xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp.h (revision 1db9f3b21e39176dd5b67cf8ac378633b172463e)
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 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
107 #include "hwloc.h"
108 #ifndef HWLOC_OBJ_NUMANODE
109 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
110 #endif
111 #ifndef HWLOC_OBJ_PACKAGE
112 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
113 #endif
114 #endif
115 
116 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
117 #include <xmmintrin.h>
118 #endif
119 
120 // The below has to be defined before including "kmp_barrier.h".
121 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
122 #define KMP_INTERNAL_FREE(p) free(p)
123 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
124 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
125 
126 #include "kmp_debug.h"
127 #include "kmp_lock.h"
128 #include "kmp_version.h"
129 #include "kmp_barrier.h"
130 #if USE_DEBUGGER
131 #include "kmp_debugger.h"
132 #endif
133 #include "kmp_i18n.h"
134 
135 #define KMP_HANDLE_SIGNALS ((KMP_OS_UNIX && !KMP_OS_WASI) || KMP_OS_WINDOWS)
136 
137 #include "kmp_wrapper_malloc.h"
138 #if KMP_OS_UNIX
139 #include <unistd.h>
140 #if !defined NSIG && defined _NSIG
141 #define NSIG _NSIG
142 #endif
143 #endif
144 
145 #if KMP_OS_LINUX
146 #pragma weak clock_gettime
147 #endif
148 
149 #if OMPT_SUPPORT
150 #include "ompt-internal.h"
151 #endif
152 
153 #if OMPD_SUPPORT
154 #include "ompd-specific.h"
155 #endif
156 
157 #ifndef UNLIKELY
158 #define UNLIKELY(x) (x)
159 #endif
160 
161 // Affinity format function
162 #include "kmp_str.h"
163 
164 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
165 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
166 // free lists of limited size.
167 #ifndef USE_FAST_MEMORY
168 #define USE_FAST_MEMORY 3
169 #endif
170 
171 #ifndef KMP_NESTED_HOT_TEAMS
172 #define KMP_NESTED_HOT_TEAMS 0
173 #define USE_NESTED_HOT_ARG(x)
174 #else
175 #if KMP_NESTED_HOT_TEAMS
176 #define USE_NESTED_HOT_ARG(x) , x
177 #else
178 #define USE_NESTED_HOT_ARG(x)
179 #endif
180 #endif
181 
182 // Assume using BGET compare_exchange instruction instead of lock by default.
183 #ifndef USE_CMP_XCHG_FOR_BGET
184 #define USE_CMP_XCHG_FOR_BGET 1
185 #endif
186 
187 // Test to see if queuing lock is better than bootstrap lock for bget
188 // #ifndef USE_QUEUING_LOCK_FOR_BGET
189 // #define USE_QUEUING_LOCK_FOR_BGET
190 // #endif
191 
192 #define KMP_NSEC_PER_SEC 1000000000L
193 #define KMP_USEC_PER_SEC 1000000L
194 #define KMP_NSEC_PER_USEC 1000L
195 
196 /*!
197 @ingroup BASIC_TYPES
198 @{
199 */
200 
201 /*!
202 Values for bit flags used in the ident_t to describe the fields.
203 */
204 enum {
205   /*! Use trampoline for internal microtasks */
206   KMP_IDENT_IMB = 0x01,
207   /*! Use c-style ident structure */
208   KMP_IDENT_KMPC = 0x02,
209   /* 0x04 is no longer used */
210   /*! Entry point generated by auto-parallelization */
211   KMP_IDENT_AUTOPAR = 0x08,
212   /*! Compiler generates atomic reduction option for kmpc_reduce* */
213   KMP_IDENT_ATOMIC_REDUCE = 0x10,
214   /*! To mark a 'barrier' directive in user code */
215   KMP_IDENT_BARRIER_EXPL = 0x20,
216   /*! To Mark implicit barriers. */
217   KMP_IDENT_BARRIER_IMPL = 0x0040,
218   KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
219   KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
220   KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
221 
222   KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
223   KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
224 
225   /*! To mark a static loop in OMPT callbacks */
226   KMP_IDENT_WORK_LOOP = 0x200,
227   /*! To mark a sections directive in OMPT callbacks */
228   KMP_IDENT_WORK_SECTIONS = 0x400,
229   /*! To mark a distribute construct in OMPT callbacks */
230   KMP_IDENT_WORK_DISTRIBUTE = 0x800,
231   /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
232       not currently used. If one day we need more bits, then we can use
233       an invalid combination of hints to mean that another, larger field
234       should be used in a different flag. */
235   KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
236   KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
237   KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
238   KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
239   KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
240   KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
241 };
242 
243 /*!
244  * The ident structure that describes a source location.
245  */
246 typedef struct ident {
247   kmp_int32 reserved_1; /**<  might be used in Fortran; see above  */
248   kmp_int32 flags; /**<  also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
249                       identifies this union member  */
250   kmp_int32 reserved_2; /**<  not really used in Fortran any more; see above */
251 #if USE_ITT_BUILD
252 /*  but currently used for storing region-specific ITT */
253 /*  contextual information. */
254 #endif /* USE_ITT_BUILD */
255   kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++  */
256   char const *psource; /**< String describing the source location.
257                        The string is composed of semi-colon separated fields
258                        which describe the source file, the function and a pair
259                        of line numbers that delimit the construct. */
260   // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
261   kmp_int32 get_openmp_version() {
262     return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
263   }
264 } ident_t;
265 /*!
266 @}
267 */
268 
269 // Some forward declarations.
270 typedef union kmp_team kmp_team_t;
271 typedef struct kmp_taskdata kmp_taskdata_t;
272 typedef union kmp_task_team kmp_task_team_t;
273 typedef union kmp_team kmp_team_p;
274 typedef union kmp_info kmp_info_p;
275 typedef union kmp_root kmp_root_p;
276 
277 template <bool C = false, bool S = true> class kmp_flag_32;
278 template <bool C = false, bool S = true> class kmp_flag_64;
279 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
280 class kmp_flag_oncore;
281 
282 #ifdef __cplusplus
283 extern "C" {
284 #endif
285 
286 /* ------------------------------------------------------------------------ */
287 
288 /* Pack two 32-bit signed integers into a 64-bit signed integer */
289 /* ToDo: Fix word ordering for big-endian machines. */
290 #define KMP_PACK_64(HIGH_32, LOW_32)                                           \
291   ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
292 
293 // Generic string manipulation macros. Assume that _x is of type char *
294 #define SKIP_WS(_x)                                                            \
295   {                                                                            \
296     while (*(_x) == ' ' || *(_x) == '\t')                                      \
297       (_x)++;                                                                  \
298   }
299 #define SKIP_DIGITS(_x)                                                        \
300   {                                                                            \
301     while (*(_x) >= '0' && *(_x) <= '9')                                       \
302       (_x)++;                                                                  \
303   }
304 #define SKIP_TOKEN(_x)                                                         \
305   {                                                                            \
306     while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
307            (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_')                     \
308       (_x)++;                                                                  \
309   }
310 #define SKIP_TO(_x, _c)                                                        \
311   {                                                                            \
312     while (*(_x) != '\0' && *(_x) != (_c))                                     \
313       (_x)++;                                                                  \
314   }
315 
316 /* ------------------------------------------------------------------------ */
317 
318 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
319 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
320 
321 /* ------------------------------------------------------------------------ */
322 /* Enumeration types */
323 
324 enum kmp_state_timer {
325   ts_stop,
326   ts_start,
327   ts_pause,
328 
329   ts_last_state
330 };
331 
332 enum dynamic_mode {
333   dynamic_default,
334 #ifdef USE_LOAD_BALANCE
335   dynamic_load_balance,
336 #endif /* USE_LOAD_BALANCE */
337   dynamic_random,
338   dynamic_thread_limit,
339   dynamic_max
340 };
341 
342 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
343  * not include it here */
344 #ifndef KMP_SCHED_TYPE_DEFINED
345 #define KMP_SCHED_TYPE_DEFINED
346 typedef enum kmp_sched {
347   kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
348   // Note: need to adjust __kmp_sch_map global array in case enum is changed
349   kmp_sched_static = 1, // mapped to kmp_sch_static_chunked           (33)
350   kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked          (35)
351   kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked           (36)
352   kmp_sched_auto = 4, // mapped to kmp_sch_auto                     (38)
353   kmp_sched_upper_std = 5, // upper bound for standard schedules
354   kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
355   kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
356 #if KMP_STATIC_STEAL_ENABLED
357   kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
358 #endif
359   kmp_sched_upper,
360   kmp_sched_default = kmp_sched_static, // default scheduling
361   kmp_sched_monotonic = 0x80000000
362 } kmp_sched_t;
363 #endif
364 
365 /*!
366  @ingroup WORK_SHARING
367  * Describes the loop schedule to be used for a parallel for loop.
368  */
369 enum sched_type : kmp_int32 {
370   kmp_sch_lower = 32, /**< lower bound for unordered values */
371   kmp_sch_static_chunked = 33,
372   kmp_sch_static = 34, /**< static unspecialized */
373   kmp_sch_dynamic_chunked = 35,
374   kmp_sch_guided_chunked = 36, /**< guided unspecialized */
375   kmp_sch_runtime = 37,
376   kmp_sch_auto = 38, /**< auto */
377   kmp_sch_trapezoidal = 39,
378 
379   /* accessible only through KMP_SCHEDULE environment variable */
380   kmp_sch_static_greedy = 40,
381   kmp_sch_static_balanced = 41,
382   /* accessible only through KMP_SCHEDULE environment variable */
383   kmp_sch_guided_iterative_chunked = 42,
384   kmp_sch_guided_analytical_chunked = 43,
385   /* accessible only through KMP_SCHEDULE environment variable */
386   kmp_sch_static_steal = 44,
387 
388   /* static with chunk adjustment (e.g., simd) */
389   kmp_sch_static_balanced_chunked = 45,
390   kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
391   kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
392 
393   /* accessible only through KMP_SCHEDULE environment variable */
394   kmp_sch_upper, /**< upper bound for unordered values */
395 
396   kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
397   kmp_ord_static_chunked = 65,
398   kmp_ord_static = 66, /**< ordered static unspecialized */
399   kmp_ord_dynamic_chunked = 67,
400   kmp_ord_guided_chunked = 68,
401   kmp_ord_runtime = 69,
402   kmp_ord_auto = 70, /**< ordered auto */
403   kmp_ord_trapezoidal = 71,
404   kmp_ord_upper, /**< upper bound for ordered values */
405 
406   /* Schedules for Distribute construct */
407   kmp_distribute_static_chunked = 91, /**< distribute static chunked */
408   kmp_distribute_static = 92, /**< distribute static unspecialized */
409 
410   /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
411      single iteration/chunk, even if the loop is serialized. For the schedule
412      types listed above, the entire iteration vector is returned if the loop is
413      serialized. This doesn't work for gcc/gcomp sections. */
414   kmp_nm_lower = 160, /**< lower bound for nomerge values */
415 
416   kmp_nm_static_chunked =
417       (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
418   kmp_nm_static = 162, /**< static unspecialized */
419   kmp_nm_dynamic_chunked = 163,
420   kmp_nm_guided_chunked = 164, /**< guided unspecialized */
421   kmp_nm_runtime = 165,
422   kmp_nm_auto = 166, /**< auto */
423   kmp_nm_trapezoidal = 167,
424 
425   /* accessible only through KMP_SCHEDULE environment variable */
426   kmp_nm_static_greedy = 168,
427   kmp_nm_static_balanced = 169,
428   /* accessible only through KMP_SCHEDULE environment variable */
429   kmp_nm_guided_iterative_chunked = 170,
430   kmp_nm_guided_analytical_chunked = 171,
431   kmp_nm_static_steal =
432       172, /* accessible only through OMP_SCHEDULE environment variable */
433 
434   kmp_nm_ord_static_chunked = 193,
435   kmp_nm_ord_static = 194, /**< ordered static unspecialized */
436   kmp_nm_ord_dynamic_chunked = 195,
437   kmp_nm_ord_guided_chunked = 196,
438   kmp_nm_ord_runtime = 197,
439   kmp_nm_ord_auto = 198, /**< auto */
440   kmp_nm_ord_trapezoidal = 199,
441   kmp_nm_upper, /**< upper bound for nomerge values */
442 
443   /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
444      we need to distinguish the three possible cases (no modifier, monotonic
445      modifier, nonmonotonic modifier), we need separate bits for each modifier.
446      The absence of monotonic does not imply nonmonotonic, especially since 4.5
447      says that the behaviour of the "no modifier" case is implementation defined
448      in 4.5, but will become "nonmonotonic" in 5.0.
449 
450      Since we're passing a full 32 bit value, we can use a couple of high bits
451      for these flags; out of paranoia we avoid the sign bit.
452 
453      These modifiers can be or-ed into non-static schedules by the compiler to
454      pass the additional information. They will be stripped early in the
455      processing in __kmp_dispatch_init when setting up schedules, so most of the
456      code won't ever see schedules with these bits set.  */
457   kmp_sch_modifier_monotonic =
458       (1 << 29), /**< Set if the monotonic schedule modifier was present */
459   kmp_sch_modifier_nonmonotonic =
460       (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
461 
462 #define SCHEDULE_WITHOUT_MODIFIERS(s)                                          \
463   (enum sched_type)(                                                           \
464       (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
465 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
466 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
467 #define SCHEDULE_HAS_NO_MODIFIERS(s)                                           \
468   (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
469 #define SCHEDULE_GET_MODIFIERS(s)                                              \
470   ((enum sched_type)(                                                          \
471       (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
472 #define SCHEDULE_SET_MODIFIERS(s, m)                                           \
473   (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
474 #define SCHEDULE_NONMONOTONIC 0
475 #define SCHEDULE_MONOTONIC 1
476 
477   kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
478 };
479 
480 // Apply modifiers on internal kind to standard kind
481 static inline void
482 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
483                                enum sched_type internal_kind) {
484   if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
485     *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
486   }
487 }
488 
489 // Apply modifiers on standard kind to internal kind
490 static inline void
491 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
492                                enum sched_type *internal_kind) {
493   if ((int)kind & (int)kmp_sched_monotonic) {
494     *internal_kind = (enum sched_type)((int)*internal_kind |
495                                        (int)kmp_sch_modifier_monotonic);
496   }
497 }
498 
499 // Get standard schedule without modifiers
500 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
501   return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
502 }
503 
504 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
505 typedef union kmp_r_sched {
506   struct {
507     enum sched_type r_sched_type;
508     int chunk;
509   };
510   kmp_int64 sched;
511 } kmp_r_sched_t;
512 
513 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
514 // internal schedule types
515 
516 enum library_type {
517   library_none,
518   library_serial,
519   library_turnaround,
520   library_throughput
521 };
522 
523 #if KMP_OS_LINUX
524 enum clock_function_type {
525   clock_function_gettimeofday,
526   clock_function_clock_gettime
527 };
528 #endif /* KMP_OS_LINUX */
529 
530 #if KMP_MIC_SUPPORTED
531 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
532 #endif
533 
534 /* -- fast reduction stuff ------------------------------------------------ */
535 
536 #undef KMP_FAST_REDUCTION_BARRIER
537 #define KMP_FAST_REDUCTION_BARRIER 1
538 
539 #undef KMP_FAST_REDUCTION_CORE_DUO
540 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
541 #define KMP_FAST_REDUCTION_CORE_DUO 1
542 #endif
543 
544 enum _reduction_method {
545   reduction_method_not_defined = 0,
546   critical_reduce_block = (1 << 8),
547   atomic_reduce_block = (2 << 8),
548   tree_reduce_block = (3 << 8),
549   empty_reduce_block = (4 << 8)
550 };
551 
552 // Description of the packed_reduction_method variable:
553 // The packed_reduction_method variable consists of two enum types variables
554 // that are packed together into 0-th byte and 1-st byte:
555 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
556 // barrier that will be used in fast reduction: bs_plain_barrier or
557 // bs_reduction_barrier
558 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
559 // be used in fast reduction;
560 // Reduction method is of 'enum _reduction_method' type and it's defined the way
561 // so that the bits of 0-th byte are empty, so no need to execute a shift
562 // instruction while packing/unpacking
563 
564 #if KMP_FAST_REDUCTION_BARRIER
565 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
566   ((reduction_method) | (barrier_type))
567 
568 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
569   ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
570 
571 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method)                      \
572   ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
573 #else
574 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
575   (reduction_method)
576 
577 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
578   (packed_reduction_method)
579 
580 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
581 #endif
582 
583 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block)  \
584   ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) ==                       \
585    (which_reduction_block))
586 
587 #if KMP_FAST_REDUCTION_BARRIER
588 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER                               \
589   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
590 
591 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER                                   \
592   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
593 #endif
594 
595 typedef int PACKED_REDUCTION_METHOD_T;
596 
597 /* -- end of fast reduction stuff ----------------------------------------- */
598 
599 #if KMP_OS_WINDOWS
600 #define USE_CBLKDATA
601 #if KMP_MSVC_COMPAT
602 #pragma warning(push)
603 #pragma warning(disable : 271 310)
604 #endif
605 #include <windows.h>
606 #if KMP_MSVC_COMPAT
607 #pragma warning(pop)
608 #endif
609 #endif
610 
611 #if KMP_OS_UNIX
612 #if !KMP_OS_WASI
613 #include <dlfcn.h>
614 #endif
615 #include <pthread.h>
616 #endif
617 
618 enum kmp_hw_t : int {
619   KMP_HW_UNKNOWN = -1,
620   KMP_HW_SOCKET = 0,
621   KMP_HW_PROC_GROUP,
622   KMP_HW_NUMA,
623   KMP_HW_DIE,
624   KMP_HW_LLC,
625   KMP_HW_L3,
626   KMP_HW_TILE,
627   KMP_HW_MODULE,
628   KMP_HW_L2,
629   KMP_HW_L1,
630   KMP_HW_CORE,
631   KMP_HW_THREAD,
632   KMP_HW_LAST
633 };
634 
635 typedef enum kmp_hw_core_type_t {
636   KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
637 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
638   KMP_HW_CORE_TYPE_ATOM = 0x20,
639   KMP_HW_CORE_TYPE_CORE = 0x40,
640   KMP_HW_MAX_NUM_CORE_TYPES = 3,
641 #else
642   KMP_HW_MAX_NUM_CORE_TYPES = 1,
643 #endif
644 } kmp_hw_core_type_t;
645 
646 #define KMP_HW_MAX_NUM_CORE_EFFS 8
647 
648 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type)                                   \
649   KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
650 #define KMP_ASSERT_VALID_HW_TYPE(type)                                         \
651   KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
652 
653 #define KMP_FOREACH_HW_TYPE(type)                                              \
654   for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST;                        \
655        type = (kmp_hw_t)((int)type + 1))
656 
657 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
658 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
659 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
660 
661 /* Only Linux* OS and Windows* OS support thread affinity. */
662 #if KMP_AFFINITY_SUPPORTED
663 
664 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
665 #if KMP_OS_WINDOWS
666 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
667 typedef struct GROUP_AFFINITY {
668   KAFFINITY Mask;
669   WORD Group;
670   WORD Reserved[3];
671 } GROUP_AFFINITY;
672 #endif /* _MSC_VER < 1600 */
673 #if KMP_GROUP_AFFINITY
674 extern int __kmp_num_proc_groups;
675 #else
676 static const int __kmp_num_proc_groups = 1;
677 #endif /* KMP_GROUP_AFFINITY */
678 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
679 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
680 
681 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
682 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
683 
684 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
685 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
686 
687 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
688                                              GROUP_AFFINITY *);
689 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
690 #endif /* KMP_OS_WINDOWS */
691 
692 #if KMP_USE_HWLOC
693 extern hwloc_topology_t __kmp_hwloc_topology;
694 extern int __kmp_hwloc_error;
695 #endif
696 
697 extern size_t __kmp_affin_mask_size;
698 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
699 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
700 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
701 #define KMP_CPU_SET_ITERATE(i, mask)                                           \
702   for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
703 #define KMP_CPU_SET(i, mask) (mask)->set(i)
704 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
705 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
706 #define KMP_CPU_ZERO(mask) (mask)->zero()
707 #define KMP_CPU_ISEMPTY(mask) (mask)->empty()
708 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
709 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
710 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
711 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
712 #define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
713 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
714 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
715 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
716 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
717 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
718 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
719 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
720 #define KMP_CPU_ALLOC_ARRAY(arr, n)                                            \
721   (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
722 #define KMP_CPU_FREE_ARRAY(arr, n)                                             \
723   __kmp_affinity_dispatch->deallocate_mask_array(arr)
724 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
725 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
726 #define __kmp_get_system_affinity(mask, abort_bool)                            \
727   (mask)->get_system_affinity(abort_bool)
728 #define __kmp_set_system_affinity(mask, abort_bool)                            \
729   (mask)->set_system_affinity(abort_bool)
730 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
731 
732 class KMPAffinity {
733 public:
734   class Mask {
735   public:
736     void *operator new(size_t n);
737     void operator delete(void *p);
738     void *operator new[](size_t n);
739     void operator delete[](void *p);
740     virtual ~Mask() {}
741     // Set bit i to 1
742     virtual void set(int i) {}
743     // Return bit i
744     virtual bool is_set(int i) const { return false; }
745     // Set bit i to 0
746     virtual void clear(int i) {}
747     // Zero out entire mask
748     virtual void zero() {}
749     // Check whether mask is empty
750     virtual bool empty() const { return true; }
751     // Copy src into this mask
752     virtual void copy(const Mask *src) {}
753     // this &= rhs
754     virtual void bitwise_and(const Mask *rhs) {}
755     // this |= rhs
756     virtual void bitwise_or(const Mask *rhs) {}
757     // this = ~this
758     virtual void bitwise_not() {}
759     // this == rhs
760     virtual bool is_equal(const Mask *rhs) const { return false; }
761     // API for iterating over an affinity mask
762     // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
763     virtual int begin() const { return 0; }
764     virtual int end() const { return 0; }
765     virtual int next(int previous) const { return 0; }
766 #if KMP_OS_WINDOWS
767     virtual int set_process_affinity(bool abort_on_error) const { return -1; }
768 #endif
769     // Set the system's affinity to this affinity mask's value
770     virtual int set_system_affinity(bool abort_on_error) const { return -1; }
771     // Set this affinity mask to the current system affinity
772     virtual int get_system_affinity(bool abort_on_error) { return -1; }
773     // Only 1 DWORD in the mask should have any procs set.
