_THREADS */
3475 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3476                                       initialization */
3477 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3478                                  used (fixed) */
3479 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3480                                (__kmpc_threadprivate_cached()) */
3481 extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3482                                     blocking (env setting) */
3483 extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3484 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3485 
3486 // Convert raw blocktime from ms to us if needed.
3487 static inline void __kmp_aux_convert_blocktime(int *bt) {
3488   if (__kmp_blocktime_units == 'm') {
3489     if (*bt > INT_MAX / 1000) {
3490       *bt = INT_MAX / 1000;
3491       KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3492     }
3493     *bt = *bt * 1000;
3494   }
3495 }
3496 
3497 #if KMP_USE_MONITOR
3498 extern int
3499     __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3500 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3501                                   blocking */
3502 #endif
3503 #ifdef KMP_ADJUST_BLOCKTIME
3504 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3505 #endif /* KMP_ADJUST_BLOCKTIME */
3506 #ifdef KMP_DFLT_NTH_CORES
3507 extern int __kmp_ncores; /* Total number of cores for threads placement */
3508 #endif
3509 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3510 extern int __kmp_abort_delay;
3511 
3512 extern int __kmp_need_register_atfork_specified;
3513 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3514                                           to install fork handler */
3515 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3516                                0 - not set, will be set at runtime
3517                                1 - using stack search
3518                                2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3519                                    X*) or TlsGetValue(Windows* OS))
3520                                3 - static TLS (__declspec(thread) __kmp_gtid),
3521                                    Linux* OS .so only.  */
3522 extern int
3523     __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3524 #ifdef KMP_TDATA_GTID
3525 extern KMP_THREAD_LOCAL int __kmp_gtid;
3526 #endif
3527 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3528 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3529 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3530 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3531 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3532 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3533 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3534 
3535 // max_active_levels for nested parallelism enabled by default via
3536 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3537 extern int __kmp_dflt_max_active_levels;
3538 // Indicates whether value of __kmp_dflt_max_active_levels was already
3539 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3540 extern bool __kmp_dflt_max_active_levels_set;
3541 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3542                                           concurrent execution per team */
3543 #if KMP_NESTED_HOT_TEAMS
3544 extern int __kmp_hot_teams_mode;
3545 extern int __kmp_hot_teams_max_level;
3546 #endif
3547 
3548 #if KMP_OS_LINUX
3549 extern enum clock_function_type __kmp_clock_function;
3550 extern int __kmp_clock_function_param;
3551 #endif /* KMP_OS_LINUX */
3552 
3553 #if KMP_MIC_SUPPORTED
3554 extern enum mic_type __kmp_mic_type;
3555 #endif
3556 
3557 #ifdef USE_LOAD_BALANCE
3558 extern double __kmp_load_balance_interval; // load balance algorithm interval
3559 #endif /* USE_LOAD_BALANCE */
3560 
3561 #if KMP_USE_ADAPTIVE_LOCKS
3562 
3563 // Parameters for the speculative lock backoff system.
3564 struct kmp_adaptive_backoff_params_t {
3565   // Number of soft retries before it counts as a hard retry.
3566   kmp_uint32 max_soft_retries;
3567   // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3568   // the right
3569   kmp_uint32 max_badness;
3570 };
3571 
3572 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3573 
3574 #if KMP_DEBUG_ADAPTIVE_LOCKS
3575 extern const char *__kmp_speculative_statsfile;
3576 #endif
3577 
3578 #endif // KMP_USE_ADAPTIVE_LOCKS
3579 
3580 extern int __kmp_display_env; /* TRUE or FALSE */
3581 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3582 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3583 extern int __kmp_nteams;
3584 extern int __kmp_teams_thread_limit;
3585 
3586 /* ------------------------------------------------------------------------- */
3587 
3588 /* the following are protected by the fork/join lock */
3589 /* write: lock  read: anytime */
3590 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3591 /* Holds old arrays of __kmp_threads until library shutdown */
3592 extern kmp_old_threads_list_t *__kmp_old_threads_list;
3593 /* read/write: lock */
3594 extern volatile kmp_team_t *__kmp_team_pool;
3595 extern volatile kmp_info_t *__kmp_thread_pool;
3596 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3597 
3598 // total num threads reachable from some root thread including all root threads
3599 extern volatile int __kmp_nth;
3600 /* total number of threads reachable from some root thread including all root
3601    threads, and those in the thread pool */
3602 extern volatile int __kmp_all_nth;
3603 extern std::atomic<int> __kmp_thread_pool_active_nth;
3604 
3605 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3606 /* end data protected by fork/join lock */
3607 /* ------------------------------------------------------------------------- */
3608 
3609 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3610 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3611 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3612 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3613 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3614 
3615 // AT: Which way is correct?
