xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp.h (revision 43a5ec4eb41567cc92586503212743d89686d78f)
1 /*! \file */
2 /*
3  * kmp.h -- KPTS runtime header file.
4  */
5 
6 //===----------------------------------------------------------------------===//
7 //
8 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9 // See https://llvm.org/LICENSE.txt for license information.
10 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef KMP_H
15 #define KMP_H
16 
17 #include "kmp_config.h"
18 
19 /* #define BUILD_PARALLEL_ORDERED 1 */
20 
21 /* This fix replaces gettimeofday with clock_gettime for better scalability on
22    the Altix.  Requires user code to be linked with -lrt. */
23 //#define FIX_SGI_CLOCK
24 
25 /* Defines for OpenMP 3.0 tasking and auto scheduling */
26 
27 #ifndef KMP_STATIC_STEAL_ENABLED
28 #define KMP_STATIC_STEAL_ENABLED 1
29 #endif
30 
31 #define TASK_CURRENT_NOT_QUEUED 0
32 #define TASK_CURRENT_QUEUED 1
33 
34 #ifdef BUILD_TIED_TASK_STACK
35 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37 // Number of entries in each task stack array
38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39 // Mask for determining index into stack block
40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41 #endif // BUILD_TIED_TASK_STACK
42 
43 #define TASK_NOT_PUSHED 1
44 #define TASK_SUCCESSFULLY_PUSHED 0
45 #define TASK_TIED 1
46 #define TASK_UNTIED 0
47 #define TASK_EXPLICIT 1
48 #define TASK_IMPLICIT 0
49 #define TASK_PROXY 1
50 #define TASK_FULL 0
51 #define TASK_DETACHABLE 1
52 #define TASK_UNDETACHABLE 0
53 
54 #define KMP_CANCEL_THREADS
55 #define KMP_THREAD_ATTR
56 
57 // Android does not have pthread_cancel.  Undefine KMP_CANCEL_THREADS if being
58 // built on Android
59 #if defined(__ANDROID__)
60 #undef KMP_CANCEL_THREADS
61 #endif
62 
63 #include <signal.h>
64 #include <stdarg.h>
65 #include <stddef.h>
66 #include <stdio.h>
67 #include <stdlib.h>
68 #include <string.h>
69 #include <limits>
70 #include <type_traits>
71 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72    Microsoft library. Some macros provided below to replace these functions  */
73 #ifndef __ABSOFT_WIN
74 #include <sys/types.h>
75 #endif
76 #include <limits.h>
77 #include <time.h>
78 
79 #include <errno.h>
80 
81 #include "kmp_os.h"
82 
83 #include "kmp_safe_c_api.h"
84 
85 #if KMP_STATS_ENABLED
86 class kmp_stats_list;
87 #endif
88 
89 #if KMP_USE_HIER_SCHED
90 // Only include hierarchical scheduling if affinity is supported
91 #undef KMP_USE_HIER_SCHED
92 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93 #endif
94 
95 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96 #include "hwloc.h"
97 #ifndef HWLOC_OBJ_NUMANODE
98 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99 #endif
100 #ifndef HWLOC_OBJ_PACKAGE
101 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102 #endif
103 #if HWLOC_API_VERSION >= 0x00020000
104 // hwloc 2.0 changed type of depth of object from unsigned to int
105 typedef int kmp_hwloc_depth_t;
106 #else
107 typedef unsigned int kmp_hwloc_depth_t;
108 #endif
109 #endif
110 
111 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
112 #include <xmmintrin.h>
113 #endif
114 
115 #include "kmp_debug.h"
116 #include "kmp_lock.h"
117 #include "kmp_version.h"
118 #if USE_DEBUGGER
119 #include "kmp_debugger.h"
120 #endif
121 #include "kmp_i18n.h"
122 
123 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
124 
125 #include "kmp_wrapper_malloc.h"
126 #if KMP_OS_UNIX
127 #include <unistd.h>
128 #if !defined NSIG && defined _NSIG
129 #define NSIG _NSIG
130 #endif
131 #endif
132 
133 #if KMP_OS_LINUX
134 #pragma weak clock_gettime
135 #endif
136 
137 #if OMPT_SUPPORT
138 #include "ompt-internal.h"
139 #endif
140 
141 #if OMPD_SUPPORT
142 #include "ompd-specific.h"
143 #endif
144 
145 #ifndef UNLIKELY
146 #define UNLIKELY(x) (x)
147 #endif
148 
149 // Affinity format function
150 #include "kmp_str.h"
151 
152 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
153 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
154 // free lists of limited size.