774     // Return the appropriate index, or -1 for an invalid mask.
775     virtual int get_proc_group() const { return -1; }
776     int get_max_cpu() const {
777       int cpu;
778       int max_cpu = -1;
779       KMP_CPU_SET_ITERATE(cpu, this) {
780         if (cpu > max_cpu)
781           max_cpu = cpu;
782       }
783       return max_cpu;
784     }
785   };
786   void *operator new(size_t n);
787   void operator delete(void *p);
788   // Need virtual destructor
789   virtual ~KMPAffinity() = default;
790   // Determine if affinity is capable
791   virtual void determine_capable(const char *env_var) {}
792   // Bind the current thread to os proc
793   virtual void bind_thread(int proc) {}
794   // Factory functions to allocate/deallocate a mask
795   virtual Mask *allocate_mask() { return nullptr; }
796   virtual void deallocate_mask(Mask *m) {}
797   virtual Mask *allocate_mask_array(int num) { return nullptr; }
798   virtual void deallocate_mask_array(Mask *m) {}
799   virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
800   static void pick_api();
801   static void destroy_api();
802   enum api_type {
803     NATIVE_OS
804 #if KMP_USE_HWLOC
805     ,
806     HWLOC
807 #endif
808   };
809   virtual api_type get_api_type() const {
810     KMP_ASSERT(0);
811     return NATIVE_OS;
812   }
813 
814 private:
815   static bool picked_api;
816 };
817 
818 typedef KMPAffinity::Mask kmp_affin_mask_t;
819 extern KMPAffinity *__kmp_affinity_dispatch;
820 
821 class kmp_affinity_raii_t {
822   kmp_affin_mask_t *mask;
823   bool restored;
824 
825 public:
826   kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
827       : restored(false) {
828     if (KMP_AFFINITY_CAPABLE()) {
829       KMP_CPU_ALLOC(mask);
830       KMP_ASSERT(mask != NULL);
831       __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
832       if (new_mask)
833         __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
834     }
835   }
836   void restore() {
837     if (!restored && KMP_AFFINITY_CAPABLE()) {
838       __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
839       KMP_CPU_FREE(mask);
840     }
841     restored = true;
842   }
843   ~kmp_affinity_raii_t() { restore(); }
844 };
845 
846 // Declare local char buffers with this size for printing debug and info
847 // messages, using __kmp_affinity_print_mask().
848 #define KMP_AFFIN_MASK_PRINT_LEN 1024
849 
850 enum affinity_type {
851   affinity_none = 0,
852   affinity_physical,
853   affinity_logical,
854   affinity_compact,
855   affinity_scatter,
856   affinity_explicit,
857   affinity_balanced,
858   affinity_disabled, // not used outsize the env var parser
859   affinity_default
860 };
861 
862 enum affinity_top_method {
863   affinity_top_method_all = 0, // try all (supported) methods, in order
864 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
865   affinity_top_method_apicid,
866   affinity_top_method_x2apicid,
867   affinity_top_method_x2apicid_1f,
868 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
869   affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
870 #if KMP_GROUP_AFFINITY
871   affinity_top_method_group,
872 #endif /* KMP_GROUP_AFFINITY */
873   affinity_top_method_flat,
874 #if KMP_USE_HWLOC
875   affinity_top_method_hwloc,
876 #endif
877   affinity_top_method_default
878 };
879 
880 #define affinity_respect_mask_default (2)
881 
882 typedef struct kmp_affinity_flags_t {
883   unsigned dups : 1;
884   unsigned verbose : 1;
885   unsigned warnings : 1;
886   unsigned respect : 2;
887   unsigned reset : 1;
888   unsigned initialized : 1;
889   unsigned core_types_gran : 1;
890   unsigned core_effs_gran : 1;
891   unsigned omp_places : 1;
892   unsigned reserved : 22;
893 } kmp_affinity_flags_t;
894 KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
895 
896 typedef struct kmp_affinity_ids_t {
897   int os_id;
898   int ids[KMP_HW_LAST];
899 } kmp_affinity_ids_t;
900 
901 typedef struct kmp_affinity_attrs_t {
902   int core_type : 8;
903   int core_eff : 8;
904   unsigned valid : 1;
905   unsigned reserved : 15;
906 } kmp_affinity_attrs_t;
907 #define KMP_AFFINITY_ATTRS_UNKNOWN                                             \
908   { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
909 
910 typedef struct kmp_affinity_t {
911   char *proclist;
912   enum affinity_type type;
913   kmp_hw_t gran;
914   int gran_levels;
915   kmp_affinity_attrs_t core_attr_gran;
916   int compact;
917   int offset;
918   kmp_affinity_flags_t flags;
919   unsigned num_masks;
920   kmp_affin_mask_t *masks;
921   kmp_affinity_ids_t *ids;
922   kmp_affinity_attrs_t *attrs;
923   unsigned num_os_id_masks;
924   kmp_affin_mask_t *os_id_masks;
925   const char *env_var;
926 } kmp_affinity_t;
927 
928 #define KMP_AFFINITY_INIT(env)                                                 \
929   {                                                                            \
930     nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
931         0, 0,                                                                  \
932         {TRUE,  FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE,      \
933          FALSE, FALSE, FALSE},                                                 \
934         0, nullptr, nullptr, nullptr, 0, nullptr, env                          \
935   }
936 
937 extern enum affinity_top_method __kmp_affinity_top_method;
938 extern kmp_affinity_t __kmp_affinity;
939 extern kmp_affinity_t __kmp_hh_affinity;
940 extern kmp_affinity_t *__kmp_affinities[2];
941 
942 extern void __kmp_affinity_bind_thread(int which);
943 
944 extern kmp_affin_mask_t *__kmp_affin_fullMask;
945 extern kmp_affin_mask_t *__kmp_affin_origMask;
946 extern char *__kmp_cpuinfo_file;
947 
948 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
949 extern int __kmp_first_osid_with_ecore;
950 #endif
951 
952 #endif /* KMP_AFFINITY_SUPPORTED */
953 
954 // This needs to be kept in sync with the values in omp.h !!!
955 typedef enum kmp_proc_bind_t {
956   proc_bind_false = 0,
957   proc_bind_true,
958   proc_bind_primary,
959   proc_bind_close,
960   proc_bind_spread,
961   proc_bind_intel, // use KMP_AFFINITY interface
962   proc_bind_default
963 } kmp_proc_bind_t;
964 
965 typedef struct kmp_nested_proc_bind_t {
966   kmp_proc_bind_t *bind_types;
967   int size;
968   int used;
969 } kmp_nested_proc_bind_t;
970 
971 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
972 extern kmp_proc_bind_t __kmp_teams_proc_bind;
973 
974 extern int __kmp_display_affinity;
975 extern char *__kmp_affinity_format;
976 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
977 #if OMPT_SUPPORT
978 extern int __kmp_tool;
979 extern char *__kmp_tool_libraries;
980 #endif // OMPT_SUPPORT
981 
982 #if KMP_AFFINITY_SUPPORTED
983 #define KMP_PLACE_ALL (-1)
984 #define KMP_PLACE_UNDEFINED (-2)
985 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
986 #define KMP_AFFINITY_NON_PROC_BIND                                             \
987   ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false ||                 \
988     __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) &&                \
989    (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
990 #endif /* KMP_AFFINITY_SUPPORTED */
991 
992 extern int __kmp_affinity_num_places;
993 
994 typedef enum kmp_cancel_kind_t {
995   cancel_noreq = 0,
996   cancel_parallel = 1,
997   cancel_loop = 2,
998   cancel_sections = 3,
999   cancel_taskgroup = 4
1000 } kmp_cancel_kind_t;
1001 
1002 // KMP_HW_SUBSET support:
1003 typedef struct kmp_hws_item {
1004   int num;
1005   int offset;
1006 } kmp_hws_item_t;
1007 
1008 extern kmp_hws_item_t __kmp_hws_socket;
1009 extern kmp_hws_item_t __kmp_hws_die;
1010 extern kmp_hws_item_t __kmp_hws_node;
1011 extern kmp_hws_item_t __kmp_hws_tile;
1012 extern kmp_hws_item_t __kmp_hws_core;
1013 extern kmp_hws_item_t __kmp_hws_proc;
1014 extern int __kmp_hws_requested;
1015 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1016 
1017 /* ------------------------------------------------------------------------ */
1018 
1019 #define KMP_PAD(type, sz)                                                      \
1020   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1021 
1022 // We need to avoid using -1 as a GTID as +1 is added to the gtid
1023 // when storing it in a lock, and the value 0 is reserved.
1024 #define KMP_GTID_DNE (-2) /* Does not exist */
1025 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1026 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1027 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
1028 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1029 
1030 /* OpenMP 5.0 Memory Management support */
1031 
1032 #ifndef __OMP_H
1033 // Duplicate type definitions from omp.h
1034 typedef uintptr_t omp_uintptr_t;
1035 
1036 typedef enum {
1037   omp_atk_sync_hint = 1,
1038   omp_atk_alignment = 2,
1039   omp_atk_access = 3,
1040   omp_atk_pool_size = 4,
1041   omp_atk_fallback = 5,
1042   omp_atk_fb_data = 6,
1043   omp_atk_pinned = 7,
1044   omp_atk_partition = 8
1045 } omp_alloctrait_key_t;
1046 
1047 typedef enum {
1048   omp_atv_false = 0,
1049   omp_atv_true = 1,
1050   omp_atv_contended = 3,
1051   omp_atv_uncontended = 4,
1052   omp_atv_serialized = 5,
1053   omp_atv_sequential = omp_atv_serialized, // (deprecated)
1054   omp_atv_private = 6,
1055   omp_atv_all = 7,
1056   omp_atv_thread = 8,
1057   omp_atv_pteam = 9,
1058   omp_atv_cgroup = 10,
1059   omp_atv_default_mem_fb = 11,
1060   omp_atv_null_fb = 12,
1061   omp_atv_abort_fb = 13,
1062   omp_atv_allocator_fb = 14,
1063   omp_atv_environment = 15,
1064   omp_atv_nearest = 16,
1065   omp_atv_blocked = 17,
1066   omp_atv_interleaved = 18
1067 } omp_alloctrait_value_t;
1068 #define omp_atv_default ((omp_uintptr_t)-1)
1069 
1070 typedef void *omp_memspace_handle_t;
1071 extern omp_memspace_handle_t const omp_default_mem_space;
1072 extern omp_memspace_handle_t const omp_large_cap_mem_space;
1073 extern omp_memspace_handle_t const omp_const_mem_space;
1074 extern omp_memspace_handle_t const omp_high_bw_mem_space;
1075 extern omp_memspace_handle_t const omp_low_lat_mem_space;
1076 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1077 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1078 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1079 
1080 typedef struct {
1081   omp_alloctrait_key_t key;
1082   omp_uintptr_t value;
1083 } omp_alloctrait_t;
1084 
1085 typedef void *omp_allocator_handle_t;
1086 extern omp_allocator_handle_t const omp_null_allocator;
1087 extern omp_allocator_handle_t const omp_default_mem_alloc;
1088 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1089 extern omp_allocator_handle_t const omp_const_mem_alloc;
1090 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1091 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1092 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1093 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1094 extern omp_allocator_handle_t const omp_thread_mem_alloc;
1095 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1096 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1097 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1098 extern omp_allocator_handle_t const kmp_max_mem_alloc;
1099 extern omp_allocator_handle_t __kmp_def_allocator;
1100 
1101 // end of duplicate type definitions from omp.h
1102 #endif
1103 
1104 extern int __kmp_memkind_available;
1105 
1106 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1107 
1108 typedef struct kmp_allocator_t {
1109   omp_memspace_handle_t memspace;
1110   void **memkind; // pointer to memkind
1111   size_t alignment;
1112   omp_alloctrait_value_t fb;
1113   kmp_allocator_t *fb_data;
1114   kmp_uint64 pool_size;
1115   kmp_uint64 pool_used;
1116   bool pinned;
1117 } kmp_allocator_t;
1118 
1119 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1120                                                     omp_memspace_handle_t,
1121                                                     int ntraits,
1122                                                     omp_alloctrait_t traits[]);
1123 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1124 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1125 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1126 // external interfaces, may be used by compiler
1127 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1128 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1129                                   omp_allocator_handle_t al);
1130 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1131                            omp_allocator_handle_t al);
1132 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1133                             omp_allocator_handle_t al,
1134                             omp_allocator_handle_t free_al);
1135 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1136 // internal interfaces, contain real implementation
1137 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1138                          omp_allocator_handle_t al);
1139 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1140                           omp_allocator_handle_t al);
1141 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1142                            omp_allocator_handle_t al,
1143                            omp_allocator_handle_t free_al);
1144 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1145 
1146 extern void __kmp_init_memkind();
1147 extern void __kmp_fini_memkind();
1148 extern void __kmp_init_target_mem();
1149 
1150 /* ------------------------------------------------------------------------ */
1151 
1152 #if ENABLE_LIBOMPTARGET
1153 extern void __kmp_init_target_task();
1154 #endif
1155 
1156 /* ------------------------------------------------------------------------ */
1157 
1158 #define KMP_UINT64_MAX                                                         \
1159   (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1160 
1161 #define KMP_MIN_NTH 1
1162 
1163 #ifndef KMP_MAX_NTH
1164 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1165 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1166 #else
1167 #ifdef __ve__
1168 // VE's pthread supports only up to 64 threads per a VE process.
1169 // Please check p. 14 of following documentation for more details.
1170 // https://sxauroratsubasa.sakura.ne.jp/documents/veos/en/VEOS_high_level_design.pdf
1171 #define KMP_MAX_NTH 64
1172 #else
1173 #define KMP_MAX_NTH INT_MAX
1174 #endif
1175 #endif
1176 #endif /* KMP_MAX_NTH */
1177 
1178 #ifdef PTHREAD_STACK_MIN
1179 #define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1180 #else
1181 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1182 #endif
1183 
1184 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1185 
1186 #if KMP_ARCH_X86
1187 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1188 #elif KMP_ARCH_X86_64
1189 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1190 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1191 #elif KMP_ARCH_VE
1192 // Minimum stack size for pthread for VE is 4MB.
1193 //   https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1194 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1195 #elif KMP_OS_AIX
1196 // The default stack size for worker threads on AIX is 4MB.
1197 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1198 #else
1199 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1200 #endif
1201 
1202 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1203 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1204 #define KMP_MAX_MALLOC_POOL_INCR                                               \
1205   (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1206 
1207 #define KMP_MIN_STKOFFSET (0)
1208 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1209 #if KMP_OS_DARWIN
1210 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1211 #else
1212 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1213 #endif
1214 
1215 #define KMP_MIN_STKPADDING (0)
1216 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1217 
1218 #define KMP_BLOCKTIME_MULTIPLIER                                               \
1219   (1000000) /* number of blocktime units per second */
1220 #define KMP_MIN_BLOCKTIME (0)
1221 #define KMP_MAX_BLOCKTIME                                                      \
1222   (INT_MAX) /* Must be this for "infinite" setting the work */
1223 
1224 /* __kmp_blocktime is in microseconds */
1225 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1226 
1227 #if KMP_USE_MONITOR
1228 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1229 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1230 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1231 
1232 /* Calculate new number of monitor wakeups for a specific block time based on
1233    previous monitor_wakeups. Only allow increasing number of wakeups */
1234 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups)                 \
1235   (((blocktime) == KMP_MAX_BLOCKTIME)   ? (monitor_wakeups)                    \
1236    : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS              \
1237    : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime)))            \
1238        ? (monitor_wakeups)                                                     \
1239        : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1240 
1241 /* Calculate number of intervals for a specific block time based on
1242    monitor_wakeups */
1243 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups)               \
1244   (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) /        \
1245    (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1246 #else
1247 #define KMP_BLOCKTIME(team, tid)                                               \
1248   (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1249 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1250 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1251 extern kmp_uint64 __kmp_ticks_per_msec;
1252 extern kmp_uint64 __kmp_ticks_per_usec;
1253 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1254 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1255 #else
1256 #define KMP_NOW() __kmp_hardware_timestamp()
1257 #endif
1258 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1259   ((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1260 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1261 #else
1262 // System time is retrieved sporadically while blocking.