3616 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3617 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3618 #define __kmp_get_team_num_threads(gtid)                                       \
3619   (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3620 
3621 static inline bool KMP_UBER_GTID(int gtid) {
3622   KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3623   KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3624   return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3625           __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3626 }
3627 
3628 static inline int __kmp_tid_from_gtid(int gtid) {
3629   KMP_DEBUG_ASSERT(gtid >= 0);
3630   return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3631 }
3632 
3633 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3634   KMP_DEBUG_ASSERT(tid >= 0 && team);
3635   return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3636 }
3637 
3638 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3639   KMP_DEBUG_ASSERT(thr);
3640   return thr->th.th_info.ds.ds_gtid;
3641 }
3642 
3643 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3644   KMP_DEBUG_ASSERT(gtid >= 0);
3645   return __kmp_threads[gtid];
3646 }
3647 
3648 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3649   KMP_DEBUG_ASSERT(gtid >= 0);
3650   return __kmp_threads[gtid]->th.th_team;
3651 }
3652 
3653 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3654   if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3655     KMP_FATAL(ThreadIdentInvalid);
3656 }
3657 
3658 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3659 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3660 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3661 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3662 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3663 #endif
3664 
3665 #if KMP_HAVE_UMWAIT
3666 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3667 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3668 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3669 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3670 #endif
3671 
3672 /* ------------------------------------------------------------------------- */
3673 
3674 extern kmp_global_t __kmp_global; /* global status */
3675 
3676 extern kmp_info_t __kmp_monitor;
3677 // For Debugging Support Library
3678 extern std::atomic<kmp_int32> __kmp_team_counter;
3679 // For Debugging Support Library
3680 extern std::atomic<kmp_int32> __kmp_task_counter;
3681 
3682 #if USE_DEBUGGER
3683 #define _KMP_GEN_ID(counter)                                                   \
3684   (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3685 #else
3686 #define _KMP_GEN_ID(counter) (~0)
3687 #endif /* USE_DEBUGGER */
3688 
3689 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3690 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3691 
3692 /* ------------------------------------------------------------------------ */
3693 
3694 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3695                                          size_t size, char const *format, ...);
3696 
3697 extern void __kmp_serial_initialize(void);
3698 extern void __kmp_middle_initialize(void);
3699 extern void __kmp_parallel_initialize(void);
3700 
3701 extern void __kmp_internal_begin(void);
3702 extern void __kmp_internal_end_library(int gtid);
3703 extern void __kmp_internal_end_thread(int gtid);
3704 extern void __kmp_internal_end_atexit(void);
3705 extern void __kmp_internal_end_dtor(void);
3706 extern void __kmp_internal_end_dest(void *);
3707 
3708 extern int __kmp_register_root(int initial_thread);
3709 extern void __kmp_unregister_root(int gtid);
3710 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3711 
3712 extern int __kmp_ignore_mppbeg(void);
3713 extern int __kmp_ignore_mppend(void);
3714 
3715 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3716 extern void __kmp_exit_single(int gtid);
3717 
3718 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3719 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3720 
3721 #ifdef USE_LOAD_BALANCE
3722 extern int __kmp_get_load_balance(int);
3723 #endif
3724 
3725 extern int __kmp_get_global_thread_id(void);
3726 extern int __kmp_get_global_thread_id_reg(void);
3727 extern void __kmp_exit_thread(int exit_status);
3728 extern void __kmp_abort(char const *format, ...);
3729 extern void __kmp_abort_thread(void);
3730 KMP_NORETURN extern void __kmp_abort_process(void);
3731 extern void __kmp_warn(char const *format, ...);
3732 
3733 extern void __kmp_set_num_threads(int new_nth, int gtid);
3734 
3735 extern bool __kmp_detect_shm();
3736 extern bool __kmp_detect_tmp();
3737 
3738 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3739 // registered.
3740 static inline kmp_info_t *__kmp_entry_thread() {
3741   int gtid = __kmp_entry_gtid();
3742 
3743   return __kmp_threads[gtid];
3744 }
3745 
3746 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3747 extern int __kmp_get_max_active_levels(int gtid);
3748 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3749 extern int __kmp_get_team_size(int gtid, int level);
3750 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3751 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3752 
3753 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3754 extern void __kmp_init_random(kmp_info_t *thread);
3755 
3756 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3757 extern void __kmp_adjust_num_threads(int new_nproc);
3758 extern void __kmp_check_stksize(size_t *val);
3759 
3760 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3761 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3762 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3763 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3764 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3765 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3766 
3767 #if USE_FAST_MEMORY
3768 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3769                                   size_t size KMP_SRC_LOC_DECL);
3770 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3771 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3772 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3773 #define __kmp_fast_allocate(this_thr, size)                                    \
3774   ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3775 #define __kmp_fast_free(this_thr, ptr)                                         \
3776   ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3777 #endif
3778 
3779 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3780 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3781                                   size_t elsize KMP_SRC_LOC_DECL);
3782 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3783                                    size_t size KMP_SRC_LOC_DECL);
3784 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3785 #define __kmp_thread_malloc(th, size)                                          \
3786   ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3787 #define __kmp_thread_calloc(th, nelem, elsize)                                 \
3788   ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3789 #define __kmp_thread_realloc(th, ptr, size)                                    \
3790   ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3791 #define __kmp_thread_free(th, ptr)                                             \
3792   ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3793 
3794 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3795 extern void __kmp_push_num_threads_list(ident_t *loc, int gtid,
3796                                         kmp_uint32 list_length,
3797                                         int *num_threads_list);
3798 extern void __kmp_set_strict_num_threads(ident_t *loc, int gtid, int sev,
3799                                          const char *msg);
3800 
3801 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3802                                  kmp_proc_bind_t proc_bind);
3803 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3804                                  int num_threads);
3805 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3806                                     int num_teams_ub, int num_threads);
3807 
3808 extern void __kmp_yield();
3809 
3810 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3811                                    enum sched_type schedule, kmp_int32 lb,
3812                                    kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3813 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3814                                     enum sched_type schedule, kmp_uint32 lb,
3815                                     kmp_uint32 ub, kmp_int32 st,
3816                                     kmp_int32 chunk);
3817 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3818                                    enum sched_type schedule, kmp_int64 lb,