155 #ifndef USE_FAST_MEMORY
156 #define USE_FAST_MEMORY 3
157 #endif
158 
159 #ifndef KMP_NESTED_HOT_TEAMS
160 #define KMP_NESTED_HOT_TEAMS 0
161 #define USE_NESTED_HOT_ARG(x)
162 #else
163 #if KMP_NESTED_HOT_TEAMS
164 #define USE_NESTED_HOT_ARG(x) , x
165 #else
166 #define USE_NESTED_HOT_ARG(x)
167 #endif
168 #endif
169 
170 // Assume using BGET compare_exchange instruction instead of lock by default.
171 #ifndef USE_CMP_XCHG_FOR_BGET
172 #define USE_CMP_XCHG_FOR_BGET 1
173 #endif
174 
175 // Test to see if queuing lock is better than bootstrap lock for bget
176 // #ifndef USE_QUEUING_LOCK_FOR_BGET
177 // #define USE_QUEUING_LOCK_FOR_BGET
178 // #endif
179 
180 #define KMP_NSEC_PER_SEC 1000000000L
181 #define KMP_USEC_PER_SEC 1000000L
182 
183 /*!
184 @ingroup BASIC_TYPES
185 @{
186 */
187 
188 /*!
189 Values for bit flags used in the ident_t to describe the fields.
190 */
191 enum {
192   /*! Use trampoline for internal microtasks */
193   KMP_IDENT_IMB = 0x01,
194   /*! Use c-style ident structure */
195   KMP_IDENT_KMPC = 0x02,
196   /* 0x04 is no longer used */
197   /*! Entry point generated by auto-parallelization */
198   KMP_IDENT_AUTOPAR = 0x08,
199   /*! Compiler generates atomic reduction option for kmpc_reduce* */
200   KMP_IDENT_ATOMIC_REDUCE = 0x10,
201   /*! To mark a 'barrier' directive in user code */
202   KMP_IDENT_BARRIER_EXPL = 0x20,
203   /*! To Mark implicit barriers. */
204   KMP_IDENT_BARRIER_IMPL = 0x0040,
205   KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
206   KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
207   KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
208 
209   KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
210   KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
211 
212   /*! To mark a static loop in OMPT callbacks */
213   KMP_IDENT_WORK_LOOP = 0x200,
214   /*! To mark a sections directive in OMPT callbacks */
215   KMP_IDENT_WORK_SECTIONS = 0x400,
216   /*! To mark a distribute construct in OMPT callbacks */
217   KMP_IDENT_WORK_DISTRIBUTE = 0x800,
218   /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
219       not currently used. If one day we need more bits, then we can use
220       an invalid combination of hints to mean that another, larger field
221       should be used in a different flag. */
222   KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
223   KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
224   KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
225   KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
226   KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
227   KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
228 };
229 
230 /*!
231  * The ident structure that describes a source location.
232  */
233 typedef struct ident {
234   kmp_int32 reserved_1; /**<  might be used in Fortran; see above  */
235   kmp_int32 flags; /**<  also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
236                       identifies this union member  */
237   kmp_int32 reserved_2; /**<  not really used in Fortran any more; see above */
238 #if USE_ITT_BUILD
239 /*  but currently used for storing region-specific ITT */
240 /*  contextual information. */
241 #endif /* USE_ITT_BUILD */
242   kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++  */
243   char const *psource; /**< String describing the source location.