1263 extern kmp_uint64 __kmp_now_nsec();
1264 #define KMP_NOW() __kmp_now_nsec()
1265 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1266   ((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1267 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1268 #endif
1269 #endif // KMP_USE_MONITOR
1270 
1271 #define KMP_MIN_STATSCOLS 40
1272 #define KMP_MAX_STATSCOLS 4096
1273 #define KMP_DEFAULT_STATSCOLS 80
1274 
1275 #define KMP_MIN_INTERVAL 0
1276 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1277 #define KMP_DEFAULT_INTERVAL 0
1278 
1279 #define KMP_MIN_CHUNK 1
1280 #define KMP_MAX_CHUNK (INT_MAX - 1)
1281 #define KMP_DEFAULT_CHUNK 1
1282 
1283 #define KMP_MIN_DISP_NUM_BUFF 1
1284 #define KMP_DFLT_DISP_NUM_BUFF 7
1285 #define KMP_MAX_DISP_NUM_BUFF 4096
1286 
1287 #define KMP_MAX_ORDERED 8
1288 
1289 #define KMP_MAX_FIELDS 32
1290 
1291 #define KMP_MAX_BRANCH_BITS 31
1292 
1293 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1294 
1295 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1296 
1297 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1298 
1299 /* Minimum number of threads before switch to TLS gtid (experimentally
1300    determined) */
1301 /* josh TODO: what about OS X* tuning? */
1302 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1303 #define KMP_TLS_GTID_MIN 5
1304 #else
1305 #define KMP_TLS_GTID_MIN INT_MAX
1306 #endif
1307 
1308 #define KMP_MASTER_TID(tid) (0 == (tid))
1309 #define KMP_WORKER_TID(tid) (0 != (tid))
1310 
1311 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1312 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1313 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1314 
1315 #ifndef TRUE
1316 #define FALSE 0
1317 #define TRUE (!FALSE)
1318 #endif
1319 
1320 /* NOTE: all of the following constants must be even */
1321 
1322 #if KMP_OS_WINDOWS
1323 #define KMP_INIT_WAIT 64U /* initial number of spin-tests   */
1324 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1325 #elif KMP_OS_LINUX
1326 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1327 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1328 #elif KMP_OS_DARWIN
1329 /* TODO: tune for KMP_OS_DARWIN */
1330 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1331 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1332 #elif KMP_OS_DRAGONFLY
1333 /* TODO: tune for KMP_OS_DRAGONFLY */
1334 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1335 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1336 #elif KMP_OS_FREEBSD
1337 /* TODO: tune for KMP_OS_FREEBSD */
1338 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1339 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1340 #elif KMP_OS_NETBSD
1341 /* TODO: tune for KMP_OS_NETBSD */
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_OPENBSD
1345 /* TODO: tune for KMP_OS_OPENBSD */
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_HURD
1349 /* TODO: tune for KMP_OS_HURD */
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_SOLARIS
1353 /* TODO: tune for KMP_OS_SOLARIS */
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_WASI
1357 /* TODO: tune for KMP_OS_WASI */
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_AIX
1361 /* TODO: tune for KMP_OS_AIX */
1362 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1363 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1364 #endif
1365 
1366 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1367 typedef struct kmp_cpuid {
1368   kmp_uint32 eax;
1369   kmp_uint32 ebx;
1370   kmp_uint32 ecx;
1371   kmp_uint32 edx;
1372 } kmp_cpuid_t;
1373 
1374 typedef struct kmp_cpuinfo_flags_t {
1375   unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1376   unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1377   unsigned hybrid : 1;
1378   unsigned reserved : 29; // Ensure size of 32 bits
1379 } kmp_cpuinfo_flags_t;
1380 
1381 typedef struct kmp_cpuinfo {
1382   int initialized; // If 0, other fields are not initialized.
1383   int signature; // CPUID(1).EAX
1384   int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1385   int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1386   // Model << 4 ) + Model)
1387   int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1388   kmp_cpuinfo_flags_t flags;
1389   int apic_id;
1390   int physical_id;
1391   int logical_id;
1392   kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1393   char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1394 } kmp_cpuinfo_t;
1395 
1396 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1397 
1398 #if KMP_OS_UNIX
1399 // subleaf is only needed for cache and topology discovery and can be set to
1400 // zero in most cases
1401 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1402   __asm__ __volatile__("cpuid"
1403                        : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1404                        : "a"(leaf), "c"(subleaf));
1405 }
1406 // Load p into FPU control word
1407 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1408   __asm__ __volatile__("fldcw %0" : : "m"(*p));
1409 }
1410 // Store FPU control word into p
1411 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1412   __asm__ __volatile__("fstcw %0" : "=m"(*p));
1413 }
1414 static inline void __kmp_clear_x87_fpu_status_word() {
1415 #if KMP_MIC
1416   // 32-bit protected mode x87 FPU state
1417   struct x87_fpu_state {
1418     unsigned cw;
1419     unsigned sw;
1420     unsigned tw;
1421     unsigned fip;
1422     unsigned fips;
1423     unsigned fdp;
1424     unsigned fds;
1425   };
1426   struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1427   __asm__ __volatile__("fstenv %0\n\t" // store FP env
1428                        "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1429                        "fldenv %0\n\t" // load FP env back
1430                        : "+m"(fpu_state), "+m"(fpu_state.sw));
1431 #else
1432   __asm__ __volatile__("fnclex");
1433 #endif // KMP_MIC
1434 }
1435 #if __SSE__
1436 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1437 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1438 #else
1439 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1440 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1441 #endif
1442 #else
1443 // Windows still has these as external functions in assembly file
1444 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1445 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1446 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1447 extern void __kmp_clear_x87_fpu_status_word();
1448 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1449 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1450 #endif // KMP_OS_UNIX
1451 
1452 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1453 
1454 // User-level Monitor/Mwait
1455 #if KMP_HAVE_UMWAIT
1456 // We always try for UMWAIT first
1457 #if KMP_HAVE_WAITPKG_INTRINSICS
1458 #if KMP_HAVE_IMMINTRIN_H
1459 #include <immintrin.h>
1460 #elif KMP_HAVE_INTRIN_H
1461 #include <intrin.h>
1462 #endif
1463 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1464 
1465 KMP_ATTRIBUTE_TARGET_WAITPKG
1466 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1467 #if !KMP_HAVE_WAITPKG_INTRINSICS
1468   uint32_t timeHi = uint32_t(counter >> 32);
1469   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1470   char flag;
1471   __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1472                    "setb   %0"
1473                    // The "=q" restraint means any register accessible as rl
1474                    //   in 32-bit mode: a, b, c, and d;
1475                    //   in 64-bit mode: any integer register
1476                    : "=q"(flag)
1477                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1478                    :);
1479   return flag;
1480 #else
1481   return _tpause(hint, counter);
1482 #endif
1483 }
1484 KMP_ATTRIBUTE_TARGET_WAITPKG
1485 static inline void __kmp_umonitor(void *cacheline) {
1486 #if !KMP_HAVE_WAITPKG_INTRINSICS
1487   __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1488                    :
1489                    : "a"(cacheline)
1490                    :);
1491 #else
1492   _umonitor(cacheline);
1493 #endif
1494 }
1495 KMP_ATTRIBUTE_TARGET_WAITPKG
1496 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1497 #if !KMP_HAVE_WAITPKG_INTRINSICS
1498   uint32_t timeHi = uint32_t(counter >> 32);
1499   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1500   char flag;
1501   __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1502                    "setb   %0"
1503                    // The "=q" restraint means any register accessible as rl
1504                    //   in 32-bit mode: a, b, c, and d;
1505                    //   in 64-bit mode: any integer register
1506                    : "=q"(flag)
1507                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1508                    :);
1509   return flag;
1510 #else
1511   return _umwait(hint, counter);
1512 #endif
1513 }
1514 #elif KMP_HAVE_MWAIT
1515 #if KMP_OS_UNIX
1516 #include <pmmintrin.h>
1517 #else
1518 #include <intrin.h>
1519 #endif
1520 #if KMP_OS_UNIX
1521 __attribute__((target("sse3")))
1522 #endif
1523 static inline void
1524 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1525   _mm_monitor(cacheline, extensions, hints);
1526 }
1527 #if KMP_OS_UNIX
1528 __attribute__((target("sse3")))
1529 #endif
1530 static inline void
1531 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1532   _mm_mwait(extensions, hints);
1533 }
1534 #endif // KMP_HAVE_UMWAIT
1535 
1536 #if KMP_ARCH_X86
1537 extern void __kmp_x86_pause(void);
1538 #elif KMP_MIC
1539 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1540 // regression after removal of extra PAUSE from spin loops. Changing
1541 // the delay from 100 to 300 showed even better performance than double PAUSE
1542 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1543 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1544 #else
1545 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1546 #endif
1547 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1548 #elif KMP_ARCH_PPC64
1549 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1550 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1551 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1552 #define KMP_CPU_PAUSE()                                                        \
1553   do {                                                                         \
1554     KMP_PPC64_PRI_LOW();                                                       \
1555     KMP_PPC64_PRI_MED();                                                       \
1556     KMP_PPC64_PRI_LOC_MB();                                                    \
1557   } while (0)
1558 #else
1559 #define KMP_CPU_PAUSE() /* nothing to do */
1560 #endif
1561 
1562 #define KMP_INIT_YIELD(count)                                                  \
1563   { (count) = __kmp_yield_init; }
1564 
1565 #define KMP_INIT_BACKOFF(time)                                                 \
1566   { (time) = __kmp_pause_init; }
1567 
1568 #define KMP_OVERSUBSCRIBED                                                     \
1569   (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1570 
1571 #define KMP_TRY_YIELD                                                          \
1572   ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1573 
1574 #define KMP_TRY_YIELD_OVERSUB                                                  \
1575   ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1576 
1577 #define KMP_YIELD(cond)                                                        \
1578   {                                                                            \
1579     KMP_CPU_PAUSE();                                                           \
1580     if ((cond) && (KMP_TRY_YIELD))                                             \
1581       __kmp_yield();                                                           \
1582   }
1583 
1584 #define KMP_YIELD_OVERSUB()                                                    \
1585   {                                                                            \
1586     KMP_CPU_PAUSE();                                                           \
1587     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1588       __kmp_yield();                                                           \
1589   }
1590 
1591 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1592 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1593 #define KMP_YIELD_SPIN(count)                                                  \
1594   {                                                                            \
1595     KMP_CPU_PAUSE();                                                           \
1596     if (KMP_TRY_YIELD) {                                                       \
1597       (count) -= 2;                                                            \
1598       if (!(count)) {                                                          \
1599         __kmp_yield();                                                         \
1600         (count) = __kmp_yield_next;                                            \
1601       }                                                                        \
1602     }                                                                          \
1603   }
1604 
1605 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1606 // (C0.2) state, which improves performance of other SMT threads on the same
1607 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1608 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1609 // available, fall back to the regular CPU pause and yield combination.
1610 #if KMP_HAVE_UMWAIT
1611 #define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1612 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1613   {                                                                            \
1614     if (__kmp_tpause_enabled) {                                                \
1615       if (KMP_OVERSUBSCRIBED) {                                                \
1616         __kmp_tpause(0, (time));                                               \
1617       } else {                                                                 \
1618         __kmp_tpause(__kmp_tpause_hint, (time));                               \
1619       }                                                                        \
1620       (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK;                          \
1621     } else {                                                                   \
1622       KMP_CPU_PAUSE();                                                         \
1623       if ((KMP_TRY_YIELD_OVERSUB)) {                                           \
1624         __kmp_yield();                                                         \
1625       } else if (__kmp_use_yield == 1) {                                       \
1626         (count) -= 2;                                                          \
1627         if (!(count)) {                                                        \
1628           __kmp_yield();                                                       \
1629           (count) = __kmp_yield_next;                                          \
1630         }                                                                      \
1631       }                                                                        \
1632     }                                                                          \
1633   }
1634 #else
1635 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1636   {                                                                            \
1637     KMP_CPU_PAUSE();                                                           \
1638     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1639       __kmp_yield();                                                           \
1640     else if (__kmp_use_yield == 1) {                                           \
1641       (count) -= 2;                                                            \
1642       if (!(count)) {                                                          \
1643         __kmp_yield();                                                         \
1644         (count) = __kmp_yield_next;                                            \
1645       }                                                                        \
1646     }                                                                          \
1647   }
1648 #endif // KMP_HAVE_UMWAIT
1649 
1650 /* ------------------------------------------------------------------------ */
1651 /* Support datatypes for the orphaned construct nesting checks.             */
1652 /* ------------------------------------------------------------------------ */
1653 
1654 /* When adding to this enum, add its corresponding string in cons_text_c[]
1655  * array in kmp_error.cpp */
1656 enum cons_type {
1657   ct_none,
1658   ct_parallel,
1659   ct_pdo,
1660   ct_pdo_ordered,
1661   ct_psections,
1662   ct_psingle,
1663   ct_critical,
1664   ct_ordered_in_parallel,
1665   ct_ordered_in_pdo,
1666   ct_master,
1667   ct_reduce,
1668   ct_barrier,
1669   ct_masked
1670 };
1671 
1672 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1673 
1674 struct cons_data {
1675   ident_t const *ident;
1676   enum cons_type type;
1677   int prev;
1678   kmp_user_lock_p
1679       name; /* address exclusively for critical section name comparison */
1680 };
1681 
1682 struct cons_header {
1683   int p_top, w_top, s_top;
1684   int stack_size, stack_top;
1685   struct cons_data *stack_data;
1686 };
1687 
1688 struct kmp_region_info {
1689   char *text;
1690   int offset[KMP_MAX_FIELDS];
1691   int length[KMP_MAX_FIELDS];
1692 };
1693 
1694 /* ---------------------------------------------------------------------- */
1695 /* ---------------------------------------------------------------------- */
1696 
1697 #if KMP_OS_WINDOWS
1698 typedef HANDLE kmp_thread_t;
1699 typedef DWORD kmp_key_t;
1700 #endif /* KMP_OS_WINDOWS */
1701 
1702 #if KMP_OS_UNIX
1703 typedef pthread_t kmp_thread_t;
1704 typedef pthread_key_t kmp_key_t;
1705 #endif
1706 
1707 extern kmp_key_t __kmp_gtid_threadprivate_key;
1708 
1709 typedef struct kmp_sys_info {
1710   long maxrss; /* the maximum resident set size utilized (in kilobytes)     */
1711   long minflt; /* the number of page faults serviced without any I/O        */
1712   long majflt; /* the number of page faults serviced that required I/O      */
1713   long nswap; /* the number of times a process was "swapped" out of memory */
1714   long inblock; /* the number of times the file system had to perform input  */
1715   long oublock; /* the number of times the file system had to perform output */
1716   long nvcsw; /* the number of times a context switch was voluntarily      */
1717   long nivcsw; /* the number of times a context switch was forced           */
1718 } kmp_sys_info_t;
1719 
1720 #if USE_ITT_BUILD
1721 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1722 // required type here. Later we will check the type meets requirements.
1723 typedef int kmp_itt_mark_t;
1724 #define KMP_ITT_DEBUG 0
1725 #endif /* USE_ITT_BUILD */
1726 
1727 typedef kmp_int32 kmp_critical_name[8];
1728 
1729 /*!
1730 @ingroup PARALLEL
1731 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1732 The arguments to the outlined function are
1733 @param global_tid the global thread identity of the thread executing the
1734 function.
1735 @param bound_tid  the local identity of the thread executing the function
1736 @param ... pointers to shared variables accessed by the function.
1737 */
1738 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1739 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1740                                  ...);
1741 
1742 /*!
1743 @ingroup THREADPRIVATE
1744 @{
1745 */
1746 /* ---------------------------------------------------------------------------
1747  */
1748 /* Threadprivate initialization/finalization function declarations */
1749 
1750 /*  for non-array objects:  __kmpc_threadprivate_register()  */
1751 
1752 /*!
1753  Pointer to the constructor function.
1754  The first argument is the <tt>this</tt> pointer
1755 */
1756 typedef void *(*kmpc_ctor)(void *);
1757 
1758 /*!
1759  Pointer to the destructor function.
1760  The first argument is the <tt>this</tt> pointer
1761 */
1762 typedef void (*kmpc_dtor)(
1763     void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1764                               compiler */
1765 /*!
1766  Pointer to an alternate constructor.
1767  The first argument is the <tt>this</tt> pointer.
1768 */
1769 typedef void *(*kmpc_cctor)(void *, void *);
1770 
1771 /* for array objects: __kmpc_threadprivate_register_vec() */
1772 /* First arg: "this" pointer */
1773 /* Last arg: number of array elements */
1774 /*!
1775  Array constructor.
1776  First argument is the <tt>this</tt> pointer
1777  Second argument the number of array elements.
1778 */
1779 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1780 /*!
1781  Pointer to the array destructor function.
1782  The first argument is the <tt>this</tt> pointer
1783  Second argument the number of array elements.
1784 */
1785 typedef void (*kmpc_dtor_vec)(void *, size_t);
1786 /*!
1787  Array constructor.
1788  First argument is the <tt>this</tt> pointer
1789  Third argument the number of array elements.
1790 */
1791 typedef void *(*kmpc_cctor_vec)(void *, void *,
1792                                 size_t); /* function unused by compiler */
1793 
1794 /*!
1795 @}
1796 */
1797 
1798 /* keeps tracked of threadprivate cache allocations for cleanup later */
1799 typedef struct kmp_cached_addr {
1800   void **addr; /* address of allocated cache */
1801   void ***compiler_cache; /* pointer to compiler's cache */
1802   void *data; /* pointer to global data */
1803   struct kmp_cached_addr *next; /* pointer to next cached address */
1804 } kmp_cached_addr_t;
1805 
1806 struct private_data {
1807   struct private_data *next; /* The next descriptor in the list      */
1808   void *data; /* The data buffer for this descriptor  */
1809   int more; /* The repeat count for this descriptor */
1810   size_t size; /* The data size for this descriptor    */
1811 };
1812 
1813 struct private_common {
1814   struct private_common *next;
1815   struct private_common *link;
1816   void *gbl_addr;
1817   void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1818   size_t cmn_size;
1819 };
1820 
1821 struct shared_common {
1822   struct shared_common *next;
1823   struct private_data *pod_init;
1824   void *obj_init;
1825   void *gbl_addr;
1826   union {
1827     kmpc_ctor ctor;
1828     kmpc_ctor_vec ctorv;
1829   } ct;
1830   union {
1831     kmpc_cctor cctor;
1832     kmpc_cctor_vec cctorv;
1833   } cct;
1834   union {
1835     kmpc_dtor dtor;
1836     kmpc_dtor_vec dtorv;
1837   } dt;
1838   size_t vec_len;
1839   int is_vec;
1840   size_t cmn_size;
1841 };
1842 
1843 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1844 #define KMP_HASH_TABLE_SIZE                                                    \
1845   (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1846 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1847 #define KMP_HASH(x)                                                            \
1848   ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1849 
1850 struct common_table {
1851   struct private_common *data[KMP_HASH_TABLE_SIZE];
1852 };
1853 
1854 struct shared_table {
1855   struct shared_common *data[KMP_HASH_TABLE_SIZE];
1856 };
1857 
1858 /* ------------------------------------------------------------------------ */
1859 
1860 #if KMP_USE_HIER_SCHED
1861 // Shared barrier data that exists inside a single unit of the scheduling
1862 // hierarchy
1863 typedef struct kmp_hier_private_bdata_t {
1864   kmp_int32 num_active;
1865   kmp_uint64 index;
1866   kmp_uint64 wait_val[2];
1867 } kmp_hier_private_bdata_t;
1868 #endif
1869 
1870 typedef struct kmp_sched_flags {
1871   unsigned ordered : 1;
1872   unsigned nomerge : 1;
1873   unsigned contains_last : 1;
1874   unsigned use_hier : 1; // Used in KMP_USE_HIER_SCHED code
1875   unsigned use_hybrid : 1; // Used in KMP_WEIGHTED_ITERATIONS_SUPPORTED code
1876   unsigned unused : 27;
1877 } kmp_sched_flags_t;
1878 
1879 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1880 
1881 #if KMP_STATIC_STEAL_ENABLED
1882 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1883   kmp_int32 count;
1884   kmp_int32 ub;
1885   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1886   kmp_int32 lb;
1887   kmp_int32 st;
1888   kmp_int32 tc;
1889   kmp_lock_t *steal_lock; // lock used for chunk stealing
1890 
1891   kmp_uint32 ordered_lower;
1892   kmp_uint32 ordered_upper;
1893 
1894   // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1895   //    a) parm3 is properly aligned and
1896   //    b) all parm1-4 are on the same cache line.
1897   // Because of parm1-4 are used together, performance seems to be better
1898   // if they are on the same cache line (not measured though).