3819                                    kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3820 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3821                                     enum sched_type schedule, kmp_uint64 lb,
3822                                     kmp_uint64 ub, kmp_int64 st,
3823                                     kmp_int64 chunk);
3824 
3825 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3826                                   kmp_int32 *p_last, kmp_int32 *p_lb,
3827                                   kmp_int32 *p_ub, kmp_int32 *p_st);
3828 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3829                                    kmp_int32 *p_last, kmp_uint32 *p_lb,
3830                                    kmp_uint32 *p_ub, kmp_int32 *p_st);
3831 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3832                                   kmp_int32 *p_last, kmp_int64 *p_lb,
3833                                   kmp_int64 *p_ub, kmp_int64 *p_st);
3834 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3835                                    kmp_int32 *p_last, kmp_uint64 *p_lb,
3836                                    kmp_uint64 *p_ub, kmp_int64 *p_st);
3837 
3838 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3839 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3840 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3841 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3842 
3843 extern void __kmpc_dispatch_deinit(ident_t *loc, kmp_int32 gtid);
3844 
3845 #ifdef KMP_GOMP_COMPAT
3846 
3847 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3848                                       enum sched_type schedule, kmp_int32 lb,
3849                                       kmp_int32 ub, kmp_int32 st,
3850                                       kmp_int32 chunk, int push_ws);
3851 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3852                                        enum sched_type schedule, kmp_uint32 lb,
3853                                        kmp_uint32 ub, kmp_int32 st,
3854                                        kmp_int32 chunk, int push_ws);
3855 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3856                                       enum sched_type schedule, kmp_int64 lb,
3857                                       kmp_int64 ub, kmp_int64 st,
3858                                       kmp_int64 chunk, int push_ws);
3859 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3860                                        enum sched_type schedule, kmp_uint64 lb,
3861                                        kmp_uint64 ub, kmp_int64 st,
3862                                        kmp_int64 chunk, int push_ws);
3863 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3864 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3865 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3866 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3867 
3868 #endif /* KMP_GOMP_COMPAT */
3869 
3870 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3871 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3872 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3873 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3874 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3875 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3876                                kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3877                                void *obj);
3878 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3879                              kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3880 
3881 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3882                           int final_spin
3883 #if USE_ITT_BUILD
3884                           ,
3885                           void *itt_sync_obj
3886 #endif
3887 );
3888 extern void __kmp_release_64(kmp_flag_64<> *flag);
3889 
3890 extern void __kmp_infinite_loop(void);
3891 
3892 extern void __kmp_cleanup(void);
3893 
3894 #if KMP_HANDLE_SIGNALS
3895 extern int __kmp_handle_signals;
3896 extern void __kmp_install_signals(int parallel_init);
3897 extern void __kmp_remove_signals(void);
3898 #endif
3899 
3900 extern void __kmp_clear_system_time(void);
3901 extern void __kmp_read_system_time(double *delta);
3902 
3903 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3904 
3905 extern void __kmp_expand_host_name(char *buffer, size_t size);
3906 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3907 
3908 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3909 extern void
3910 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3911 #endif
3912 
3913 extern void
3914 __kmp_runtime_initialize(void); /* machine specific initialization */
3915 extern void __kmp_runtime_destroy(void);
3916 
3917 #if KMP_AFFINITY_SUPPORTED
3918 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3919                                        kmp_affin_mask_t *mask);
3920 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3921                                                   kmp_affin_mask_t *mask);
3922 extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3923 extern void __kmp_affinity_uninitialize(void);
3924 extern void __kmp_affinity_set_init_mask(
3925     int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3926 void __kmp_affinity_bind_init_mask(int gtid);
3927 extern void __kmp_affinity_bind_place(int gtid);
3928 extern void __kmp_affinity_determine_capable(const char *env_var);
3929 extern int __kmp_aux_set_affinity(void **mask);
3930 extern int __kmp_aux_get_affinity(void **mask);
3931 extern int __kmp_aux_get_affinity_max_proc();
3932 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3933 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3934 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3935 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3936 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
3937 extern int __kmp_get_first_osid_with_ecore(void);
3938 #endif
3939 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY ||     \
3940     KMP_OS_AIX
3941 extern int kmp_set_thread_affinity_mask_initial(void);
3942 #endif
3943 static inline void __kmp_assign_root_init_mask() {
3944   int gtid = __kmp_entry_gtid();
3945   kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3946   if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3947     __kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3948     __kmp_affinity_bind_init_mask(gtid);
3949     r->r.r_affinity_assigned = TRUE;
3950   }
3951 }
3952 static inline void __kmp_reset_root_init_mask(int gtid) {
3953   if (!KMP_AFFINITY_CAPABLE())
3954     return;
3955   kmp_info_t *th = __kmp_threads[gtid];
3956   kmp_root_t *r = th->th.th_root;
3957   if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3958     __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3959     KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3960     r->r.r_affinity_assigned = FALSE;
3961   }
3962 }
3963 #else /* KMP_AFFINITY_SUPPORTED */
3964 #define __kmp_assign_root_init_mask() /* Nothing */
3965 static inline void __kmp_reset_root_init_mask(int gtid) {}
3966 #endif /* KMP_AFFINITY_SUPPORTED */
3967 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3968 // format string is for affinity, so platforms that do not support
3969 // affinity can still use the other fields, e.g., %n for num_threads
3970 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3971                                          kmp_str_buf_t *buffer);
3972 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3973 
3974 extern void __kmp_cleanup_hierarchy();
3975 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3976 
3977 #if KMP_USE_FUTEX
3978 
3979 extern int __kmp_futex_determine_capable(void);
3980 
3981 #endif // KMP_USE_FUTEX
3982 
3983 extern void __kmp_gtid_set_specific(int gtid);
3984 extern int __kmp_gtid_get_specific(void);
3985 
3986 extern double __kmp_read_cpu_time(void);
3987 
3988 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3989 
3990 #if KMP_USE_MONITOR
3991 extern void __kmp_create_monitor(kmp_info_t *th);
3992 #endif
3993 
3994 extern void *__kmp_launch_thread(kmp_info_t *thr);
3995 
3996 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3997 
3998 #if KMP_OS_WINDOWS
3999 extern int __kmp_still_running(kmp_info_t *th);
4000 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
4001 extern void __kmp_free_handle(kmp_thread_t tHandle);
4002 #endif
4003 
4004 #if KMP_USE_MONITOR
4005 extern void __kmp_reap_monitor(kmp_info_t *th);
4006 #endif
4007 extern void __kmp_reap_worker(kmp_info_t *th);
4008 extern void __kmp_terminate_thread(int gtid);
4009 
4010 extern int __kmp_try_suspend_mx(kmp_info_t *th);
4011 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
4012 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
4013 
4014 extern void __kmp_elapsed(double *);
4015 extern void __kmp_elapsed_tick(double *);
4016 
4017 extern void __kmp_enable(int old_state);
4018 extern void __kmp_disable(int *old_state);
4019 
4020 extern void __kmp_thread_sleep(int millis);
4021 