244                        The string is composed of semi-colon separated fields
245                        which describe the source file, the function and a pair
246                        of line numbers that delimit the construct. */
247   // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
248   kmp_int32 get_openmp_version() {
249     return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
250   }
251 } ident_t;
252 /*!
253 @}
254 */
255 
256 // Some forward declarations.
257 typedef union kmp_team kmp_team_t;
258 typedef struct kmp_taskdata kmp_taskdata_t;
259 typedef union kmp_task_team kmp_task_team_t;
260 typedef union kmp_team kmp_team_p;
261 typedef union kmp_info kmp_info_p;
262 typedef union kmp_root kmp_root_p;
263 
264 template <bool C = false, bool S = true> class kmp_flag_32;
265 template <bool C = false, bool S = true> class kmp_flag_64;
266 class kmp_flag_oncore;
267 
268 #ifdef __cplusplus
269 extern "C" {
270 #endif
271 
272 /* ------------------------------------------------------------------------ */
273 
274 /* Pack two 32-bit signed integers into a 64-bit signed integer */
275 /* ToDo: Fix word ordering for big-endian machines. */
276 #define KMP_PACK_64(HIGH_32, LOW_32)                                           \
277   ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
278 
279 // Generic string manipulation macros. Assume that _x is of type char *
280 #define SKIP_WS(_x)                                                            \
281   {                                                                            \
282     while (*(_x) == ' ' || *(_x) == '\t')                                      \
283       (_x)++;                                                                  \
284   }
285 #define SKIP_DIGITS(_x)                                                        \
286   {                                                                            \
287     while (*(_x) >= '0' && *(_x) <= '9')                                       \
288       (_x)++;                                                                  \
289   }
290 #define SKIP_TOKEN(_x)                                                         \
291   {                                                                            \
292     while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
293            (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_')                     \
294       (_x)++;                                                                  \
295   }
296 #define SKIP_TO(_x, _c)                                                        \
297   {                                                                            \
298     while (*(_x) != '\0' && *(_x) != (_c))                                     \
299       (_x)++;                                                                  \
300   }
301 
302 /* ------------------------------------------------------------------------ */
303 
304 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
305 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
306 
307 /* ------------------------------------------------------------------------ */
308 /* Enumeration types */
309 
310 enum kmp_state_timer {
311   ts_stop,
312   ts_start,
313   ts_pause,
314 
315   ts_last_state
316 };
317 
318 enum dynamic_mode {
319   dynamic_default,
320 #ifdef USE_LOAD_BALANCE
321   dynamic_load_balance,
322 #endif /* USE_LOAD_BALANCE */
323   dynamic_random,
324   dynamic_thread_limit,
325   dynamic_max
326 };
327 
328 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
329  * not include it here */
330 #ifndef KMP_SCHED_TYPE_DEFINED
331 #define KMP_SCHED_TYPE_DEFINED
332 typedef enum kmp_sched {
333   kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
334   // Note: need to adjust __kmp_sch_map global array in case enum is changed
335   kmp_sched_static = 1, // mapped to kmp_sch_static_chunked           (33)
336   kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked          (35)
337   kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked           (36)
338   kmp_sched_auto = 4, // mapped to kmp_sch_auto                     (38)
339   kmp_sched_upper_std = 5, // upper bound for standard schedules
340   kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
341   kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
342 #if KMP_STATIC_STEAL_ENABLED
343   kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
344 #endif
345   kmp_sched_upper,
346   kmp_sched_default = kmp_sched_static, // default scheduling
347   kmp_sched_monotonic = 0x80000000
348 } kmp_sched_t;
349 #endif
350 
351 /*!
352  @ingroup WORK_SHARING
353  * Describes the loop schedule to be used for a parallel for loop.