1899 
1900   struct KMP_ALIGN(32) {
1901     kmp_int32 parm1;
1902     kmp_int32 parm2;
1903     kmp_int32 parm3;
1904     kmp_int32 parm4;
1905   };
1906 
1907 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1908   kmp_uint32 pchunks;
1909   kmp_uint32 num_procs_with_pcore;
1910   kmp_int32 first_thread_with_ecore;
1911 #endif
1912 #if KMP_OS_WINDOWS
1913   kmp_int32 last_upper;
1914 #endif /* KMP_OS_WINDOWS */
1915 } dispatch_private_info32_t;
1916 
1917 #if CACHE_LINE <= 128
1918 KMP_BUILD_ASSERT(sizeof(dispatch_private_info32_t) <= 128);
1919 #endif
1920 
1921 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1922   kmp_int64 count; // current chunk number for static & static-steal scheduling
1923   kmp_int64 ub; /* upper-bound */
1924   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1925   kmp_int64 lb; /* lower-bound */
1926   kmp_int64 st; /* stride */
1927   kmp_int64 tc; /* trip count (number of iterations) */
1928   kmp_lock_t *steal_lock; // lock used for chunk stealing
1929 
1930   kmp_uint64 ordered_lower;
1931   kmp_uint64 ordered_upper;
1932   /* parm[1-4] are used in different ways by different scheduling algorithms */
1933 
1934   // KMP_ALIGN(32) ensures ( if the KMP_ALIGN macro is turned on )
1935   //    a) parm3 is properly aligned and
1936   //    b) all parm1-4 are in the same cache line.
1937   // Because of parm1-4 are used together, performance seems to be better
1938   // if they are in the same line (not measured though).
1939   struct KMP_ALIGN(32) {
1940     kmp_int64 parm1;
1941     kmp_int64 parm2;
1942     kmp_int64 parm3;
1943     kmp_int64 parm4;
1944   };
1945 
1946 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1947   kmp_uint64 pchunks;
1948   kmp_uint64 num_procs_with_pcore;
1949   kmp_int64 first_thread_with_ecore;
1950 #endif
1951 
1952 #if KMP_OS_WINDOWS
1953   kmp_int64 last_upper;
1954 #endif /* KMP_OS_WINDOWS */
1955 } dispatch_private_info64_t;
1956 
1957 #if CACHE_LINE <= 128
1958 KMP_BUILD_ASSERT(sizeof(dispatch_private_info64_t) <= 128);
1959 #endif
1960 
1961 #else /* KMP_STATIC_STEAL_ENABLED */
1962 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1963   kmp_int32 lb;
1964   kmp_int32 ub;
1965   kmp_int32 st;
1966   kmp_int32 tc;
1967 
1968   kmp_int32 parm1;
1969   kmp_int32 parm2;
1970   kmp_int32 parm3;
1971   kmp_int32 parm4;
1972 
1973   kmp_int32 count;
1974 
1975   kmp_uint32 ordered_lower;
1976   kmp_uint32 ordered_upper;
1977 #if KMP_OS_WINDOWS
1978   kmp_int32 last_upper;
1979 #endif /* KMP_OS_WINDOWS */
1980 } dispatch_private_info32_t;
1981 
1982 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1983   kmp_int64 lb; /* lower-bound */
1984   kmp_int64 ub; /* upper-bound */
1985   kmp_int64 st; /* stride */
1986   kmp_int64 tc; /* trip count (number of iterations) */
1987 
1988   /* parm[1-4] are used in different ways by different scheduling algorithms */
1989   kmp_int64 parm1;
1990   kmp_int64 parm2;
1991   kmp_int64 parm3;
1992   kmp_int64 parm4;
1993 
1994   kmp_int64 count; /* current chunk number for static scheduling */
1995 
1996   kmp_uint64 ordered_lower;
1997   kmp_uint64 ordered_upper;
1998 #if KMP_OS_WINDOWS
1999   kmp_int64 last_upper;
2000 #endif /* KMP_OS_WINDOWS */
2001 } dispatch_private_info64_t;
2002 #endif /* KMP_STATIC_STEAL_ENABLED */
2003 
2004 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
2005   union private_info {
2006     dispatch_private_info32_t p32;
2007     dispatch_private_info64_t p64;
2008   } u;
2009   enum sched_type schedule; /* scheduling algorithm */
2010   kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
2011   std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
2012   kmp_int32 ordered_bumped;
2013   // Stack of buffers for nest of serial regions
2014   struct dispatch_private_info *next;
2015   kmp_int32 type_size; /* the size of types in private_info */
2016 #if KMP_USE_HIER_SCHED
2017   kmp_int32 hier_id;
2018   void *parent; /* hierarchical scheduling parent pointer */
2019 #endif
2020   enum cons_type pushed_ws;
2021 } dispatch_private_info_t;
2022 
2023 typedef struct dispatch_shared_info32 {
2024   /* chunk index under dynamic, number of idle threads under static-steal;
2025      iteration index otherwise */
2026   volatile kmp_uint32 iteration;
2027   volatile kmp_int32 num_done;
2028   volatile kmp_uint32 ordered_iteration;
2029   // Dummy to retain the structure size after making ordered_iteration scalar
2030   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
2031 } dispatch_shared_info32_t;
2032 
2033 typedef struct dispatch_shared_info64 {
2034   /* chunk index under dynamic, number of idle threads under static-steal;
2035      iteration index otherwise */
2036   volatile kmp_uint64 iteration;
2037   volatile kmp_int64 num_done;
2038   volatile kmp_uint64 ordered_iteration;
2039   // Dummy to retain the structure size after making ordered_iteration scalar
2040   kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
2041 } dispatch_shared_info64_t;
2042 
2043 typedef struct dispatch_shared_info {
2044   union shared_info {
2045     dispatch_shared_info32_t s32;
2046     dispatch_shared_info64_t s64;
2047   } u;
2048   volatile kmp_uint32 buffer_index;
2049   volatile kmp_int32 doacross_buf_idx; // teamwise index
2050   volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
2051   kmp_int32 doacross_num_done; // count finished threads
2052 #if KMP_USE_HIER_SCHED
2053   void *hier;
2054 #endif
2055 #if KMP_USE_HWLOC
2056   // When linking with libhwloc, the ORDERED EPCC test slows down on big
2057   // machines (> 48 cores). Performance analysis showed that a cache thrash
2058   // was occurring and this padding helps alleviate the problem.
2059   char padding[64];
2060 #endif
2061 } dispatch_shared_info_t;
2062 
2063 typedef struct kmp_disp {
2064   /* Vector for ORDERED SECTION */
2065   void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2066   /* Vector for END ORDERED SECTION */
2067   void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2068 
2069   dispatch_shared_info_t *th_dispatch_sh_current;
2070   dispatch_private_info_t *th_dispatch_pr_current;
2071 
2072   dispatch_private_info_t *th_disp_buffer;
2073   kmp_uint32 th_disp_index;
2074   kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2075   volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2076   kmp_int64 *th_doacross_info; // info on loop bounds
2077 #if KMP_USE_INTERNODE_ALIGNMENT
2078   char more_padding[INTERNODE_CACHE_LINE];
2079 #endif
2080 } kmp_disp_t;
2081 
2082 /* ------------------------------------------------------------------------ */
2083 /* Barrier stuff */
2084 
2085 /* constants for barrier state update */
2086 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2087 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2088 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2089 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2090 
2091 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2092 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2093 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2094 
2095 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2096 #error "Barrier sleep bit must be smaller than barrier bump bit"
2097 #endif
2098 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2099 #error "Barrier unused bit must be smaller than barrier bump bit"
2100 #endif
2101 
2102 // Constants for release barrier wait state: currently, hierarchical only
2103 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2104 #define KMP_BARRIER_OWN_FLAG                                                   \
2105   1 // Normal state; worker waiting on own b_go flag in release
2106 #define KMP_BARRIER_PARENT_FLAG                                                \
2107   2 // Special state; worker waiting on parent's b_go flag in release
2108 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG                                         \
2109   3 // Special state; tells worker to shift from parent to own b_go
2110 #define KMP_BARRIER_SWITCHING                                                  \
2111   4 // Special state; worker resets appropriate flag on wake-up
2112 
2113 #define KMP_NOT_SAFE_TO_REAP                                                   \
2114   0 // Thread th_reap_state: not safe to reap (tasking)
2115 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2116 
2117 // The flag_type describes the storage used for the flag.
2118 enum flag_type {
2119   flag32, /**< atomic 32 bit flags */
2120   flag64, /**< 64 bit flags */
2121   atomic_flag64, /**< atomic 64 bit flags */
2122   flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */
2123   flag_unset
2124 };
2125 
2126 enum barrier_type {
2127   bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2128                            barriers if enabled) */
2129   bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2130 #if KMP_FAST_REDUCTION_BARRIER
2131   bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2132 #endif // KMP_FAST_REDUCTION_BARRIER
2133   bs_last_barrier /* Just a placeholder to mark the end */
2134 };
2135 
2136 // to work with reduction barriers just like with plain barriers
2137 #if !KMP_FAST_REDUCTION_BARRIER
2138 #define bs_reduction_barrier bs_plain_barrier
2139 #endif // KMP_FAST_REDUCTION_BARRIER
2140 
2141 typedef enum kmp_bar_pat { /* Barrier communication patterns */
2142                            bp_linear_bar =
2143                                0, /* Single level (degenerate) tree */
2144                            bp_tree_bar =
2145                                1, /* Balanced tree with branching factor 2^n */
2146                            bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2147                                                 branching factor 2^n */
2148                            bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2149                            bp_dist_bar = 4, /* Distributed barrier */
2150                            bp_last_bar /* Placeholder to mark the end */
2151 } kmp_bar_pat_e;
2152 
2153 #define KMP_BARRIER_ICV_PUSH 1
2154 
2155 /* Record for holding the values of the internal controls stack records */
2156 typedef struct kmp_internal_control {
2157   int serial_nesting_level; /* corresponds to the value of the
2158                                th_team_serialized field */
2159   kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2160                        thread) */
2161   kmp_int8
2162       bt_set; /* internal control for whether blocktime is explicitly set */
2163   int blocktime; /* internal control for blocktime */
2164 #if KMP_USE_MONITOR
2165   int bt_intervals; /* internal control for blocktime intervals */
2166 #endif
2167   int nproc; /* internal control for #threads for next parallel region (per
2168                 thread) */
2169   int thread_limit; /* internal control for thread-limit-var */
2170   int task_thread_limit; /* internal control for thread-limit-var of a task*/
2171   int max_active_levels; /* internal control for max_active_levels */
2172   kmp_r_sched_t
2173       sched; /* internal control for runtime schedule {sched,chunk} pair */
2174   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
2175   kmp_int32 default_device; /* internal control for default device */
2176   struct kmp_internal_control *next;
2177 } kmp_internal_control_t;
2178 
2179 static inline void copy_icvs(kmp_internal_control_t *dst,
2180                              kmp_internal_control_t *src) {
2181   *dst = *src;
2182 }
2183 
2184 /* Thread barrier needs volatile barrier fields */
2185 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2186   // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2187   // uses of it). It is not explicitly aligned below, because we *don't* want
2188   // it to be padded -- instead, we fit b_go into the same cache line with
2189   // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2190   kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2191   // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2192   // same NGO store
2193   volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2194   KMP_ALIGN_CACHE volatile kmp_uint64
2195       b_arrived; // STATE => task reached synch point.
2196   kmp_uint32 *skip_per_level;
2197   kmp_uint32 my_level;
2198   kmp_int32 parent_tid;
2199   kmp_int32 old_tid;
2200   kmp_uint32 depth;
2201   struct kmp_bstate *parent_bar;
2202   kmp_team_t *team;
2203   kmp_uint64 leaf_state;
2204   kmp_uint32 nproc;
2205   kmp_uint8 base_leaf_kids;
2206   kmp_uint8 leaf_kids;
2207   kmp_uint8 offset;
2208   kmp_uint8 wait_flag;
2209   kmp_uint8 use_oncore_barrier;
2210 #if USE_DEBUGGER
2211   // The following field is intended for the debugger solely. Only the worker
2212   // thread itself accesses this field: the worker increases it by 1 when it
2213   // arrives to a barrier.
2214   KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2215 #endif /* USE_DEBUGGER */
2216 } kmp_bstate_t;
2217 
2218 union KMP_ALIGN_CACHE kmp_barrier_union {
2219   double b_align; /* use worst case alignment */
2220   char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2221   kmp_bstate_t bb;
2222 };
2223 
2224 typedef union kmp_barrier_union kmp_balign_t;
2225 
2226 /* Team barrier needs only non-volatile arrived counter */
2227 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2228   double b_align; /* use worst case alignment */
2229   char b_pad[CACHE_LINE];
2230   struct {
2231     kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2232 #if USE_DEBUGGER
2233     // The following two fields are indended for the debugger solely. Only
2234     // primary thread of the team accesses these fields: the first one is
2235     // increased by 1 when the primary thread arrives to a barrier, the second
2236     // one is increased by one when all the threads arrived.
2237     kmp_uint b_master_arrived;
2238     kmp_uint b_team_arrived;
2239 #endif
2240   };
2241 };
2242 
2243 typedef union kmp_barrier_team_union kmp_balign_team_t;
2244 
2245 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2246    threads when a condition changes.  This is to workaround an NPTL bug where
2247    padding was added to pthread_cond_t which caused the initialization routine
2248    to write outside of the structure if compiled on pre-NPTL threads.  */
2249 #if KMP_OS_WINDOWS
2250 typedef struct kmp_win32_mutex {
2251   /* The Lock */
2252   CRITICAL_SECTION cs;
2253 } kmp_win32_mutex_t;
2254 
2255 typedef struct kmp_win32_cond {
2256   /* Count of the number of waiters. */
2257   int waiters_count_;
2258 
2259   /* Serialize access to <waiters_count_> */
2260   kmp_win32_mutex_t waiters_count_lock_;
2261 
2262   /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2263   int release_count_;
2264 
2265   /* Keeps track of the current "generation" so that we don't allow */
2266   /* one thread to steal all the "releases" from the broadcast. */
2267   int wait_generation_count_;
2268 
2269   /* A manual-reset event that's used to block and release waiting threads. */
2270   HANDLE event_;
2271 } kmp_win32_cond_t;
2272 #endif
2273 
2274 #if KMP_OS_UNIX
2275 
2276 union KMP_ALIGN_CACHE kmp_cond_union {
2277   double c_align;
2278   char c_pad[CACHE_LINE];
2279   pthread_cond_t c_cond;
2280 };
2281 
2282 typedef union kmp_cond_union kmp_cond_align_t;
2283 
2284 union KMP_ALIGN_CACHE kmp_mutex_union {
2285   double m_align;
2286   char m_pad[CACHE_LINE];
2287   pthread_mutex_t m_mutex;
2288 };
2289 
2290 typedef union kmp_mutex_union kmp_mutex_align_t;
2291 
2292 #endif /* KMP_OS_UNIX */
2293 
2294 typedef struct kmp_desc_base {
2295   void *ds_stackbase;
2296   size_t ds_stacksize;
2297   int ds_stackgrow;
2298   kmp_thread_t ds_thread;
2299   volatile int ds_tid;
2300   int ds_gtid;
2301 #if KMP_OS_WINDOWS
2302   volatile int ds_alive;
2303   DWORD ds_thread_id;
2304 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2305    However, debugger support (libomp_db) cannot work with handles, because they
2306    uncomparable. For example, debugger requests info about thread with handle h.
2307    h is valid within debugger process, and meaningless within debugee process.
2308    Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2309    within debugee process, but it is a *new* handle which does *not* equal to
2310    any other handle in debugee... The only way to compare handles is convert
2311    them to system-wide ids. GetThreadId() function is available only in
2312    Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2313    on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2314    thread id by call to GetCurrentThreadId() from within the thread and save it
2315    to let libomp_db identify threads.  */
2316 #endif /* KMP_OS_WINDOWS */
2317 } kmp_desc_base_t;
2318 
2319 typedef union KMP_ALIGN_CACHE kmp_desc {
2320   double ds_align; /* use worst case alignment */
2321   char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2322   kmp_desc_base_t ds;
2323 } kmp_desc_t;
2324 
2325 typedef struct kmp_local {
2326   volatile int this_construct; /* count of single's encountered by thread */
2327   void *reduce_data;
2328 #if KMP_USE_BGET
2329   void *bget_data;
2330   void *bget_list;
2331 #if !USE_CMP_XCHG_FOR_BGET
2332 #ifdef USE_QUEUING_LOCK_FOR_BGET
2333   kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2334 #else
2335   kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2336 // bootstrap lock so we can use it at library
2337 // shutdown.
2338 #endif /* USE_LOCK_FOR_BGET */
2339 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2340 #endif /* KMP_USE_BGET */
2341 
2342   PACKED_REDUCTION_METHOD_T
2343   packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2344                               __kmpc_end_reduce*() */
2345 
2346 } kmp_local_t;
2347 
2348 #define KMP_CHECK_UPDATE(a, b)                                                 \
2349   if ((a) != (b))                                                              \
2350   (a) = (b)
2351 #define KMP_CHECK_UPDATE_SYNC(a, b)                                            \
2352   if ((a) != (b))                                                              \
2353   TCW_SYNC_PTR((a), (b))
2354 
2355 #define get__blocktime(xteam, xtid)                                            \
2356   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2357 #define get__bt_set(xteam, xtid)                                               \
2358   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2359 #if KMP_USE_MONITOR
2360 #define get__bt_intervals(xteam, xtid)                                         \
2361   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2362 #endif
2363 
2364 #define get__dynamic_2(xteam, xtid)                                            \
2365   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2366 #define get__nproc_2(xteam, xtid)                                              \
2367   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2368 #define get__sched_2(xteam, xtid)                                              \
2369   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2370 
2371 #define set__blocktime_team(xteam, xtid, xval)                                 \
2372   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) =     \
2373        (xval))
2374 
2375 #if KMP_USE_MONITOR
2376 #define set__bt_intervals_team(xteam, xtid, xval)                              \
2377   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) =  \
2378        (xval))
2379 #endif
2380 
2381 #define set__bt_set_team(xteam, xtid, xval)                                    \
2382   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2383 
2384 #define set__dynamic(xthread, xval)                                            \
2385   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2386 #define get__dynamic(xthread)                                                  \
2387   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2388 
2389 #define set__nproc(xthread, xval)                                              \
2390   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2391 
2392 #define set__thread_limit(xthread, xval)                                       \
2393   (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2394 
2395 #define set__max_active_levels(xthread, xval)                                  \
2396   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2397 
2398 #define get__max_active_levels(xthread)                                        \
2399   ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2400 
2401 #define set__sched(xthread, xval)                                              \
2402   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2403 
2404 #define set__proc_bind(xthread, xval)                                          \
2405   (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2406 #define get__proc_bind(xthread)                                                \
2407   ((xthread)->th.th_current_task->td_icvs.proc_bind)
2408 
2409 // OpenMP tasking data structures
2410 
2411 typedef enum kmp_tasking_mode {
2412   tskm_immediate_exec = 0,
2413   tskm_extra_barrier = 1,
2414   tskm_task_teams = 2,
2415   tskm_max = 2
2416 } kmp_tasking_mode_t;
2417 
2418 extern kmp_tasking_mode_t
2419     __kmp_tasking_mode; /* determines how/when to execute tasks */
2420 extern int __kmp_task_stealing_constraint;
2421 extern int __kmp_enable_task_throttling;
2422 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2423 // specified, defaults to 0 otherwise
2424 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2425 extern kmp_int32 __kmp_max_task_priority;
2426 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2427 extern kmp_uint64 __kmp_taskloop_min_tasks;
2428 
2429 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2430    taskdata first */
2431 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2432 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2433 
2434 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2435 // were spawned and queued since the previous barrier release.
2436 #define KMP_TASKING_ENABLED(task_team)                                         \
2437   (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2438 /*!
2439 @ingroup BASIC_TYPES
2440 @{
2441 */
2442 
2443 /*!
2444  */
2445 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2446 
2447 typedef union kmp_cmplrdata {
2448   kmp_int32 priority; /**< priority specified by user for the task */
2449   kmp_routine_entry_t
2450       destructors; /* pointer to function to invoke deconstructors of
2451                       firstprivate C++ objects */
2452   /* future data */
2453 } kmp_cmplrdata_t;
2454 
2455 /*  sizeof_kmp_task_t passed as arg to kmpc_omp_task call  */
2456 /*!