4022 extern void __kmp_common_initialize(void);
4023 extern void __kmp_common_destroy(void);
4024 extern void __kmp_common_destroy_gtid(int gtid);
4025 
4026 #if KMP_OS_UNIX
4027 extern void __kmp_register_atfork(void);
4028 #endif
4029 extern void __kmp_suspend_initialize(void);
4030 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
4031 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
4032 
4033 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4034                                          int tid);
4035 extern kmp_team_t *
4036 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4037 #if OMPT_SUPPORT
4038                     ompt_data_t ompt_parallel_data,
4039 #endif
4040                     kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
4041                     int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
4042 extern void __kmp_free_thread(kmp_info_t *);
4043 extern void __kmp_free_team(kmp_root_t *,
4044                             kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
4045 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
4046 
4047 /* ------------------------------------------------------------------------ */
4048 
4049 extern void __kmp_initialize_bget(kmp_info_t *th);
4050 extern void __kmp_finalize_bget(kmp_info_t *th);
4051 
4052 KMP_EXPORT void *kmpc_malloc(size_t size);
4053 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
4054 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
4055 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
4056 KMP_EXPORT void kmpc_free(void *ptr);
4057 
4058 /* declarations for internal use */
4059 
4060 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
4061                          size_t reduce_size, void *reduce_data,
4062                          void (*reduce)(void *, void *));
4063 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
4064 extern int __kmp_barrier_gomp_cancel(int gtid);
4065 
4066 /*!
4067  * Tell the fork call which compiler generated the fork call, and therefore how
4068  * to deal with the call.
4069  */
4070 enum fork_context_e {
4071   fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
4072                        microtask internally. */
4073   fork_context_intel, /**< Called from Intel generated code.  */
4074   fork_context_last
4075 };
4076 extern int __kmp_fork_call(ident_t *loc, int gtid,
4077                            enum fork_context_e fork_context, kmp_int32 argc,
4078                            microtask_t microtask, launch_t invoker,
4079                            kmp_va_list ap);
4080 
4081 extern void __kmp_join_call(ident_t *loc, int gtid
4082 #if OMPT_SUPPORT
4083                             ,
4084                             enum fork_context_e fork_context
4085 #endif
4086                             ,
4087                             int exit_teams = 0);
4088 
4089 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
4090 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
4091 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
4092 extern int __kmp_invoke_task_func(int gtid);
4093 extern void __kmp_run_before_invoked_task(int gtid, int tid,
4094                                           kmp_info_t *this_thr,
4095                                           kmp_team_t *team);
4096 extern void __kmp_run_after_invoked_task(int gtid, int tid,
4097                                          kmp_info_t *this_thr,
4098                                          kmp_team_t *team);
4099 
4100 // should never have been exported
4101 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
4102 extern int __kmp_invoke_teams_master(int gtid);
4103 extern void __kmp_teams_master(int gtid);
4104 extern int __kmp_aux_get_team_num();
4105 extern int __kmp_aux_get_num_teams();
4106 extern void __kmp_save_internal_controls(kmp_info_t *thread);
4107 extern void __kmp_user_set_library(enum library_type arg);
4108 extern void __kmp_aux_set_library(enum library_type arg);
4109 extern void __kmp_aux_set_stacksize(size_t arg);
4110 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
4111 extern void __kmp_aux_set_defaults(char const *str, size_t len);
4112 
4113 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
4114 void kmpc_set_blocktime(int arg);
4115 void ompc_set_nested(int flag);
4116 void ompc_set_dynamic(int flag);
4117 void ompc_set_num_threads(int arg);
4118 
4119 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
4120                                               kmp_team_t *team, int tid);
4121 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
4122 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4123                                     kmp_tasking_flags_t *flags,
4124                                     size_t sizeof_kmp_task_t,
4125                                     size_t sizeof_shareds,
4126                                     kmp_routine_entry_t task_entry);
4127 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4128                                      kmp_team_t *team, int tid,
4129                                      int set_curr_task);
4130 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4131 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4132 
4133 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4134                                                        int gtid,
4135                                                        kmp_task_t *task);
4136 extern void __kmp_fulfill_event(kmp_event_t *event);
4137 
4138 extern void __kmp_free_task_team(kmp_info_t *thread,
4139                                  kmp_task_team_t *task_team);
4140 extern void __kmp_reap_task_teams(void);
4141 extern void __kmp_push_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4142 extern void __kmp_pop_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4143 extern void __kmp_wait_to_unref_task_teams(void);
4144 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team);
4145 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4146 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4147 #if USE_ITT_BUILD
4148                                  ,
4149                                  void *itt_sync_obj
4150 #endif /* USE_ITT_BUILD */
4151                                  ,
4152                                  int wait = 1);
4153 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4154                                   int gtid);
4155 #if KMP_DEBUG
4156 #define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr)                         \
4157   KMP_DEBUG_ASSERT(                                                            \
4158       __kmp_tasking_mode != tskm_task_teams || team->t.t_nproc == 1 ||         \
4159       thr->th.th_task_team == team->t.t_task_team[thr->th.th_task_state])
4160 #else
4161 #define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr) /* Nothing */
4162 #endif
4163 
4164 extern int __kmp_is_address_mapped(void *addr);
4165 extern kmp_uint64 __kmp_hardware_timestamp(void);
4166 
4167 #if KMP_OS_UNIX
4168 extern int __kmp_read_from_file(char const *path, char const *format, ...);
4169 #endif
4170 
4171 /* ------------------------------------------------------------------------ */
4172 //
4173 // Assembly routines that have no compiler intrinsic replacement
4174 //
4175 
4176 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4177                                   void *argv[]
4178 #if OMPT_SUPPORT
4179                                   ,
4180                                   void **exit_frame_ptr
4181 #endif
4182 );
4183 
4184 /* ------------------------------------------------------------------------ */
4185 
4186 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4187 KMP_EXPORT void __kmpc_end(ident_t *);
4188 
4189 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4190                                                   kmpc_ctor_vec ctor,
4191                                                   kmpc_cctor_vec cctor,
4192                                                   kmpc_dtor_vec dtor,
4193                                                   size_t vector_length);
4194 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4195                                               kmpc_ctor ctor, kmpc_cctor cctor,
4196                                               kmpc_dtor dtor);
4197 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4198                                       void *data, size_t size);
4199 
4200 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4201 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4202 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4203 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4204 
4205 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4206 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4207                                  kmpc_micro microtask, ...);
4208 KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4209                                     kmpc_micro microtask, kmp_int32 cond,
4210                                     void *args);
4211 
4212 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4213 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4214 
4215 KMP_EXPORT void __kmpc_flush(ident_t *);
4216 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4217 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4218 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4219 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4220                                    kmp_int32 filter);
4221 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4222 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4223 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4224 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4225                                 kmp_critical_name *);
4226 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4227                                     kmp_critical_name *);
4228 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4229                                           kmp_critical_name *, uint32_t hint);
4230 
4231 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4232 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4233 
4234 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4235                                                   kmp_int32 global_tid);
4236 
4237 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4238 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4239 
4240 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4241 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4242                                          kmp_int32 numberOfSections);
4243 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4244 
4245 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4246                                      kmp_int32 schedtype, kmp_int32 *plastiter,
4247                                      kmp_int *plower, kmp_int *pupper,
4248                                      kmp_int *pstride, kmp_int incr,
4249                                      kmp_int chunk);
4250 
4251 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4252 
4253 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4254                                    size_t cpy_size, void *cpy_data,
4255                                    void (*cpy_func)(void *, void *),
4256                                    kmp_int32 didit);
4257 
4258 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4259                                           void *cpy_data);
4260 
4261 extern void KMPC_SET_NUM_THREADS(int arg);
4262 extern void KMPC_SET_DYNAMIC(int flag);
4263 extern void KMPC_SET_NESTED(int flag);
4264 
4265 /* OMP 3.0 tasking interface routines */
4266 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4267                                      kmp_task_t *new_task);
4268 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4269                                              kmp_int32 flags,
4270                                              size_t sizeof_kmp_task_t,
4271                                              size_t sizeof_shareds,
4272                                              kmp_routine_entry_t task_entry);
4273 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4274     ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4275     size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4276 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4277                                           kmp_task_t *task);
4278 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4279                                              kmp_task_t *task);
4280 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4281                                            kmp_task_t *new_task);
4282 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4283 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4284                                           int end_part);
4285 
4286 #if TASK_UNUSED
4287 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4288 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4289                               kmp_task_t *task);
4290 #endif // TASK_UNUSED
4291 
4292 /* ------------------------------------------------------------------------ */
4293 
4294 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4295 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4296 
4297 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4298     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4299     kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4300     kmp_depend_info_t *noalias_dep_list);
4301 
4302 KMP_EXPORT kmp_base_depnode_t *__kmpc_task_get_depnode(kmp_task_t *task);
4303 
4304 KMP_EXPORT kmp_depnode_list_t *__kmpc_task_get_successors(kmp_task_t *task);
4305 
4306 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4307                                      kmp_int32 ndeps,
4308                                      kmp_depend_info_t *dep_list,
4309                                      kmp_int32 ndeps_noalias,
4310                                      kmp_depend_info_t *noalias_dep_list);
4311 /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4312  *                               Placeholder for taskwait with nowait clause.*/
4313 KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4314                                             kmp_int32 ndeps,
4315                                             kmp_depend_info_t *dep_list,
4316                                             kmp_int32 ndeps_noalias,
4317                                             kmp_depend_info_t *noalias_dep_list,
4318                                             kmp_int32 has_no_wait);
4319 
4320 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4321                                 bool serialize_immediate);
4322 
4323 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4324                                    kmp_int32 cncl_kind);
4325 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4326                                               kmp_int32 cncl_kind);
4327 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4328 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4329 
4330 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4331 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4332 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4333                                 kmp_int32 if_val, kmp_uint64 *lb,
4334                                 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4335                                 kmp_int32 sched, kmp_uint64 grainsize,
4336                                 void *task_dup);
4337 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4338                                   kmp_task_t *task, kmp_int32 if_val,
4339                                   kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4340                                   kmp_int32 nogroup, kmp_int32 sched,
4341                                   kmp_uint64 grainsize, kmp_int32 modifier,
4342                                   void *task_dup);
4343 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4344 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4345 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4346 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4347                                                      int is_ws, int num,
4348                                                      void *data);
4349 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4350                                               int num, void *data);
4351 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4352                                                     int is_ws);
4353 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4354     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4355     kmp_task_affinity_info_t *affin_list);
4356 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4357 KMP_EXPORT int __kmp_get_max_teams(void);
4358 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4359 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4360 
4361 /* Interface target task integration */
4362 KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4363 KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4364 
4365 /* Lock interface routines (fast versions with gtid passed in) */
4366 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4367                                  void **user_lock);
4368 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4369                                       void **user_lock);
4370 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4371                                     void **user_lock);
4372 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4373                                          void **user_lock);
4374 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4375 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4376                                      void **user_lock);