2457  */
2458 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2459   void *shareds; /**< pointer to block of pointers to shared vars   */
2460   kmp_routine_entry_t
2461       routine; /**< pointer to routine to call for executing task */
2462   kmp_int32 part_id; /**< part id for the task                          */
2463   kmp_cmplrdata_t
2464       data1; /* Two known optional additions: destructors and priority */
2465   kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2466   /* future data */
2467   /*  private vars  */
2468 } kmp_task_t;
2469 
2470 /*!
2471 @}
2472 */
2473 
2474 typedef struct kmp_taskgroup {
2475   std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2476   std::atomic<kmp_int32>
2477       cancel_request; // request for cancellation of this taskgroup
2478   struct kmp_taskgroup *parent; // parent taskgroup
2479   // Block of data to perform task reduction
2480   void *reduce_data; // reduction related info
2481   kmp_int32 reduce_num_data; // number of data items to reduce
2482   uintptr_t *gomp_data; // gomp reduction data
2483 } kmp_taskgroup_t;
2484 
2485 // forward declarations
2486 typedef union kmp_depnode kmp_depnode_t;
2487 typedef struct kmp_depnode_list kmp_depnode_list_t;
2488 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2489 
2490 // macros for checking dep flag as an integer
2491 #define KMP_DEP_IN 0x1
2492 #define KMP_DEP_OUT 0x2
2493 #define KMP_DEP_INOUT 0x3
2494 #define KMP_DEP_MTX 0x4
2495 #define KMP_DEP_SET 0x8
2496 #define KMP_DEP_ALL 0x80
2497 // Compiler sends us this info:
2498 typedef struct kmp_depend_info {
2499   kmp_intptr_t base_addr;
2500   size_t len;
2501   union {
2502     kmp_uint8 flag; // flag as an unsigned char
2503     struct { // flag as a set of 8 bits
2504 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
2505       /* Same fields as in the #else branch, but in reverse order */
2506       unsigned all : 1;
2507       unsigned unused : 3;
2508       unsigned set : 1;
2509       unsigned mtx : 1;
2510       unsigned out : 1;
2511       unsigned in : 1;
2512 #else
2513       unsigned in : 1;
2514       unsigned out : 1;
2515       unsigned mtx : 1;
2516       unsigned set : 1;
2517       unsigned unused : 3;
2518       unsigned all : 1;
2519 #endif
2520     } flags;
2521   };
2522 } kmp_depend_info_t;
2523 
2524 // Internal structures to work with task dependencies:
2525 struct kmp_depnode_list {
2526   kmp_depnode_t *node;
2527   kmp_depnode_list_t *next;
2528 };
2529 
2530 // Max number of mutexinoutset dependencies per node
2531 #define MAX_MTX_DEPS 4
2532 
2533 typedef struct kmp_base_depnode {
2534   kmp_depnode_list_t *successors; /* used under lock */
2535   kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2536   kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2537   kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2538   kmp_lock_t lock; /* guards shared fields: task, successors */
2539 #if KMP_SUPPORT_GRAPH_OUTPUT
2540   kmp_uint32 id;
2541 #endif
2542   std::atomic<kmp_int32> npredecessors;
2543   std::atomic<kmp_int32> nrefs;
2544 } kmp_base_depnode_t;
2545 
2546 union KMP_ALIGN_CACHE kmp_depnode {
2547   double dn_align; /* use worst case alignment */
2548   char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2549   kmp_base_depnode_t dn;
2550 };
2551 
2552 struct kmp_dephash_entry {
2553   kmp_intptr_t addr;
2554   kmp_depnode_t *last_out;
2555   kmp_depnode_list_t *last_set;
2556   kmp_depnode_list_t *prev_set;
2557   kmp_uint8 last_flag;
2558   kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2559   kmp_dephash_entry_t *next_in_bucket;
2560 };
2561 
2562 typedef struct kmp_dephash {
2563   kmp_dephash_entry_t **buckets;
2564   size_t size;
2565   kmp_depnode_t *last_all;
2566   size_t generation;
2567   kmp_uint32 nelements;
2568   kmp_uint32 nconflicts;
2569 } kmp_dephash_t;
2570 
2571 typedef struct kmp_task_affinity_info {
2572   kmp_intptr_t base_addr;
2573   size_t len;
2574   struct {
2575     bool flag1 : 1;
2576     bool flag2 : 1;
2577     kmp_int32 reserved : 30;
2578   } flags;
2579 } kmp_task_affinity_info_t;
2580 
2581 typedef enum kmp_event_type_t {
2582   KMP_EVENT_UNINITIALIZED = 0,
2583   KMP_EVENT_ALLOW_COMPLETION = 1
2584 } kmp_event_type_t;
2585 
2586 typedef struct {
2587   kmp_event_type_t type;
2588   kmp_tas_lock_t lock;
2589   union {
2590     kmp_task_t *task;
2591   } ed;
2592 } kmp_event_t;
2593 
2594 #if OMPX_TASKGRAPH
2595 // Initial number of allocated nodes while recording
2596 #define INIT_MAPSIZE 50
2597 
2598 typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2599   unsigned nowait : 1;
2600   unsigned re_record : 1;
2601   unsigned reserved : 30;
2602 } kmp_taskgraph_flags_t;
2603 
2604 /// Represents a TDG node
2605 typedef struct kmp_node_info {
2606   kmp_task_t *task; // Pointer to the actual task
2607   kmp_int32 *successors; // Array of the succesors ids
2608   kmp_int32 nsuccessors; // Number of succesors of the node
2609   std::atomic<kmp_int32>
2610       npredecessors_counter; // Number of predessors on the fly
2611   kmp_int32 npredecessors; // Total number of predecessors
2612   kmp_int32 successors_size; // Number of allocated succesors ids
2613   kmp_taskdata_t *parent_task; // Parent implicit task
2614 } kmp_node_info_t;
2615 
2616 /// Represent a TDG's current status
2617 typedef enum kmp_tdg_status {
2618   KMP_TDG_NONE = 0,
2619   KMP_TDG_RECORDING = 1,
2620   KMP_TDG_READY = 2
2621 } kmp_tdg_status_t;
2622 
2623 /// Structure that contains a TDG
2624 typedef struct kmp_tdg_info {
2625   kmp_int32 tdg_id; // Unique idenfifier of the TDG
2626   kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2627   kmp_int32 map_size; // Number of allocated TDG nodes
2628   kmp_int32 num_roots; // Number of roots tasks int the TDG
2629   kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2630   kmp_node_info_t *record_map; // Array of TDG nodes
2631   kmp_tdg_status_t tdg_status =
2632       KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2633   std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2634   kmp_bootstrap_lock_t
2635       graph_lock; // Protect graph attributes when updated via taskloop_recur
2636   // Taskloop reduction related
2637   void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2638                           // __kmpc_taskred_init
2639   kmp_int32 rec_num_taskred;
2640 } kmp_tdg_info_t;
2641 
2642 extern int __kmp_tdg_dot;
2643 extern kmp_int32 __kmp_max_tdgs;
2644 extern kmp_tdg_info_t **__kmp_global_tdgs;
2645 extern kmp_int32 __kmp_curr_tdg_idx;
2646 extern kmp_int32 __kmp_successors_size;
2647 extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2648 extern kmp_int32 __kmp_num_tdg;
2649 #endif
2650 
2651 #ifdef BUILD_TIED_TASK_STACK
2652 
2653 /* Tied Task stack definitions */
2654 typedef struct kmp_stack_block {
2655   kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2656   struct kmp_stack_block *sb_next;
2657   struct kmp_stack_block *sb_prev;
2658 } kmp_stack_block_t;
2659 
2660 typedef struct kmp_task_stack {
2661   kmp_stack_block_t ts_first_block; // first block of stack entries
2662   kmp_taskdata_t **ts_top; // pointer to the top of stack
2663   kmp_int32 ts_entries; // number of entries on the stack
2664 } kmp_task_stack_t;
2665 
2666 #endif // BUILD_TIED_TASK_STACK
2667 
2668 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2669 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
2670   /* Same fields as in the #else branch, but in reverse order */
2671 #if OMPX_TASKGRAPH
2672   unsigned reserved31 : 6;
2673   unsigned onced : 1;
2674 #else
2675   unsigned reserved31 : 7;
2676 #endif
2677   unsigned native : 1;
2678   unsigned freed : 1;
2679   unsigned complete : 1;
2680   unsigned executing : 1;
2681   unsigned started : 1;
2682   unsigned team_serial : 1;
2683   unsigned tasking_ser : 1;
2684   unsigned task_serial : 1;
2685   unsigned tasktype : 1;
2686   unsigned reserved : 8;
2687   unsigned hidden_helper : 1;
2688   unsigned detachable : 1;
2689   unsigned priority_specified : 1;
2690   unsigned proxy : 1;
2691   unsigned destructors_thunk : 1;
2692   unsigned merged_if0 : 1;
2693   unsigned final : 1;
2694   unsigned tiedness : 1;
2695 #else
2696   /* Compiler flags */ /* Total compiler flags must be 16 bits */
2697   unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2698   unsigned final : 1; /* task is final(1) so execute immediately */
2699   unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2700                               code path */
2701   unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2702                                      invoke destructors from the runtime */
2703   unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2704                          context of the RTL) */
2705   unsigned priority_specified : 1; /* set if the compiler provides priority
2706                                       setting for the task */
2707   unsigned detachable : 1; /* 1 == can detach */
2708   unsigned hidden_helper : 1; /* 1 == hidden helper task */
2709   unsigned reserved : 8; /* reserved for compiler use */
2710 
2711   /* Library flags */ /* Total library flags must be 16 bits */
2712   unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2713   unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2714   unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2715   // (1) or may be deferred (0)
2716   unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2717   // (0) [>= 2 threads]
2718   /* If either team_serial or tasking_ser is set, task team may be NULL */
2719   /* Task State Flags: */
2720   unsigned started : 1; /* 1==started, 0==not started     */
2721   unsigned executing : 1; /* 1==executing, 0==not executing */
2722   unsigned complete : 1; /* 1==complete, 0==not complete   */
2723   unsigned freed : 1; /* 1==freed, 0==allocated        */
2724   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2725 #if OMPX_TASKGRAPH
2726   unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2727   unsigned reserved31 : 6; /* reserved for library use */
2728 #else
2729   unsigned reserved31 : 7; /* reserved for library use */
2730 #endif
2731 #endif
2732 } kmp_tasking_flags_t;
2733 
2734 typedef struct kmp_target_data {
2735   void *async_handle; // libomptarget async handle for task completion query
2736 } kmp_target_data_t;
2737 
2738 struct kmp_taskdata { /* aligned during dynamic allocation       */
2739   kmp_int32 td_task_id; /* id, assigned by debugger                */
2740   kmp_tasking_flags_t td_flags; /* task flags                              */
2741   kmp_team_t *td_team; /* team for this task                      */
2742   kmp_info_p *td_alloc_thread; /* thread that allocated data structures   */
2743   /* Currently not used except for perhaps IDB */
2744   kmp_taskdata_t *td_parent; /* parent task                             */
2745   kmp_int32 td_level; /* task nesting level                      */
2746   std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2747   ident_t *td_ident; /* task identifier                         */
2748   // Taskwait data.
2749   ident_t *td_taskwait_ident;
2750   kmp_uint32 td_taskwait_counter;
2751   kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2752   KMP_ALIGN_CACHE kmp_internal_control_t
2753       td_icvs; /* Internal control variables for the task */
2754   KMP_ALIGN_CACHE std::atomic<kmp_int32>
2755       td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2756                                    deallocated */
2757   std::atomic<kmp_int32>
2758       td_incomplete_child_tasks; /* Child tasks not yet complete */
2759   kmp_taskgroup_t
2760       *td_taskgroup; // Each task keeps pointer to its current taskgroup
2761   kmp_dephash_t
2762       *td_dephash; // Dependencies for children tasks are tracked from here
2763   kmp_depnode_t
2764       *td_depnode; // Pointer to graph node if this task has dependencies
2765   kmp_task_team_t *td_task_team;
2766   size_t td_size_alloc; // Size of task structure, including shareds etc.
2767 #if defined(KMP_GOMP_COMPAT)
2768   // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2769   kmp_int32 td_size_loop_bounds;
2770 #endif
2771   kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2772 #if defined(KMP_GOMP_COMPAT)
2773   // GOMP sends in a copy function for copy constructors
2774   void (*td_copy_func)(void *, void *);
2775 #endif
2776   kmp_event_t td_allow_completion_event;
2777 #if OMPT_SUPPORT
2778   ompt_task_info_t ompt_task_info;
2779 #endif
2780 #if OMPX_TASKGRAPH
2781   bool is_taskgraph = 0; // whether the task is within a TDG
2782   kmp_tdg_info_t *tdg; // used to associate task with a TDG
2783 #endif
2784   kmp_target_data_t td_target_data;
2785 }; // struct kmp_taskdata
2786 
2787 // Make sure padding above worked
2788 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2789 
2790 // Data for task team but per thread
2791 typedef struct kmp_base_thread_data {
2792   kmp_info_p *td_thr; // Pointer back to thread info
2793   // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2794   // queued?
2795   kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2796   kmp_taskdata_t *
2797       *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2798   kmp_int32 td_deque_size; // Size of deck
2799   kmp_uint32 td_deque_head; // Head of deque (will wrap)
2800   kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2801   kmp_int32 td_deque_ntasks; // Number of tasks in deque
2802   // GEH: shouldn't this be volatile since used in while-spin?
2803   kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2804 #ifdef BUILD_TIED_TASK_STACK
2805   kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2806 // scheduling constraint
2807 #endif // BUILD_TIED_TASK_STACK
2808 } kmp_base_thread_data_t;
2809 
2810 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2811 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2812 
2813 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2814 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2815 
2816 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2817   kmp_base_thread_data_t td;
2818   double td_align; /* use worst case alignment */
2819   char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2820 } kmp_thread_data_t;
2821 
2822 typedef struct kmp_task_pri {
2823   kmp_thread_data_t td;
2824   kmp_int32 priority;
2825   kmp_task_pri *next;
2826 } kmp_task_pri_t;
2827 
2828 // Data for task teams which are used when tasking is enabled for the team
2829 typedef struct kmp_base_task_team {
2830   kmp_bootstrap_lock_t
2831       tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2832   /* must be bootstrap lock since used at library shutdown*/
2833 
2834   // TODO: check performance vs kmp_tas_lock_t
2835   kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2836   kmp_task_pri_t *tt_task_pri_list;
2837 
2838   kmp_task_team_t *tt_next; /* For linking the task team free list */
2839   kmp_thread_data_t
2840       *tt_threads_data; /* Array of per-thread structures for task team */
2841   /* Data survives task team deallocation */
2842   kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2843                                executing this team? */
2844   /* TRUE means tt_threads_data is set up and initialized */
2845   kmp_int32 tt_nproc; /* #threads in team           */
2846   kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2847   kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2848   kmp_int32 tt_untied_task_encountered;
2849   std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2850   // There is hidden helper thread encountered in this task team so that we must
2851   // wait when waiting on task team
2852   kmp_int32 tt_hidden_helper_task_encountered;
2853 
2854   KMP_ALIGN_CACHE
2855   std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2856 
2857   KMP_ALIGN_CACHE
2858   volatile kmp_uint32
2859       tt_active; /* is the team still actively executing tasks */
2860 } kmp_base_task_team_t;
2861 
2862 union KMP_ALIGN_CACHE kmp_task_team {
2863   kmp_base_task_team_t tt;
2864   double tt_align; /* use worst case alignment */
2865   char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2866 };
2867 
2868 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2869 // Free lists keep same-size free memory slots for fast memory allocation
2870 // routines
2871 typedef struct kmp_free_list {
2872   void *th_free_list_self; // Self-allocated tasks free list
2873   void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2874   // threads
2875   void *th_free_list_other; // Non-self free list (to be returned to owner's
2876   // sync list)
2877 } kmp_free_list_t;
2878 #endif
2879 #if KMP_NESTED_HOT_TEAMS
2880 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2881 // are not put in teams pool, and they don't put threads in threads pool.
2882 typedef struct kmp_hot_team_ptr {
2883   kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2884   kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2885 } kmp_hot_team_ptr_t;
2886 #endif
2887 typedef struct kmp_teams_size {
2888   kmp_int32 nteams; // number of teams in a league
2889   kmp_int32 nth; // number of threads in each team of the league
2890 } kmp_teams_size_t;
2891 
2892 // This struct stores a thread that acts as a "root" for a contention
2893 // group. Contention groups are rooted at kmp_root threads, but also at
2894 // each primary thread of each team created in the teams construct.
2895 // This struct therefore also stores a thread_limit associated with
2896 // that contention group, and a counter to track the number of threads
2897 // active in that contention group. Each thread has a list of these: CG
2898 // root threads have an entry in their list in which cg_root refers to
2899 // the thread itself, whereas other workers in the CG will have a
2900 // single entry where cg_root is same as the entry containing their CG
2901 // root. When a thread encounters a teams construct, it will add a new
2902 // entry to the front of its list, because it now roots a new CG.
2903 typedef struct kmp_cg_root {
2904   kmp_info_p *cg_root; // "root" thread for a contention group
2905   // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2906   // thread_limit clause for teams primary threads
2907   kmp_int32 cg_thread_limit;
2908   kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2909   struct kmp_cg_root *up; // pointer to higher level CG root in list
2910 } kmp_cg_root_t;
2911 
2912 // OpenMP thread data structures
2913 
2914 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2915   /* Start with the readonly data which is cache aligned and padded. This is
2916      written before the thread starts working by the primary thread. Uber
2917      masters may update themselves later. Usage does not consider serialized
2918      regions.  */
2919   kmp_desc_t th_info;
2920   kmp_team_p *th_team; /* team we belong to */
2921   kmp_root_p *th_root; /* pointer to root of task hierarchy */
2922   kmp_info_p *th_next_pool; /* next available thread in the pool */
2923   kmp_disp_t *th_dispatch; /* thread's dispatch data */
2924   int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2925 
2926   /* The following are cached from the team info structure */
2927   /* TODO use these in more places as determined to be needed via profiling */
2928   int th_team_nproc; /* number of threads in a team */
2929   kmp_info_p *th_team_master; /* the team's primary thread */
2930   int th_team_serialized; /* team is serialized */
2931   microtask_t th_teams_microtask; /* save entry address for teams construct */
2932   int th_teams_level; /* save initial level of teams construct */
2933 /* it is 0 on device but may be any on host */
2934 
2935 /* The blocktime info is copied from the team struct to the thread struct */
2936 /* at the start of a barrier, and the values stored in the team are used  */
2937 /* at points in the code where the team struct is no longer guaranteed    */
2938 /* to exist (from the POV of worker threads).                             */
2939 #if KMP_USE_MONITOR
2940   int th_team_bt_intervals;
2941   int th_team_bt_set;
2942 #else
2943   kmp_uint64 th_team_bt_intervals;
2944 #endif
2945 
2946 #if KMP_AFFINITY_SUPPORTED
2947   kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2948   kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2949   kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2950 #endif
2951   omp_allocator_handle_t th_def_allocator; /* default allocator */
2952   /* The data set by the primary thread at reinit, then R/W by the worker */
2953   KMP_ALIGN_CACHE int
2954       th_set_nproc; /* if > 0, then only use this request for the next fork */
2955 #if KMP_NESTED_HOT_TEAMS
2956   kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2957 #endif
2958   kmp_proc_bind_t
2959       th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2960   kmp_teams_size_t
2961       th_teams_size; /* number of teams/threads in teams construct */
2962 #if KMP_AFFINITY_SUPPORTED
2963   int th_current_place; /* place currently bound to */
2964   int th_new_place; /* place to bind to in par reg */
2965   int th_first_place; /* first place in partition */
2966   int th_last_place; /* last place in partition */
2967 #endif
2968   int th_prev_level; /* previous level for affinity format */
2969   int th_prev_num_threads; /* previous num_threads for affinity format */
2970 #if USE_ITT_BUILD
2971   kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2972   kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2973   kmp_uint64 th_frame_time; /* frame timestamp */
2974 #endif /* USE_ITT_BUILD */
2975   kmp_local_t th_local;
2976   struct private_common *th_pri_head;
2977 
2978   /* Now the data only used by the worker (after initial allocation) */
2979   /* TODO the first serial team should actually be stored in the info_t
2980      structure.  this will help reduce initial allocation overhead */
2981   KMP_ALIGN_CACHE kmp_team_p
2982       *th_serial_team; /*serialized team held in reserve*/
2983 
2984 #if OMPT_SUPPORT
2985   ompt_thread_info_t ompt_thread_info;
2986 #endif
2987 
2988   /* The following are also read by the primary thread during reinit */
2989   struct common_table *th_pri_common;
2990 
2991   volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2992   /* while awaiting queuing lock acquire */
2993 
2994   volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2995   flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2996 
2997   ident_t *th_ident;
2998   unsigned th_x; // Random number generator data
2999   unsigned th_a; // Random number generator data
3000 
3001   /* Tasking-related data for the thread */
3002   kmp_task_team_t *th_task_team; // Task team struct
3003   kmp_taskdata_t *th_current_task; // Innermost Task being executed
3004   kmp_uint8 th_task_state; // alternating 0/1 for task team identification
3005   kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
3006   // at nested levels
3007   kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
3008   kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
3009   kmp_uint32 th_reap_state; // Non-zero indicates thread is not
3010   // tasking, thus safe to reap
3011 
3012   /* More stuff for keeping track of active/sleeping threads (this part is
3013      written by the worker thread) */
3014   kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
3015   int th_active; // ! sleeping; 32 bits for TCR/TCW
3016   std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
3017   // 0 = not used in team; 1 = used in team;
3018   // 2 = transitioning to not used in team; 3 = transitioning to used in team
3019   struct cons_header *th_cons; // used for consistency check
3020 #if KMP_USE_HIER_SCHED
3021   // used for hierarchical scheduling
3022   kmp_hier_private_bdata_t *th_hier_bar_data;
3023 #endif
3024 
3025   /* Add the syncronizing data which is cache aligned and padded. */
3026   KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
3027 
3028   KMP_ALIGN_CACHE volatile kmp_int32
3029       th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
3030 
3031 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
3032 #define NUM_LISTS 4
3033   kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
3034 // allocation routines
3035 #endif
3036 
3037 #if KMP_OS_WINDOWS
3038   kmp_win32_cond_t th_suspend_cv;
3039   kmp_win32_mutex_t th_suspend_mx;
3040   std::atomic<int> th_suspend_init;
3041 #endif
3042 #if KMP_OS_UNIX
3043   kmp_cond_align_t th_suspend_cv;
3044   kmp_mutex_align_t th_suspend_mx;
3045   std::atomic<int> th_suspend_init_count;
3046 #endif
3047 
3048 #if USE_ITT_BUILD
3049   kmp_itt_mark_t th_itt_mark_single;
3050 // alignment ???
3051 #endif /* USE_ITT_BUILD */
3052 #if KMP_STATS_ENABLED
3053   kmp_stats_list *th_stats;
3054 #endif
3055 #if KMP_OS_UNIX
3056   std::atomic<bool> th_blocking;
3057 #endif
3058   kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
3059 } kmp_base_info_t;
3060 
3061 typedef union KMP_ALIGN_CACHE kmp_info {
3062   double th_align; /* use worst case alignment */
3063   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
3064   kmp_base_info_t th;
3065 } kmp_info_t;
3066 
3067 // OpenMP thread team data structures
3068 
3069 typedef struct kmp_base_data {
3070   volatile kmp_uint32 t_value;
3071 } kmp_base_data_t;
3072 
3073 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
3074   double dt_align; /* use worst case alignment */
3075   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3076   kmp_base_data_t dt;
3077 } kmp_sleep_team_t;
3078 
3079 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
3080   double dt_align; /* use worst case alignment */
3081   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3082   kmp_base_data_t dt;
3083 } kmp_ordered_team_t;
3084 
3085 typedef int (*launch_t)(int gtid);
3086 
3087 /* Minimum number of ARGV entries to malloc if necessary */
3088 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3089 
3090 // Set up how many argv pointers will fit in cache lines containing
3091 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3092 // larger value for more space between the primary write/worker read section and
3093 // read/write by all section seems to buy more performance on EPCC PARALLEL.
3094 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3095 #define KMP_INLINE_ARGV_BYTES                                                  \
3096   (4 * CACHE_LINE -                                                            \
3097    ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) +               \
3098      sizeof(kmp_int16) + sizeof(kmp_uint32)) %                                 \
3099     CACHE_LINE))
3100 #else
3101 #define KMP_INLINE_ARGV_BYTES                                                  \
3102   (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3103 #endif
3104 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3105 
3106 typedef struct KMP_ALIGN_CACHE kmp_base_team {
3107   // Synchronization Data
3108   // ---------------------------------------------------------------------------
3109   KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3110   kmp_balign_team_t t_bar[bs_last_barrier];
3111   std::atomic<int> t_construct; // count of single directive encountered by team
3112   char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3113 
3114   // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3115   std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3116   std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3117 
3118   // Primary thread only
3119   // ---------------------------------------------------------------------------
3120   KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3121   int t_master_this_cons; // "this_construct" single counter of primary thread
3122   // in parent team
3123   ident_t *t_ident; // if volatile, have to change too much other crud to
3124   // volatile too
3125   kmp_team_p *t_parent; // parent team
3126   kmp_team_p *t_next_pool; // next free team in the team pool
3127   kmp_disp_t *t_dispatch; // thread's dispatch data
3128   kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3129   kmp_proc_bind_t t_proc_bind; // bind type for par region
3130 #if USE_ITT_BUILD
3131   kmp_uint64 t_region_time; // region begin timestamp
3132 #endif /* USE_ITT_BUILD */
3133 
3134   // Primary thread write, workers read
3135   // --------------------------------------------------------------------------
3136   KMP_ALIGN_CACHE void **t_argv;
3137   int t_argc;
3138   int t_nproc; // number of threads in team
3139   microtask_t t_pkfn;
3140   launch_t t_invoke; // procedure to launch the microtask
3141 
3142 #if OMPT_SUPPORT
3143   ompt_team_info_t ompt_team_info;
3144   ompt_lw_taskteam_t *ompt_serialized_team_info;
3145 #endif
3146 
3147 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3148   kmp_int8 t_fp_control_saved;
3149   kmp_int8 t_pad2b;
3150   kmp_int16 t_x87_fpu_control_word; // FP control regs
3151   kmp_uint32 t_mxcsr;
3152 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3153 
3154   void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3155 
3156   KMP_ALIGN_CACHE kmp_info_t **t_threads;
3157   kmp_taskdata_t
3158       *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3159   int t_level; // nested parallel level
3160 
3161   KMP_ALIGN_CACHE int t_max_argc;
3162   int t_max_nproc; // max threads this team can handle (dynamically expandable)
3163   int t_serialized; // levels deep of serialized teams
3164   dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3165   int t_id; // team's id, assigned by debugger.
3166   int t_active_level; // nested active parallel level
3167   kmp_r_sched_t t_sched; // run-time schedule for the team
3168 #if KMP_AFFINITY_SUPPORTED
3169   int t_first_place; // first & last place in parent thread's partition.
3170   int t_last_place; // Restore these values to primary thread after par region.
3171 #endif // KMP_AFFINITY_SUPPORTED
3172   int t_display_affinity;
3173   int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3174   // omp_set_num_threads() call
3175   omp_allocator_handle_t t_def_allocator; /* default allocator */
3176 
3177 // Read/write by workers as well
3178 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3179   // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3180   // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3181   // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3182   // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3183   char dummy_padding[1024];
3184 #endif
3185   // Internal control stack for additional nested teams.
3186   KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3187   // for SERIALIZED teams nested 2 or more levels deep
3188   // typed flag to store request state of cancellation
3189   std::atomic<kmp_int32> t_cancel_request;
3190   int t_master_active; // save on fork, restore on join
3191   void *t_copypriv_data; // team specific pointer to copyprivate data array
3192 #if KMP_OS_WINDOWS
3193   std::atomic<kmp_uint32> t_copyin_counter;
3194 #endif
3195 #if USE_ITT_BUILD
3196   void *t_stack_id; // team specific stack stitching id (for ittnotify)
3197 #endif /* USE_ITT_BUILD */
3198   distributedBarrier *b; // Distributed barrier data associated with team
3199 } kmp_base_team_t;
3200 
3201 union KMP_ALIGN_CACHE kmp_team {
3202   kmp_base_team_t t;
3203   double t_align; /* use worst case alignment */
3204   char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3205 };
3206 
3207 typedef union KMP_ALIGN_CACHE kmp_time_global {
3208   double dt_align; /* use worst case alignment */
3209   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3210   kmp_base_data_t dt;
3211 } kmp_time_global_t;
3212 
3213 typedef struct kmp_base_global {
3214   /* cache-aligned */
3215   kmp_time_global_t g_time;
3216 
3217   /* non cache-aligned */
3218   volatile int g_abort;
3219   volatile int g_done;
3220 
3221   int g_dynamic;
3222   enum dynamic_mode g_dynamic_mode;
3223 } kmp_base_global_t;
3224 
3225 typedef union KMP_ALIGN_CACHE kmp_global {
3226   kmp_base_global_t g;
3227   double g_align; /* use worst case alignment */
3228   char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3229 } kmp_global_t;
3230 
3231 typedef struct kmp_base_root {
3232   // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3233   // (r_in_parallel>= 0)
3234   // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3235   // the synch overhead or keeping r_active
3236   volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3237   // keeps a count of active parallel regions per root
3238   std::atomic<int> r_in_parallel;
3239   // GEH: This is misnamed, should be r_active_levels
3240   kmp_team_t *r_root_team;
3241   kmp_team_t *r_hot_team;
3242   kmp_info_t *r_uber_thread;
3243   kmp_lock_t r_begin_lock;
3244   volatile int r_begin;
3245   int r_blocktime; /* blocktime for this root and descendants */
3246 #if KMP_AFFINITY_SUPPORTED
3247   int r_affinity_assigned;
3248 #endif // KMP_AFFINITY_SUPPORTED
3249 } kmp_base_root_t;
3250 
3251 typedef union KMP_ALIGN_CACHE kmp_root {
3252   kmp_base_root_t r;
3253   double r_align; /* use worst case alignment */
3254   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3255 } kmp_root_t;
3256 
3257 struct fortran_inx_info {
3258   kmp_int32 data;
3259 };
3260 
3261 // This list type exists to hold old __kmp_threads arrays so that
3262 // old references to them may complete while reallocation takes place when
3263 // expanding the array. The items in this list are kept alive until library
3264 // shutdown.
3265 typedef struct kmp_old_threads_list_t {
3266   kmp_info_t **threads;
3267   struct kmp_old_threads_list_t *next;
3268 } kmp_old_threads_list_t;
3269 
3270 /* ------------------------------------------------------------------------ */
3271 
3272 extern int __kmp_settings;
3273 extern int __kmp_duplicate_library_ok;
3274 #if USE_ITT_BUILD
3275 extern int __kmp_forkjoin_frames;
3276 extern int __kmp_forkjoin_frames_mode;
3277 #endif
3278 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3279 extern int __kmp_determ_red;
3280 
3281 #ifdef KMP_DEBUG
3282 extern int kmp_a_debug;
3283 extern int kmp_b_debug;
3284 extern int kmp_c_debug;
3285 extern int kmp_d_debug;
3286 extern int kmp_e_debug;
3287 extern int kmp_f_debug;
3288 #endif /* KMP_DEBUG */
3289 
3290 /* For debug information logging using rotating buffer */
3291 #define KMP_DEBUG_BUF_LINES_INIT 512
3292 #define KMP_DEBUG_BUF_LINES_MIN 1
3293 
3294 #define KMP_DEBUG_BUF_CHARS_INIT 128
3295 #define KMP_DEBUG_BUF_CHARS_MIN 2
3296 
3297 extern int
3298     __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3299 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3300 extern int
3301     __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3302 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3303                                       entry pointer */
3304 
3305 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3306 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3307                                               printed in buffer so far */
3308 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3309                                           recommended in warnings */
3310 /* end rotating debug buffer */
3311 
3312 #ifdef KMP_DEBUG
3313 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3314 
3315 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3316 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3317 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3318 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3319 extern int __kmp_par_range_lb;
3320 extern int __kmp_par_range_ub;
3321 #endif
3322 
3323 /* For printing out dynamic storage map for threads and teams */
3324 extern int
3325     __kmp_storage_map; /* True means print storage map for threads and teams */
3326 extern int __kmp_storage_map_verbose; /* True means storage map includes
3327                                          placement info */
3328 extern int __kmp_storage_map_verbose_specified;
3329 
3330 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3331 extern kmp_cpuinfo_t __kmp_cpuinfo;
3332 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3333 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3334 static inline bool __kmp_is_hybrid_cpu() { return true; }
3335 #else
3336 static inline bool __kmp_is_hybrid_cpu() { return false; }
3337 #endif
3338 
3339 extern volatile int __kmp_init_serial;
3340 extern volatile int __kmp_init_gtid;
3341 extern volatile int __kmp_init_common;
3342 extern volatile int __kmp_need_register_serial;
3343 extern volatile int __kmp_init_middle;
3344 extern volatile int __kmp_init_parallel;
3345 #if KMP_USE_MONITOR
3346 extern volatile int __kmp_init_monitor;
3347 #endif
3348 extern volatile int __kmp_init_user_locks;
3349 extern volatile int __kmp_init_hidden_helper_threads;
3350 extern int __kmp_init_counter;
3351 extern int __kmp_root_counter;
3352 extern int __kmp_version;
3353 
3354 /* list of address of allocated caches for commons */
3355 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3356 
3357 /* Barrier algorithm types and options */
3358 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3359 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3360 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3361 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3362 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3363 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3364 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3365 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3366 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3367 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3368 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3369 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3370 
3371 /* Global Locks */
3372 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3373 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3374 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3375 extern kmp_bootstrap_lock_t
3376     __kmp_exit_lock; /* exit() is not always thread-safe */
3377 #if KMP_USE_MONITOR
3378 extern kmp_bootstrap_lock_t
3379     __kmp_monitor_lock; /* control monitor thread creation */
3380 #endif
3381 extern kmp_bootstrap_lock_t
3382     __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3383                              __kmp_threads expansion to co-exist */
3384 
3385 extern kmp_lock_t __kmp_global_lock; /* control OS/global access  */
3386 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access  */
3387 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3388 
3389 extern enum library_type __kmp_library;
3390 
3391 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3392 extern enum sched_type __kmp_static; /* default static scheduling method */
3393 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3394 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3395 extern int __kmp_chunk; /* default runtime chunk size */
3396 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3397 
3398 extern size_t __kmp_stksize; /* stack size per thread         */
3399 #if KMP_USE_MONITOR
3400 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3401 #endif
3402 extern size_t __kmp_stkoffset; /* stack offset per thread       */
3403 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3404 
3405 extern size_t
3406     __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3407 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3408 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3409 extern int __kmp_env_checks; /* was KMP_CHECKS specified?    */
3410 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3411 extern int __kmp_generate_warnings; /* should we issue warnings? */
3412 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3413 
3414 #ifdef DEBUG_SUSPEND
3415 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3416 #endif
3417 
3418 extern kmp_int32 __kmp_use_yield;
3419 extern kmp_int32 __kmp_use_yield_exp_set;
3420 extern kmp_uint32 __kmp_yield_init;
3421 extern kmp_uint32 __kmp_yield_next;
3422 extern kmp_uint64 __kmp_pause_init;
3423 
3424 /* ------------------------------------------------------------------------- */
3425 extern int __kmp_allThreadsSpecified;
3426 
3427 extern size_t __kmp_align_alloc;
3428 /* following data protected by initialization routines */
3429 extern int __kmp_xproc; /* number of processors in the system */
3430 extern int __kmp_avail_proc; /* number of processors available to the process */
3431 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3432 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3433 // maximum total number of concurrently-existing threads on device
3434 extern int __kmp_max_nth;
3435 // maximum total number of concurrently-existing threads in a contention group
3436 extern int __kmp_cg_max_nth;
3437 extern int __kmp_task_max_nth; // max threads used in a task
3438 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3439 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3440                                       __kmp_root */
3441 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3442                                    region a la OMP_NUM_THREADS */
3443 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3444                                       initialization */
3445 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3446                                  used (fixed) */
3447 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3448                                (__kmpc_threadprivate_cached()) */
3449 extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3450                                     blocking (env setting) */
3451 extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3452 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3453 
3454 // Convert raw blocktime from ms to us if needed.
3455 static inline void __kmp_aux_convert_blocktime(int *bt) {
3456   if (__kmp_blocktime_units == 'm') {
3457     if (*bt > INT_MAX / 1000) {
3458       *bt = INT_MAX / 1000;
3459       KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3460     }
3461     *bt = *bt * 1000;
3462   }
3463 }
3464 
3465 #if KMP_USE_MONITOR
3466 extern int
3467     __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3468 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3469                                   blocking */
3470 #endif
3471 #ifdef KMP_ADJUST_BLOCKTIME
3472 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3473 #endif /* KMP_ADJUST_BLOCKTIME */
3474 #ifdef KMP_DFLT_NTH_CORES
3475 extern int __kmp_ncores; /* Total number of cores for threads placement */
3476 #endif
3477 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3478 extern int __kmp_abort_delay;
3479 
3480 extern int __kmp_need_register_atfork_specified;
3481 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3482                                           to install fork handler */
3483 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3484                                0 - not set, will be set at runtime
3485                                1 - using stack search
3486                                2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3487                                    X*) or TlsGetValue(Windows* OS))
3488                                3 - static TLS (__declspec(thread) __kmp_gtid),
3489                                    Linux* OS .so only.  */
3490 extern int
3491     __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3492 #ifdef KMP_TDATA_GTID
3493 extern KMP_THREAD_LOCAL int __kmp_gtid;
3494 #endif
3495 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3496 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3497 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3498 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3499 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3500 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3501 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3502 
3503 // max_active_levels for nested parallelism enabled by default via
3504 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3505 extern int __kmp_dflt_max_active_levels;
3506 // Indicates whether value of __kmp_dflt_max_active_levels was already
3507 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3508 extern bool __kmp_dflt_max_active_levels_set;
3509 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3510                                           concurrent execution per team */
3511 #if KMP_NESTED_HOT_TEAMS
3512 extern int __kmp_hot_teams_mode;
3513 extern int __kmp_hot_teams_max_level;
3514 #endif
3515 
3516 #if KMP_OS_LINUX
3517 extern enum clock_function_type __kmp_clock_function;
3518 extern int __kmp_clock_function_param;
3519 #endif /* KMP_OS_LINUX */
3520 
3521 #if KMP_MIC_SUPPORTED
3522 extern enum mic_type __kmp_mic_type;
3523 #endif
3524 
3525 #ifdef USE_LOAD_BALANCE
3526 extern double __kmp_load_balance_interval; // load balance algorithm interval
3527 #endif /* USE_LOAD_BALANCE */
3528 
3529 // OpenMP 3.1 - Nested num threads array
3530 typedef struct kmp_nested_nthreads_t {
3531   int *nth;
3532   int size;
3533   int used;
3534 } kmp_nested_nthreads_t;
3535 
3536 extern kmp_nested_nthreads_t __kmp_nested_nth;
3537 
3538 #if KMP_USE_ADAPTIVE_LOCKS
3539 
3540 // Parameters for the speculative lock backoff system.
3541 struct kmp_adaptive_backoff_params_t {
3542   // Number of soft retries before it counts as a hard retry.