4377 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4378                                   void **user_lock);
4379 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4380                                        void **user_lock);
4381 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4382 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4383                                      void **user_lock);
4384 
4385 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4386                                            void **user_lock, uintptr_t hint);
4387 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4388                                                 void **user_lock,
4389                                                 uintptr_t hint);
4390 
4391 #if OMPX_TASKGRAPH
4392 // Taskgraph's Record & Replay mechanism
4393 // __kmp_tdg_is_recording: check whether a given TDG is recording
4394 // status: the tdg's current status
4395 static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4396   return status == KMP_TDG_RECORDING;
4397 }
4398 
4399 KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4400                                               kmp_int32 input_flags,
4401                                               kmp_int32 tdg_id);
4402 KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4403                                        kmp_int32 input_flags, kmp_int32 tdg_id);
4404 #endif
4405 /* Interface to fast scalable reduce methods routines */
4406 
4407 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4408     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4409     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4410     kmp_critical_name *lck);
4411 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4412                                          kmp_critical_name *lck);
4413 KMP_EXPORT kmp_int32 __kmpc_reduce(
4414     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4415     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4416     kmp_critical_name *lck);
4417 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4418                                   kmp_critical_name *lck);
4419 
4420 /* Internal fast reduction routines */
4421 
4422 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4423     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4424     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4425     kmp_critical_name *lck);
4426 
4427 // this function is for testing set/get/determine reduce method
4428 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4429 
4430 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4431 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4432 
4433 // C++ port
4434 // missing 'extern "C"' declarations
4435 
4436 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4437 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4438 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4439                                         kmp_int32 num_threads);
4440 KMP_EXPORT void __kmpc_push_num_threads_strict(ident_t *loc,
4441                                                kmp_int32 global_tid,
4442                                                kmp_int32 num_threads,
4443                                                int severity,
4444                                                const char *message);
4445 
4446 KMP_EXPORT void __kmpc_push_num_threads_list(ident_t *loc, kmp_int32 global_tid,
4447                                              kmp_uint32 list_length,
4448                                              kmp_int32 *num_threads_list);
4449 KMP_EXPORT void __kmpc_push_num_threads_list_strict(
4450     ident_t *loc, kmp_int32 global_tid, kmp_uint32 list_length,
4451     kmp_int32 *num_threads_list, int severity, const char *message);
4452 
4453 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4454                                       int proc_bind);
4455 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4456                                       kmp_int32 num_teams,
4457                                       kmp_int32 num_threads);
4458 KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4459                                         kmp_int32 thread_limit);
4460 /* Function for OpenMP 5.1 num_teams clause */
4461 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4462                                          kmp_int32 num_teams_lb,
4463                                          kmp_int32 num_teams_ub,
4464                                          kmp_int32 num_threads);
4465 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4466                                   kmpc_micro microtask, ...);
4467 struct kmp_dim { // loop bounds info casted to kmp_int64
4468   kmp_int64 lo; // lower
4469   kmp_int64 up; // upper
4470   kmp_int64 st; // stride
4471 };
4472 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4473                                      kmp_int32 num_dims,
4474                                      const struct kmp_dim *dims);
4475 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4476                                      const kmp_int64 *vec);
4477 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4478                                      const kmp_int64 *vec);
4479 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4480 
4481 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4482                                              void *data, size_t size,
4483                                              void ***cache);
4484 
4485 // The routines below are not exported.
4486 // Consider making them 'static' in corresponding source files.
4487 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4488                                            void *data_addr, size_t pc_size);
4489 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4490                                                 void *data_addr,
4491                                                 size_t pc_size);
4492 void __kmp_threadprivate_resize_cache(int newCapacity);
4493 void __kmp_cleanup_threadprivate_caches();
4494 
4495 // ompc_, kmpc_ entries moved from omp.h.
4496 #if KMP_OS_WINDOWS
4497 #define KMPC_CONVENTION __cdecl
4498 #else
4499 #define KMPC_CONVENTION
4500 #endif
4501 
4502 #ifndef __OMP_H
4503 typedef enum omp_sched_t {
4504   omp_sched_static = 1,
4505   omp_sched_dynamic = 2,
4506   omp_sched_guided = 3,
4507   omp_sched_auto = 4
4508 } omp_sched_t;
4509 typedef void *kmp_affinity_mask_t;
4510 #endif
4511 
4512 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4513 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4514 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4515 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4516 KMP_EXPORT int KMPC_CONVENTION
4517 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4518 KMP_EXPORT int KMPC_CONVENTION
4519 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4520 KMP_EXPORT int KMPC_CONVENTION
4521 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4522 
4523 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4524 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4525 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4526 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4527 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4528 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4529 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4530 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4531 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4532                                               char const *format);
4533 
4534 enum kmp_target_offload_kind {
4535   tgt_disabled = 0,
4536   tgt_default = 1,
4537   tgt_mandatory = 2
4538 };
4539 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4540 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4541 extern kmp_target_offload_kind_t __kmp_target_offload;
4542 extern int __kmpc_get_target_offload();
4543 
4544 // Constants used in libomptarget
4545 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4546 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4547 
4548 // OMP Pause Resource
4549 
4550 // The following enum is used both to set the status in __kmp_pause_status, and
4551 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4552 typedef enum kmp_pause_status_t {
4553   kmp_not_paused = 0, // status is not paused, or, requesting resume
4554   kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4555   kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4556 } kmp_pause_status_t;
4557 
4558 // This stores the pause state of the runtime
4559 extern kmp_pause_status_t __kmp_pause_status;
4560 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4561 extern int __kmp_pause_resource(kmp_pause_status_t level);
4562 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4563 extern void __kmp_resume_if_soft_paused();
4564 // Hard resume simply resets the status to not paused. Library will appear to
4565 // be uninitialized after hard pause. Let OMP constructs trigger required
4566 // initializations.