3543   kmp_uint32 max_soft_retries;
3544   // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3545   // the right
3546   kmp_uint32 max_badness;
3547 };
3548 
3549 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3550 
3551 #if KMP_DEBUG_ADAPTIVE_LOCKS
3552 extern const char *__kmp_speculative_statsfile;
3553 #endif
3554 
3555 #endif // KMP_USE_ADAPTIVE_LOCKS
3556 
3557 extern int __kmp_display_env; /* TRUE or FALSE */
3558 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3559 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3560 extern int __kmp_nteams;
3561 extern int __kmp_teams_thread_limit;
3562 
3563 /* ------------------------------------------------------------------------- */
3564 
3565 /* the following are protected by the fork/join lock */
3566 /* write: lock  read: anytime */
3567 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3568 /* Holds old arrays of __kmp_threads until library shutdown */
3569 extern kmp_old_threads_list_t *__kmp_old_threads_list;
3570 /* read/write: lock */
3571 extern volatile kmp_team_t *__kmp_team_pool;
3572 extern volatile kmp_info_t *__kmp_thread_pool;
3573 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3574 
3575 // total num threads reachable from some root thread including all root threads
3576 extern volatile int __kmp_nth;
3577 /* total number of threads reachable from some root thread including all root
3578    threads, and those in the thread pool */
3579 extern volatile int __kmp_all_nth;
3580 extern std::atomic<int> __kmp_thread_pool_active_nth;
3581 
3582 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3583 /* end data protected by fork/join lock */
3584 /* ------------------------------------------------------------------------- */
3585 
3586 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3587 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3588 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3589 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3590 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3591 
3592 // AT: Which way is correct?
3593 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3594 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3595 #define __kmp_get_team_num_threads(gtid)                                       \
3596   (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3597 
3598 static inline bool KMP_UBER_GTID(int gtid) {
3599   KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3600   KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3601   return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3602           __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3603 }
3604 
3605 static inline int __kmp_tid_from_gtid(int gtid) {
3606   KMP_DEBUG_ASSERT(gtid >= 0);
3607   return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3608 }
3609 
3610 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3611   KMP_DEBUG_ASSERT(tid >= 0 && team);
3612   return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3613 }
3614 
3615 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3616   KMP_DEBUG_ASSERT(thr);
3617   return thr->th.th_info.ds.ds_gtid;
3618 }
3619 
3620 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3621   KMP_DEBUG_ASSERT(gtid >= 0);
3622   return __kmp_threads[gtid];
3623 }
3624 
3625 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3626   KMP_DEBUG_ASSERT(gtid >= 0);
3627   return __kmp_threads[gtid]->th.th_team;
3628 }
3629 
3630 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3631   if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3632     KMP_FATAL(ThreadIdentInvalid);
3633 }
3634 
3635 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3636 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3637 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3638 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3639 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3640 #endif
3641 
3642 #if KMP_HAVE_UMWAIT
3643 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3644 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3645 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3646 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3647 #endif
3648 
3649 /* ------------------------------------------------------------------------- */
3650 
3651 extern kmp_global_t __kmp_global; /* global status */
3652 
3653 extern kmp_info_t __kmp_monitor;
3654 // For Debugging Support Library
3655 extern std::atomic<kmp_int32> __kmp_team_counter;
3656 // For Debugging Support Library
3657 extern std::atomic<kmp_int32> __kmp_task_counter;
3658 
3659 #if USE_DEBUGGER
3660 #define _KMP_GEN_ID(counter)                                                   \
3661   (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3662 #else
3663 #define _KMP_GEN_ID(counter) (~0)
3664 #endif /* USE_DEBUGGER */
3665 
3666 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3667 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3668 
3669 /* ------------------------------------------------------------------------ */
3670 
3671 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3672                                          size_t size, char const *format, ...);
3673 
3674 extern void __kmp_serial_initialize(void);
3675 extern void __kmp_middle_initialize(void);
3676 extern void __kmp_parallel_initialize(void);
3677 
3678 extern void __kmp_internal_begin(void);
3679 extern void __kmp_internal_end_library(int gtid);
3680 extern void __kmp_internal_end_thread(int gtid);
3681 extern void __kmp_internal_end_atexit(void);
3682 extern void __kmp_internal_end_dtor(void);
3683 extern void __kmp_internal_end_dest(void *);
3684 
3685 extern int __kmp_register_root(int initial_thread);
3686 extern void __kmp_unregister_root(int gtid);
3687 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3688 
3689 extern int __kmp_ignore_mppbeg(void);
3690 extern int __kmp_ignore_mppend(void);
3691 
3692 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3693 extern void __kmp_exit_single(int gtid);
3694 
3695 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3696 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3697 
3698 #ifdef USE_LOAD_BALANCE
3699 extern int __kmp_get_load_balance(int);
3700 #endif
3701 
3702 extern int __kmp_get_global_thread_id(void);
3703 extern int __kmp_get_global_thread_id_reg(void);
3704 extern void __kmp_exit_thread(int exit_status);
3705 extern void __kmp_abort(char const *format, ...);
3706 extern void __kmp_abort_thread(void);
3707 KMP_NORETURN extern void __kmp_abort_process(void);
3708 extern void __kmp_warn(char const *format, ...);
3709 
3710 extern void __kmp_set_num_threads(int new_nth, int gtid);
3711 
3712 extern bool __kmp_detect_shm();
3713 extern bool __kmp_detect_tmp();
3714 
3715 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3716 // registered.
3717 static inline kmp_info_t *__kmp_entry_thread() {
3718   int gtid = __kmp_entry_gtid();
3719 
3720   return __kmp_threads[gtid];
3721 }
3722 
3723 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3724 extern int __kmp_get_max_active_levels(int gtid);
3725 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3726 extern int __kmp_get_team_size(int gtid, int level);
3727 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3728 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3729 
3730 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3731 extern void __kmp_init_random(kmp_info_t *thread);
3732 
3733 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3734 extern void __kmp_adjust_num_threads(int new_nproc);
3735 extern void __kmp_check_stksize(size_t *val);
3736 
3737 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3738 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3739 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3740 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3741 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3742 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3743 
3744 #if USE_FAST_MEMORY
3745 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3746                                   size_t size KMP_SRC_LOC_DECL);
3747 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3748 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3749 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3750 #define __kmp_fast_allocate(this_thr, size)                                    \
3751   ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3752 #define __kmp_fast_free(this_thr, ptr)                                         \
3753   ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3754 #endif
3755 
3756 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3757 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3758                                   size_t elsize KMP_SRC_LOC_DECL);
3759 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3760                                    size_t size KMP_SRC_LOC_DECL);
3761 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3762 #define __kmp_thread_malloc(th, size)                                          \
3763   ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3764 #define __kmp_thread_calloc(th, nelem, elsize)                                 \
3765   ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3766 #define __kmp_thread_realloc(th, ptr, size)                                    \
3767   ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3768 #define __kmp_thread_free(th, ptr)                                             \
3769   ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3770 
3771 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3772 
3773 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3774                                  kmp_proc_bind_t proc_bind);
3775 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3776                                  int num_threads);
3777 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3778                                     int num_teams_ub, int num_threads);
3779 
3780 extern void __kmp_yield();
3781 
3782 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3783                                    enum sched_type schedule, kmp_int32 lb,
3784                                    kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3785 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3786                                     enum sched_type schedule, kmp_uint32 lb,
3787                                     kmp_uint32 ub, kmp_int32 st,
3788                                     kmp_int32 chunk);
3789 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3790                                    enum sched_type schedule, kmp_int64 lb,
3791                                    kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3792 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3793                                     enum sched_type schedule, kmp_uint64 lb,
3794                                     kmp_uint64 ub, kmp_int64 st,
3795                                     kmp_int64 chunk);
3796 
3797 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3798                                   kmp_int32 *p_last, kmp_int32 *p_lb,
3799                                   kmp_int32 *p_ub, kmp_int32 *p_st);
3800 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3801                                    kmp_int32 *p_last, kmp_uint32 *p_lb,
3802                                    kmp_uint32 *p_ub, kmp_int32 *p_st);
3803 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3804                                   kmp_int32 *p_last, kmp_int64 *p_lb,
3805                                   kmp_int64 *p_ub, kmp_int64 *p_st);
3806 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3807                                    kmp_int32 *p_last, kmp_uint64 *p_lb,
3808                                    kmp_uint64 *p_ub, kmp_int64 *p_st);
3809 
3810 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3811 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3812 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3813 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3814 
3815 #ifdef KMP_GOMP_COMPAT
3816 
3817 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3818                                       enum sched_type schedule, kmp_int32 lb,
3819                                       kmp_int32 ub, kmp_int32 st,
3820                                       kmp_int32 chunk, int push_ws);
3821 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3822                                        enum sched_type schedule, kmp_uint32 lb,
3823                                        kmp_uint32 ub, kmp_int32 st,
3824                                        kmp_int32 chunk, int push_ws);
3825 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3826                                       enum sched_type schedule, kmp_int64 lb,
3827                                       kmp_int64 ub, kmp_int64 st,
3828                                       kmp_int64 chunk, int push_ws);
3829 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3830                                        enum sched_type schedule, kmp_uint64 lb,
3831                                        kmp_uint64 ub, kmp_int64 st,
3832                                        kmp_int64 chunk, int push_ws);
3833 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3834 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3835 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3836 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3837 
3838 #endif /* KMP_GOMP_COMPAT */
3839 
3840 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3841 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3842 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3843 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3844 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3845 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3846                                kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3847                                void *obj);
3848 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3849                              kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3850 
3851 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3852                           int final_spin
3853 #if USE_ITT_BUILD
3854                           ,
3855                           void *itt_sync_obj
3856 #endif
3857 );
3858 extern void __kmp_release_64(kmp_flag_64<> *flag);
3859 
3860 extern void __kmp_infinite_loop(void);
3861 
3862 extern void __kmp_cleanup(void);
3863 
3864 #if KMP_HANDLE_SIGNALS
3865 extern int __kmp_handle_signals;
3866 extern void __kmp_install_signals(int parallel_init);
3867 extern void __kmp_remove_signals(void);
3868 #endif
3869 
3870 extern void __kmp_clear_system_time(void);
3871 extern void __kmp_read_system_time(double *delta);
3872 
3873 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3874 
3875 extern void __kmp_expand_host_name(char *buffer, size_t size);
3876 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3877 
3878 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3879 extern void
3880 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3881 #endif
3882 
3883 extern void
3884 __kmp_runtime_initialize(void); /* machine specific initialization */
3885 extern void __kmp_runtime_destroy(void);
3886 
3887 #if KMP_AFFINITY_SUPPORTED
3888 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3889                                        kmp_affin_mask_t *mask);
3890 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3891                                                   kmp_affin_mask_t *mask);
3892 extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3893 extern void __kmp_affinity_uninitialize(void);
3894 extern void __kmp_affinity_set_init_mask(
3895     int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3896 void __kmp_affinity_bind_init_mask(int gtid);
3897 extern void __kmp_affinity_bind_place(int gtid);
3898 extern void __kmp_affinity_determine_capable(const char *env_var);
3899 extern int __kmp_aux_set_affinity(void **mask);
3900 extern int __kmp_aux_get_affinity(void **mask);
3901 extern int __kmp_aux_get_affinity_max_proc();
3902 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3903 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3904 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3905 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3906 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
3907 extern int __kmp_get_first_osid_with_ecore(void);
3908 #endif
3909 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3910 extern int kmp_set_thread_affinity_mask_initial(void);
3911 #endif
3912 static inline void __kmp_assign_root_init_mask() {
3913   int gtid = __kmp_entry_gtid();
3914   kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3915   if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3916     __kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3917     __kmp_affinity_bind_init_mask(gtid);
3918     r->r.r_affinity_assigned = TRUE;
3919   }
3920 }
3921 static inline void __kmp_reset_root_init_mask(int gtid) {
3922   if (!KMP_AFFINITY_CAPABLE())
3923     return;
3924   kmp_info_t *th = __kmp_threads[gtid];
3925   kmp_root_t *r = th->th.th_root;
3926   if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3927     __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3928     KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3929     r->r.r_affinity_assigned = FALSE;
3930   }
3931 }
3932 #else /* KMP_AFFINITY_SUPPORTED */
3933 #define __kmp_assign_root_init_mask() /* Nothing */
3934 static inline void __kmp_reset_root_init_mask(int gtid) {}
3935 #endif /* KMP_AFFINITY_SUPPORTED */
3936 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3937 // format string is for affinity, so platforms that do not support
3938 // affinity can still use the other fields, e.g., %n for num_threads
3939 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3940                                          kmp_str_buf_t *buffer);
3941 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3942 
3943 extern void __kmp_cleanup_hierarchy();
3944 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3945 
3946 #if KMP_USE_FUTEX
3947 
3948 extern int __kmp_futex_determine_capable(void);
3949 
3950 #endif // KMP_USE_FUTEX
3951 
3952 extern void __kmp_gtid_set_specific(int gtid);
3953 extern int __kmp_gtid_get_specific(void);
3954 
3955 extern double __kmp_read_cpu_time(void);
3956 
3957 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3958 
3959 #if KMP_USE_MONITOR
3960 extern void __kmp_create_monitor(kmp_info_t *th);
3961 #endif
3962 
3963 extern void *__kmp_launch_thread(kmp_info_t *thr);
3964 
3965 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3966 
3967 #if KMP_OS_WINDOWS
3968 extern int __kmp_still_running(kmp_info_t *th);
3969 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3970 extern void __kmp_free_handle(kmp_thread_t tHandle);
3971 #endif
3972 
3973 #if KMP_USE_MONITOR
3974 extern void __kmp_reap_monitor(kmp_info_t *th);
3975 #endif
3976 extern void __kmp_reap_worker(kmp_info_t *th);
3977 extern void __kmp_terminate_thread(int gtid);
3978 
3979 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3980 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3981 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3982 
3983 extern void __kmp_elapsed(double *);
3984 extern void __kmp_elapsed_tick(double *);
3985 
3986 extern void __kmp_enable(int old_state);
3987 extern void __kmp_disable(int *old_state);
3988 
3989 extern void __kmp_thread_sleep(int millis);
3990 
3991 extern void __kmp_common_initialize(void);
3992 extern void __kmp_common_destroy(void);
3993 extern void __kmp_common_destroy_gtid(int gtid);
3994 
3995 #if KMP_OS_UNIX
3996 extern void __kmp_register_atfork(void);
3997 #endif
3998 extern void __kmp_suspend_initialize(void);
3999 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
4000 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
4001 
4002 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4003                                          int tid);
4004 extern kmp_team_t *
4005 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4006 #if OMPT_SUPPORT
4007                     ompt_data_t ompt_parallel_data,
4008 #endif
4009                     kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
4010                     int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
4011 extern void __kmp_free_thread(kmp_info_t *);
4012 extern void __kmp_free_team(kmp_root_t *,
4013                             kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
4014 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
4015 
4016 /* ------------------------------------------------------------------------ */
4017 
4018 extern void __kmp_initialize_bget(kmp_info_t *th);
4019 extern void __kmp_finalize_bget(kmp_info_t *th);
4020 
4021 KMP_EXPORT void *kmpc_malloc(size_t size);
4022 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
4023 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
4024 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
4025 KMP_EXPORT void kmpc_free(void *ptr);
4026 
4027 /* declarations for internal use */
4028 
4029 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
4030                          size_t reduce_size, void *reduce_data,
4031                          void (*reduce)(void *, void *));
4032 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
4033 extern int __kmp_barrier_gomp_cancel(int gtid);
4034 
4035 /*!
4036  * Tell the fork call which compiler generated the fork call, and therefore how
4037  * to deal with the call.
4038  */
4039 enum fork_context_e {
4040   fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
4041                        microtask internally. */
4042   fork_context_intel, /**< Called from Intel generated code.  */
4043   fork_context_last
4044 };
4045 extern int __kmp_fork_call(ident_t *loc, int gtid,
4046                            enum fork_context_e fork_context, kmp_int32 argc,
4047                            microtask_t microtask, launch_t invoker,
4048                            kmp_va_list ap);
4049 
4050 extern void __kmp_join_call(ident_t *loc, int gtid
4051 #if OMPT_SUPPORT
4052                             ,
4053                             enum fork_context_e fork_context
4054 #endif
4055                             ,
4056                             int exit_teams = 0);
4057 
4058 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
4059 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
4060 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
4061 extern int __kmp_invoke_task_func(int gtid);
4062 extern void __kmp_run_before_invoked_task(int gtid, int tid,
4063                                           kmp_info_t *this_thr,
4064                                           kmp_team_t *team);
4065 extern void __kmp_run_after_invoked_task(int gtid, int tid,
4066                                          kmp_info_t *this_thr,
4067                                          kmp_team_t *team);
4068 
4069 // should never have been exported
4070 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
4071 extern int __kmp_invoke_teams_master(int gtid);
4072 extern void __kmp_teams_master(int gtid);
4073 extern int __kmp_aux_get_team_num();
4074 extern int __kmp_aux_get_num_teams();
4075 extern void __kmp_save_internal_controls(kmp_info_t *thread);
4076 extern void __kmp_user_set_library(enum library_type arg);
4077 extern void __kmp_aux_set_library(enum library_type arg);
4078 extern void __kmp_aux_set_stacksize(size_t arg);
4079 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
4080 extern void __kmp_aux_set_defaults(char const *str, size_t len);
4081 
4082 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
4083 void kmpc_set_blocktime(int arg);
4084 void ompc_set_nested(int flag);
4085 void ompc_set_dynamic(int flag);
4086 void ompc_set_num_threads(int arg);
4087 
4088 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
4089                                               kmp_team_t *team, int tid);
4090 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
4091 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4092                                     kmp_tasking_flags_t *flags,
4093                                     size_t sizeof_kmp_task_t,
4094                                     size_t sizeof_shareds,
4095                                     kmp_routine_entry_t task_entry);
4096 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4097                                      kmp_team_t *team, int tid,
4098                                      int set_curr_task);
4099 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4100 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4101 
4102 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4103                                                        int gtid,
4104                                                        kmp_task_t *task);
4105 extern void __kmp_fulfill_event(kmp_event_t *event);
4106 
4107 extern void __kmp_free_task_team(kmp_info_t *thread,
4108                                  kmp_task_team_t *task_team);
4109 extern void __kmp_reap_task_teams(void);
4110 extern void __kmp_wait_to_unref_task_teams(void);
4111 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
4112                                   int always);
4113 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4114 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4115 #if USE_ITT_BUILD
4116                                  ,
4117                                  void *itt_sync_obj
4118 #endif /* USE_ITT_BUILD */
4119                                  ,
4120                                  int wait = 1);
4121 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4122                                   int gtid);
4123 
4124 extern int __kmp_is_address_mapped(void *addr);
4125 extern kmp_uint64 __kmp_hardware_timestamp(void);
4126 
4127 #if KMP_OS_UNIX
4128 extern int __kmp_read_from_file(char const *path, char const *format, ...);
4129 #endif
4130 
4131 /* ------------------------------------------------------------------------ */
4132 //
4133 // Assembly routines that have no compiler intrinsic replacement
4134 //
4135 
4136 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4137                                   void *argv[]
4138 #if OMPT_SUPPORT
4139                                   ,
4140                                   void **exit_frame_ptr
4141 #endif
4142 );
4143 
4144 /* ------------------------------------------------------------------------ */
4145 
4146 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4147 KMP_EXPORT void __kmpc_end(ident_t *);
4148 
4149 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4150                                                   kmpc_ctor_vec ctor,
4151                                                   kmpc_cctor_vec cctor,
4152                                                   kmpc_dtor_vec dtor,
4153                                                   size_t vector_length);
4154 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4155                                               kmpc_ctor ctor, kmpc_cctor cctor,
4156                                               kmpc_dtor dtor);
4157 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4158                                       void *data, size_t size);
4159 
4160 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4161 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4162 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4163 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4164 
4165 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4166 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4167                                  kmpc_micro microtask, ...);
4168 KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4169                                     kmpc_micro microtask, kmp_int32 cond,
4170                                     void *args);
4171 
4172 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4173 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4174 
4175 KMP_EXPORT void __kmpc_flush(ident_t *);
4176 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4177 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4178 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4179 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4180                                    kmp_int32 filter);
4181 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4182 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4183 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4184 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4185                                 kmp_critical_name *);
4186 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4187                                     kmp_critical_name *);
4188 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4189                                           kmp_critical_name *, uint32_t hint);
4190 
4191 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4192 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4193 
4194 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4195                                                   kmp_int32 global_tid);
4196 
4197 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4198 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4199 
4200 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4201 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4202                                          kmp_int32 numberOfSections);
4203 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4204 
4205 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4206                                      kmp_int32 schedtype, kmp_int32 *plastiter,
4207                                      kmp_int *plower, kmp_int *pupper,
4208                                      kmp_int *pstride, kmp_int incr,
4209                                      kmp_int chunk);
4210 
4211 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4212 
4213 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4214                                    size_t cpy_size, void *cpy_data,
4215                                    void (*cpy_func)(void *, void *),
4216                                    kmp_int32 didit);
4217 
4218 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4219                                           void *cpy_data);
4220 
4221 extern void KMPC_SET_NUM_THREADS(int arg);
4222 extern void KMPC_SET_DYNAMIC(int flag);
4223 extern void KMPC_SET_NESTED(int flag);
4224 
4225 /* OMP 3.0 tasking interface routines */
4226 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4227                                      kmp_task_t *new_task);
4228 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4229                                              kmp_int32 flags,
4230                                              size_t sizeof_kmp_task_t,
4231                                              size_t sizeof_shareds,
4232                                              kmp_routine_entry_t task_entry);
4233 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4234     ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4235     size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4236 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4237                                           kmp_task_t *task);
4238 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4239                                              kmp_task_t *task);
4240 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4241                                            kmp_task_t *new_task);
4242 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4243 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4244                                           int end_part);
4245 
4246 #if TASK_UNUSED
4247 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4248 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4249                               kmp_task_t *task);
4250 #endif // TASK_UNUSED
4251 
4252 /* ------------------------------------------------------------------------ */
4253 
4254 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4255 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4256 
4257 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4258     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4259     kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4260     kmp_depend_info_t *noalias_dep_list);
4261 
4262 KMP_EXPORT kmp_base_depnode_t *__kmpc_task_get_depnode(kmp_task_t *task);
4263 
4264 KMP_EXPORT kmp_depnode_list_t *__kmpc_task_get_successors(kmp_task_t *task);
4265 
4266 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4267                                      kmp_int32 ndeps,
4268                                      kmp_depend_info_t *dep_list,
4269                                      kmp_int32 ndeps_noalias,
4270                                      kmp_depend_info_t *noalias_dep_list);
4271 /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4272  *                               Placeholder for taskwait with nowait clause.*/
4273 KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4274                                             kmp_int32 ndeps,
4275                                             kmp_depend_info_t *dep_list,
4276                                             kmp_int32 ndeps_noalias,
4277                                             kmp_depend_info_t *noalias_dep_list,
4278                                             kmp_int32 has_no_wait);
4279 
4280 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4281                                 bool serialize_immediate);
4282 
4283 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4284                                    kmp_int32 cncl_kind);
4285 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4286                                               kmp_int32 cncl_kind);
4287 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4288 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4289 
4290 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4291 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4292 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4293                                 kmp_int32 if_val, kmp_uint64 *lb,
4294                                 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4295                                 kmp_int32 sched, kmp_uint64 grainsize,
4296                                 void *task_dup);
4297 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4298                                   kmp_task_t *task, kmp_int32 if_val,
4299                                   kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4300                                   kmp_int32 nogroup, kmp_int32 sched,
4301                                   kmp_uint64 grainsize, kmp_int32 modifier,
4302                                   void *task_dup);
4303 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4304 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4305 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4306 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4307                                                      int is_ws, int num,
4308                                                      void *data);
4309 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4310                                               int num, void *data);
4311 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4312                                                     int is_ws);
4313 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4314     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4315     kmp_task_affinity_info_t *affin_list);
4316 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4317 KMP_EXPORT int __kmp_get_max_teams(void);
4318 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4319 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4320 
4321 /* Interface target task integration */
4322 KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4323 KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4324 
4325 /* Lock interface routines (fast versions with gtid passed in) */
4326 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4327                                  void **user_lock);
4328 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4329                                       void **user_lock);
4330 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4331                                     void **user_lock);
4332 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4333                                          void **user_lock);
4334 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4335 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4336                                      void **user_lock);
4337 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4338                                   void **user_lock);
4339 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4340                                        void **user_lock);
4341 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4342 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4343                                      void **user_lock);
4344 
4345 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4346                                            void **user_lock, uintptr_t hint);
4347 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4348                                                 void **user_lock,
4349                                                 uintptr_t hint);
4350 
4351 #if OMPX_TASKGRAPH
4352 // Taskgraph's Record & Replay mechanism
4353 // __kmp_tdg_is_recording: check whether a given TDG is recording
4354 // status: the tdg's current status
4355 static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4356   return status == KMP_TDG_RECORDING;
4357 }
4358 
4359 KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4360                                               kmp_int32 input_flags,
4361                                               kmp_int32 tdg_id);
4362 KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4363                                        kmp_int32 input_flags, kmp_int32 tdg_id);
4364 #endif
4365 /* Interface to fast scalable reduce methods routines */
4366 
4367 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4368     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4369     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4370     kmp_critical_name *lck);
4371 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4372                                          kmp_critical_name *lck);
4373 KMP_EXPORT kmp_int32 __kmpc_reduce(
4374     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4375     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4376     kmp_critical_name *lck);
4377 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4378                                   kmp_critical_name *lck);
4379 
4380 /* Internal fast reduction routines */
4381 
4382 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4383     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4384     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4385     kmp_critical_name *lck);
4386 
4387 // this function is for testing set/get/determine reduce method
4388 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4389 
4390 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4391 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4392 
4393 // C++ port
4394 // missing 'extern "C"' declarations
4395 
4396 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4397 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4398 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4399                                         kmp_int32 num_threads);
4400 
4401 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4402                                       int proc_bind);
4403 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4404                                       kmp_int32 num_teams,
4405                                       kmp_int32 num_threads);
4406 KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4407                                         kmp_int32 thread_limit);
4408 /* Function for OpenMP 5.1 num_teams clause */
4409 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4410                                          kmp_int32 num_teams_lb,
4411                                          kmp_int32 num_teams_ub,
4412                                          kmp_int32 num_threads);
4413 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4414                                   kmpc_micro microtask, ...);
4415 struct kmp_dim { // loop bounds info casted to kmp_int64
4416   kmp_int64 lo; // lower
4417   kmp_int64 up; // upper
4418   kmp_int64 st; // stride
4419 };
4420 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4421                                      kmp_int32 num_dims,
4422                                      const struct kmp_dim *dims);
4423 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4424                                      const kmp_int64 *vec);
4425 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4426                                      const kmp_int64 *vec);
4427 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4428 
4429 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4430                                              void *data, size_t size,
4431                                              void ***cache);
4432 
4433 // The routines below are not exported.
4434 // Consider making them 'static' in corresponding source files.
4435 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4436                                            void *data_addr, size_t pc_size);
4437 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4438                                                 void *data_addr,
4439                                                 size_t pc_size);
4440 void __kmp_threadprivate_resize_cache(int newCapacity);
4441 void __kmp_cleanup_threadprivate_caches();
4442 
4443 // ompc_, kmpc_ entries moved from omp.h.
4444 #if KMP_OS_WINDOWS
4445 #define KMPC_CONVENTION __cdecl
4446 #else
4447 #define KMPC_CONVENTION
4448 #endif
4449 
4450 #ifndef __OMP_H
4451 typedef enum omp_sched_t {
4452   omp_sched_static = 1,
4453   omp_sched_dynamic = 2,
4454   omp_sched_guided = 3,
4455   omp_sched_auto = 4
4456 } omp_sched_t;
4457 typedef void *kmp_affinity_mask_t;
4458 #endif
4459 
4460 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4461 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4462 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4463 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4464 KMP_EXPORT int KMPC_CONVENTION
4465 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4466 KMP_EXPORT int KMPC_CONVENTION
4467 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4468 KMP_EXPORT int KMPC_CONVENTION
4469 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4470 
4471 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4472 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4473 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4474 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4475 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4476 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4477 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4478 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4479 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4480                                               char const *format);
4481 
4482 enum kmp_target_offload_kind {
4483   tgt_disabled = 0,
4484   tgt_default = 1,
4485   tgt_mandatory = 2
4486 };
4487 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4488 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4489 extern kmp_target_offload_kind_t __kmp_target_offload;
4490 extern int __kmpc_get_target_offload();
4491 
4492 // Constants used in libomptarget
4493 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4494 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4495 
4496 // OMP Pause Resource
4497 
4498 // The following enum is used both to set the status in __kmp_pause_status, and
4499 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4500 typedef enum kmp_pause_status_t {
4501   kmp_not_paused = 0, // status is not paused, or, requesting resume
4502   kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4503   kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4504 } kmp_pause_status_t;
4505 
4506 // This stores the pause state of the runtime
4507 extern kmp_pause_status_t __kmp_pause_status;
4508 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4509 extern int __kmp_pause_resource(kmp_pause_status_t level);
4510 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4511 extern void __kmp_resume_if_soft_paused();
4512 // Hard resume simply resets the status to not paused. Library will appear to
4513 // be uninitialized after hard pause. Let OMP constructs trigger required
4514 // initializations.
4515 static inline void __kmp_resume_if_hard_paused() {
4516   if (__kmp_pause_status == kmp_hard_paused) {
4517     __kmp_pause_status = kmp_not_paused;
4518   }
4519 }
4520 
4521 extern void __kmp_omp_display_env(int verbose);
4522 
4523 // 1: it is initializing hidden helper team
4524 extern volatile int __kmp_init_hidden_helper;
4525 // 1: the hidden helper team is done
4526 extern volatile int __kmp_hidden_helper_team_done;
4527 // 1: enable hidden helper task
4528 extern kmp_int32 __kmp_enable_hidden_helper;
4529 // Main thread of hidden helper team
4530 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4531 // Descriptors for the hidden helper threads
4532 extern kmp_info_t **__kmp_hidden_helper_threads;
4533 // Number of hidden helper threads
4534 extern kmp_int32 __kmp_hidden_helper_threads_num;
4535 // Number of hidden helper tasks that have not been executed yet
4536 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4537 
4538 extern void __kmp_hidden_helper_initialize();
4539 extern void __kmp_hidden_helper_threads_initz_routine();
4540 extern void __kmp_do_initialize_hidden_helper_threads();
4541 extern void __kmp_hidden_helper_threads_initz_wait();
4542 extern void __kmp_hidden_helper_initz_release();
4543 extern void __kmp_hidden_helper_threads_deinitz_wait();
4544 extern void __kmp_hidden_helper_threads_deinitz_release();
4545 extern void __kmp_hidden_helper_main_thread_wait();
4546 extern void __kmp_hidden_helper_worker_thread_wait();
4547 extern void __kmp_hidden_helper_worker_thread_signal();
4548 extern void __kmp_hidden_helper_main_thread_release();
4549 
4550 // Check whether a given thread is a hidden helper thread
4551 #define KMP_HIDDEN_HELPER_THREAD(gtid)                                         \
4552   ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4553 
4554 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)                                  \
4555   ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4556 
4557 #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid)                                    \
4558   ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4559 
4560 #define KMP_HIDDEN_HELPER_TEAM(team)                                           \
4561   (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4562 
4563 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4564 // main thread, is skipped.
4565 #define KMP_GTID_TO_SHADOW_GTID(gtid)                                          \
4566   ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4567 
4568 // Return the adjusted gtid value by subtracting from gtid the number
4569 // of hidden helper threads. This adjusted value is the gtid the thread would
4570 // have received if there were no hidden helper threads.
4571 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4572   int adjusted_gtid = gtid;
4573   if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4574       gtid - __kmp_hidden_helper_threads_num >= 0) {
4575     adjusted_gtid -= __kmp_hidden_helper_threads_num;
4576   }
4577   return adjusted_gtid;
4578 }
4579 
4580 // Support for error directive
4581 typedef enum kmp_severity_t {
4582   severity_warning = 1,
4583   severity_fatal = 2
4584 } kmp_severity_t;
4585 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4586 
4587 // Support for scope directive
4588 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4589 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4590 
4591 #ifdef __cplusplus
4592 }
4593 #endif
4594 
4595 template <bool C, bool S>
4596 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4597 template <bool C, bool S>
4598 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4599 template <bool C, bool S>
4600 extern void __kmp_atomic_suspend_64(int th_gtid,
4601                                     kmp_atomic_flag_64<C, S> *flag);
4602 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4603 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4604 template <bool C, bool S>
4605 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4606 template <bool C, bool S>
4607 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4608 template <bool C, bool S>
4609 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4610 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4611 #endif
4612 template <bool C, bool S>
4613 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4614 template <bool C, bool S>
4615 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4616 template <bool C, bool S>
4617 extern void __kmp_atomic_resume_64(int target_gtid,
4618                                    kmp_atomic_flag_64<C, S> *flag);
4619 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4620 
4621 template <bool C, bool S>
4622 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4623                            kmp_flag_32<C, S> *flag, int final_spin,
4624                            int *thread_finished,
4625 #if USE_ITT_BUILD
4626                            void *itt_sync_obj,
4627 #endif /* USE_ITT_BUILD */
4628                            kmp_int32 is_constrained);
4629 template <bool C, bool S>
4630 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4631                            kmp_flag_64<C, S> *flag, int final_spin,
4632                            int *thread_finished,
4633 #if USE_ITT_BUILD
4634                            void *itt_sync_obj,
4635 #endif /* USE_ITT_BUILD */
4636                            kmp_int32 is_constrained);
4637 template <bool C, bool S>
4638 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4639                                   kmp_atomic_flag_64<C, S> *flag,
4640                                   int final_spin, int *thread_finished,
4641 #if USE_ITT_BUILD
4642                                   void *itt_sync_obj,
4643 #endif /* USE_ITT_BUILD */
4644                                   kmp_int32 is_constrained);
4645 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4646                                kmp_flag_oncore *flag, int final_spin,
4647                                int *thread_finished,
4648 #if USE_ITT_BUILD
4649                                void *itt_sync_obj,
4650 #endif /* USE_ITT_BUILD */
4651                                kmp_int32 is_constrained);
4652 
4653 extern int __kmp_nesting_mode;
4654 extern int __kmp_nesting_mode_nlevels;
4655 extern int *__kmp_nesting_nth_level;
4656 extern void __kmp_init_nesting_mode();
4657 extern void __kmp_set_nesting_mode_threads();
4658 
4659 /// This class safely opens and closes a C-style FILE* object using RAII
4660 /// semantics. There are also methods which allow using stdout or stderr as
4661 /// the underlying FILE* object. With the implicit conversion operator to
4662 /// FILE*, an object with this type can be used in any function which takes
4663 /// a FILE* object e.g., fprintf().
4664 /// No close method is needed at use sites.
4665 class kmp_safe_raii_file_t {
4666   FILE *f;
4667 
4668   void close() {
4669     if (f && f != stdout && f != stderr) {
4670       fclose(f);
4671       f = nullptr;
4672     }
4673   }
4674 
4675 public:
4676   kmp_safe_raii_file_t() : f(nullptr) {}
4677   kmp_safe_raii_file_t(const char *filename, const char *mode,
4678                        const char *env_var = nullptr)
4679       : f(nullptr) {
4680     open(filename, mode, env_var);
4681   }
4682   ~kmp_safe_raii_file_t() { close(); }
4683 
4684   /// Open filename using mode. This is automatically closed in the destructor.
4685   /// The env_var parameter indicates the environment variable the filename
4686   /// came from if != nullptr.
4687   void open(const char *filename, const char *mode,
4688             const char *env_var = nullptr) {
4689     KMP_ASSERT(!f);
4690     f = fopen(filename, mode);
4691     if (!f) {
4692       int code = errno;
4693       if (env_var) {
4694         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4695                     KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4696       } else {
4697         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4698                     __kmp_msg_null);
4699       }
4700     }
4701   }
4702   /// Instead of erroring out, return non-zero when
4703   /// unsuccessful fopen() for any reason
4704   int try_open(const char *filename, const char *mode) {
4705     KMP_ASSERT(!f);
4706     f = fopen(filename, mode);
4707     if (!f)
4708       return errno;
4709     return 0;
4710   }
4711   /// Set the FILE* object to stdout and output there
4712   /// No open call should happen before this call.
4713   void set_stdout() {
4714     KMP_ASSERT(!f);
4715     f = stdout;
4716   }
4717   /// Set the FILE* object to stderr and output there
4718   /// No open call should happen before this call.
4719   void set_stderr() {
4720     KMP_ASSERT(!f);
4721     f = stderr;
4722   }
4723   operator bool() { return bool(f); }
4724   operator FILE *() { return f; }
4725 };
4726 
4727 template <typename SourceType, typename TargetType,
4728           bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4729           bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4730           bool isSourceSigned = std::is_signed<SourceType>::value,
4731           bool isTargetSigned = std::is_signed<TargetType>::value>
4732 struct kmp_convert {};
4733 
4734 // Both types are signed; Source smaller
4735 template <typename SourceType, typename TargetType>
4736 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4737   static TargetType to(SourceType src) { return (TargetType)src; }
4738 };
4739 // Source equal
4740 template <typename SourceType, typename TargetType>
4741 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4742   static TargetType to(SourceType src) { return src; }
4743 };
4744 // Source bigger
4745 template <typename SourceType, typename TargetType>
4746 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4747   static TargetType to(SourceType src) {
4748     KMP_ASSERT(src <= static_cast<SourceType>(
4749                           (std::numeric_limits<TargetType>::max)()));
4750     KMP_ASSERT(src >= static_cast<SourceType>(
4751                           (std::numeric_limits<TargetType>::min)()));
4752     return (TargetType)src;
4753   }
4754 };
4755 
4756 // Source signed, Target unsigned
4757 // Source smaller
4758 template <typename SourceType, typename TargetType>
4759 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4760   static TargetType to(SourceType src) {
4761     KMP_ASSERT(src >= 0);
4762     return (TargetType)src;
4763   }
4764 };
4765 // Source equal
4766 template <typename SourceType, typename TargetType>
4767 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4768   static TargetType to(SourceType src) {
4769     KMP_ASSERT(src >= 0);
4770     return (TargetType)src;
4771   }
4772 };
4773 // Source bigger
4774 template <typename SourceType, typename TargetType>
4775 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4776   static TargetType to(SourceType src) {
4777     KMP_ASSERT(src >= 0);
4778     KMP_ASSERT(src <= static_cast<SourceType>(
4779                           (std::numeric_limits<TargetType>::max)()));
4780     return (TargetType)src;
4781   }
4782 };
4783 
4784 // Source unsigned, Target signed
4785 // Source smaller
4786 template <typename SourceType, typename TargetType>
4787 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4788   static TargetType to(SourceType src) { return (TargetType)src; }
4789 };
4790 // Source equal
4791 template <typename SourceType, typename TargetType>
4792 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4793   static TargetType to(SourceType src) {
4794     KMP_ASSERT(src <= static_cast<SourceType>(
4795                           (std::numeric_limits<TargetType>::max)()));
4796     return (TargetType)src;
4797   }
4798 };
4799 // Source bigger
4800 template <typename SourceType, typename TargetType>
4801 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4802   static TargetType to(SourceType src) {
4803     KMP_ASSERT(src <= static_cast<SourceType>(
4804                           (std::numeric_limits<TargetType>::max)()));
4805     return (TargetType)src;
4806   }
4807 };
4808 
4809 // Source unsigned, Target unsigned
4810 // Source smaller
4811 template <typename SourceType, typename TargetType>
4812 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4813   static TargetType to(SourceType src) { return (TargetType)src; }
4814 };
4815 // Source equal
4816 template <typename SourceType, typename TargetType>
4817 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4818   static TargetType to(SourceType src) { return src; }
4819 };
4820 // Source bigger
4821 template <typename SourceType, typename TargetType>
4822 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4823   static TargetType to(SourceType src) {
4824     KMP_ASSERT(src <= static_cast<SourceType>(
4825                           (std::numeric_limits<TargetType>::max)()));
4826     return (TargetType)src;
4827   }
4828 };
4829 
4830 template <typename T1, typename T2>
4831 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4832   *dest = kmp_convert<T1, T2>::to(src);
4833 }
4834 
4835 #endif /* KMP_H */
4836