4567 static inline void __kmp_resume_if_hard_paused() {
4568   if (__kmp_pause_status == kmp_hard_paused) {
4569     __kmp_pause_status = kmp_not_paused;
4570   }
4571 }
4572 
4573 extern void __kmp_omp_display_env(int verbose);
4574 
4575 // 1: it is initializing hidden helper team
4576 extern volatile int __kmp_init_hidden_helper;
4577 // 1: the hidden helper team is done
4578 extern volatile int __kmp_hidden_helper_team_done;
4579 // 1: enable hidden helper task
4580 extern kmp_int32 __kmp_enable_hidden_helper;
4581 // Main thread of hidden helper team
4582 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4583 // Descriptors for the hidden helper threads
4584 extern kmp_info_t **__kmp_hidden_helper_threads;
4585 // Number of hidden helper threads
4586 extern kmp_int32 __kmp_hidden_helper_threads_num;
4587 // Number of hidden helper tasks that have not been executed yet
4588 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4589 
4590 extern void __kmp_hidden_helper_initialize();
4591 extern void __kmp_hidden_helper_threads_initz_routine();
4592 extern void __kmp_do_initialize_hidden_helper_threads();
4593 extern void __kmp_hidden_helper_threads_initz_wait();
4594 extern void __kmp_hidden_helper_initz_release();
4595 extern void __kmp_hidden_helper_threads_deinitz_wait();
4596 extern void __kmp_hidden_helper_threads_deinitz_release();
4597 extern void __kmp_hidden_helper_main_thread_wait();
4598 extern void __kmp_hidden_helper_worker_thread_wait();
4599 extern void __kmp_hidden_helper_worker_thread_signal();
4600 extern void __kmp_hidden_helper_main_thread_release();
4601 
4602 // Check whether a given thread is a hidden helper thread
4603 #define KMP_HIDDEN_HELPER_THREAD(gtid)                                         \
4604   ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4605 
4606 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)                                  \
4607   ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4608 
4609 #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)                                    \
4610   ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4611 
4612 #define KMP_HIDDEN_HELPER_TEAM(team)                                           \
4613   (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4614 
4615 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4616 // main thread, is skipped.
4617 #define KMP_GTID_TO_SHADOW_GTID(gtid)                                          \
4618   ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4619 
4620 // Return the adjusted gtid value by subtracting from gtid the number
4621 // of hidden helper threads. This adjusted value is the gtid the thread would
4622 // have received if there were no hidden helper threads.
4623 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4624   int adjusted_gtid = gtid;
4625   if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4626       gtid - __kmp_hidden_helper_threads_num >= 0) {
4627     adjusted_gtid -= __kmp_hidden_helper_threads_num;
4628   }
4629   return adjusted_gtid;
4630 }
4631 
4632 // Support for error directive
4633 typedef enum kmp_severity_t {
4634   severity_warning = 1,
4635   severity_fatal = 2
4636 } kmp_severity_t;
4637 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4638 
4639 // Support for scope directive
4640 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4641 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4642 
4643 #ifdef __cplusplus
4644 }
4645 #endif
4646 
4647 template <bool C, bool S>
4648 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4649 template <bool C, bool S>
4650 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4651 template <bool C, bool S>
4652 extern void __kmp_atomic_suspend_64(int th_gtid,
4653                                     kmp_atomic_flag_64<C, S> *flag);
4654 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4655 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4656 template <bool C, bool S>
4657 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4658 template <bool C, bool S>
4659 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4660 template <bool C, bool S>
4661 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4662 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4663 #endif
4664 template <bool C, bool S>
4665 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4666 template <bool C, bool S>
4667 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4668 template <bool C, bool S>
4669 extern void __kmp_atomic_resume_64(int target_gtid,
4670                                    kmp_atomic_flag_64<C, S> *flag);
4671 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4672 
4673 template <bool C, bool S>
4674 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4675                            kmp_flag_32<C, S> *flag, int final_spin,
4676                            int *thread_finished,
4677 #if USE_ITT_BUILD
4678                            void *itt_sync_obj,
4679 #endif /* USE_ITT_BUILD */
4680                            kmp_int32 is_constrained);
4681 template <bool C, bool S>
4682 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4683                            kmp_flag_64<C, S> *flag, int final_spin,
4684                            int *thread_finished,
4685 #if USE_ITT_BUILD
4686                            void *itt_sync_obj,
4687 #endif /* USE_ITT_BUILD */
4688                            kmp_int32 is_constrained);
4689 template <bool C, bool S>
4690 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4691                                   kmp_atomic_flag_64<C, S> *flag,
4692                                   int final_spin, int *thread_finished,
4693 #if USE_ITT_BUILD
4694                                   void *itt_sync_obj,
4695 #endif /* USE_ITT_BUILD */
4696                                   kmp_int32 is_constrained);
4697 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4698                                kmp_flag_oncore *flag, int final_spin,
4699                                int *thread_finished,
4700 #if USE_ITT_BUILD
4701                                void *itt_sync_obj,
4702 #endif /* USE_ITT_BUILD */
4703                                kmp_int32 is_constrained);
4704 
4705 extern int __kmp_nesting_mode;
4706 extern int __kmp_nesting_mode_nlevels;
4707 extern int *__kmp_nesting_nth_level;
4708 extern void __kmp_init_nesting_mode();
4709 extern void __kmp_set_nesting_mode_threads();
4710 
4711 /// This class safely opens and closes a C-style FILE* object using RAII
4712 /// semantics. There are also methods which allow using stdout or stderr as
4713 /// the underlying FILE* object. With the implicit conversion operator to
4714 /// FILE*, an object with this type can be used in any function which takes
4715 /// a FILE* object e.g., fprintf().
4716 /// No close method is needed at use sites.
4717 class kmp_safe_raii_file_t {
4718   FILE *f;
4719 
4720   void close() {
4721     if (f && f != stdout && f != stderr) {
4722       fclose(f);
4723       f = nullptr;
4724     }
4725   }
4726 
4727 public:
4728   kmp_safe_raii_file_t() : f(nullptr) {}
4729   kmp_safe_raii_file_t(const char *filename, const char *mode,
4730                        const char *env_var = nullptr)
4731       : f(nullptr) {
4732     open(filename, mode, env_var);
4733   }
4734   ~kmp_safe_raii_file_t() { close(); }
4735 
4736   /// Open filename using mode. This is automatically closed in the destructor.
4737   /// The env_var parameter indicates the environment variable the filename
4738   /// came from if != nullptr.
4739   void open(const char *filename, const char *mode,
4740             const char *env_var = nullptr) {
4741     KMP_ASSERT(!f);
4742     f = fopen(filename, mode);
4743     if (!f) {
4744       int code = errno;
4745       if (env_var) {
4746         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4747                     KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4748       } else {
4749         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4750                     __kmp_msg_null);
4751       }
4752     }
4753   }
4754   /// Instead of erroring out, return non-zero when
4755   /// unsuccessful fopen() for any reason
4756   int try_open(const char *filename, const char *mode) {
4757     KMP_ASSERT(!f);
4758     f = fopen(filename, mode);
4759     if (!f)
4760       return errno;
4761     return 0;
4762   }
4763   /// Set the FILE* object to stdout and output there
4764   /// No open call should happen before this call.
4765   void set_stdout() {
4766     KMP_ASSERT(!f);
4767     f = stdout;
4768   }
4769   /// Set the FILE* object to stderr and output there
4770   /// No open call should happen before this call.
4771   void set_stderr() {
4772     KMP_ASSERT(!f);
4773     f = stderr;
4774   }
4775   operator bool() { return bool(f); }
4776   operator FILE *() { return f; }
4777 };
4778 
4779 template <typename SourceType, typename TargetType,
4780           bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4781           bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4782           bool isSourceSigned = std::is_signed<SourceType>::value,
4783           bool isTargetSigned = std::is_signed<TargetType>::value>
4784 struct kmp_convert {};
4785 
4786 // Both types are signed; Source smaller
4787 template <typename SourceType, typename TargetType>
4788 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4789   static TargetType to(SourceType src) { return (TargetType)src; }
4790 };
4791 // Source equal
4792 template <typename SourceType, typename TargetType>
4793 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4794   static TargetType to(SourceType src) { return src; }
4795 };
4796 // Source bigger
4797 template <typename SourceType, typename TargetType>
4798 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4799   static TargetType to(SourceType src) {
4800     KMP_ASSERT(src <= static_cast<SourceType>(
4801                           (std::numeric_limits<TargetType>::max)()));
4802     KMP_ASSERT(src >= static_cast<SourceType>(
4803                           (std::numeric_limits<TargetType>::min)()));
4804     return (TargetType)src;
4805   }
4806 };
4807 
4808 // Source signed, Target unsigned
4809 // Source smaller
4810 template <typename SourceType, typename TargetType>
4811 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4812   static TargetType to(SourceType src) {
4813     KMP_ASSERT(src >= 0);
4814     return (TargetType)src;
4815   }
4816 };
4817 // Source equal
4818 template <typename SourceType, typename TargetType>
4819 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4820   static TargetType to(SourceType src) {
4821     KMP_ASSERT(src >= 0);
4822     return (TargetType)src;
4823   }
4824 };
4825 // Source bigger
4826 template <typename SourceType, typename TargetType>
4827 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4828   static TargetType to(SourceType src) {
4829     KMP_ASSERT(src >= 0);
4830     KMP_ASSERT(src <= static_cast<SourceType>(
4831                           (std::numeric_limits<TargetType>::max)()));
4832     return (TargetType)src;
4833   }
4834 };
4835 
4836 // Source unsigned, Target signed
4837 // Source smaller
4838 template <typename SourceType, typename TargetType>
4839 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4840   static TargetType to(SourceType src) { return (TargetType)src; }
4841 };
4842 // Source equal
4843 template <typename SourceType, typename TargetType>
4844 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4845   static TargetType to(SourceType src) {
4846     KMP_ASSERT(src <= static_cast<SourceType>(
4847                           (std::numeric_limits<TargetType>::max)()));
4848     return (TargetType)src;
4849   }
4850 };
4851 // Source bigger
4852 template <typename SourceType, typename TargetType>
4853 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4854   static TargetType to(SourceType src) {
4855     KMP_ASSERT(src <= static_cast<SourceType>(
4856                           (std::numeric_limits<TargetType>::max)()));
4857     return (TargetType)src;
4858   }
4859 };
4860 
4861 // Source unsigned, Target unsigned
4862 // Source smaller
4863 template <typename SourceType, typename TargetType>
4864 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4865   static TargetType to(SourceType src) { return (TargetType)src; }
4866 };
4867 // Source equal
4868 template <typename SourceType, typename TargetType>
4869 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4870   static TargetType to(SourceType src) { return src; }
4871 };
4872 // Source bigger
4873 template <typename SourceType, typename TargetType>
4874 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4875   static TargetType to(SourceType src) {
4876     KMP_ASSERT(src <= static_cast<SourceType>(
4877                           (std::numeric_limits<TargetType>::max)()));
4878     return (TargetType)src;
4879   }
4880 };
4881 
4882 template <typename T1, typename T2>
4883 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4884   *dest = kmp_convert<T1, T2>::to(src);
4885 }
4886 
4887 #endif /* KMP_H */
4888