354  */
355 enum sched_type : kmp_int32 {
356   kmp_sch_lower = 32, /**< lower bound for unordered values */
357   kmp_sch_static_chunked = 33,
358   kmp_sch_static = 34, /**< static unspecialized */
359   kmp_sch_dynamic_chunked = 35,
360   kmp_sch_guided_chunked = 36, /**< guided unspecialized */
361   kmp_sch_runtime = 37,
362   kmp_sch_auto = 38, /**< auto */
363   kmp_sch_trapezoidal = 39,
364 
365   /* accessible only through KMP_SCHEDULE environment variable */
366   kmp_sch_static_greedy = 40,
367   kmp_sch_static_balanced = 41,
368   /* accessible only through KMP_SCHEDULE environment variable */
369   kmp_sch_guided_iterative_chunked = 42,
370   kmp_sch_guided_analytical_chunked = 43,
371   /* accessible only through KMP_SCHEDULE environment variable */
372   kmp_sch_static_steal = 44,
373 
374   /* static with chunk adjustment (e.g., simd) */
375   kmp_sch_static_balanced_chunked = 45,
376   kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
377   kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
378 
379   /* accessible only through KMP_SCHEDULE environment variable */
380   kmp_sch_upper, /**< upper bound for unordered values */
381 
382   kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
383   kmp_ord_static_chunked = 65,
384   kmp_ord_static = 66, /**< ordered static unspecialized */
385   kmp_ord_dynamic_chunked = 67,
386   kmp_ord_guided_chunked = 68,
387   kmp_ord_runtime = 69,
388   kmp_ord_auto = 70, /**< ordered auto */
389   kmp_ord_trapezoidal = 71,
390   kmp_ord_upper, /**< upper bound for ordered values */
391 
392   /* Schedules for Distribute construct */
393   kmp_distribute_static_chunked = 91, /**< distribute static chunked */
394   kmp_distribute_static = 92, /**< distribute static unspecialized */
395 
396   /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
397      single iteration/chunk, even if the loop is serialized. For the schedule
398      types listed above, the entire iteration vector is returned if the loop is
399      serialized. This doesn't work for gcc/gcomp sections. */
400   kmp_nm_lower = 160, /**< lower bound for nomerge values */
401 
402   kmp_nm_static_chunked =
403       (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
404   kmp_nm_static = 162, /**< static unspecialized */
405   kmp_nm_dynamic_chunked = 163,
406   kmp_nm_guided_chunked = 164, /**< guided unspecialized */
407   kmp_nm_runtime = 165,
408   kmp_nm_auto = 166, /**< auto */
409   kmp_nm_trapezoidal = 167,
410 
411   /* accessible only through KMP_SCHEDULE environment variable */
412   kmp_nm_static_greedy = 168,
413   kmp_nm_static_balanced = 169,
414   /* accessible only through KMP_SCHEDULE environment variable */
415   kmp_nm_guided_iterative_chunked = 170,
416   kmp_nm_guided_analytical_chunked = 171,
417   kmp_nm_static_steal =
418       172, /* accessible only through OMP_SCHEDULE environment variable */
419 
420   kmp_nm_ord_static_chunked = 193,
421   kmp_nm_ord_static = 194, /**< ordered static unspecialized */
422   kmp_nm_ord_dynamic_chunked = 195,
423   kmp_nm_ord_guided_chunked = 196,
424   kmp_nm_ord_runtime = 197,
425   kmp_nm_ord_auto = 198, /**< auto */
426   kmp_nm_ord_trapezoidal = 199,
427   kmp_nm_upper, /**< upper bound for nomerge values */
428 
429   /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
430      we need to distinguish the three possible cases (no modifier, monotonic
431      modifier, nonmonotonic modifier), we need separate bits for each modifier.
432      The absence of monotonic does not imply nonmonotonic, especially since 4.5
433      says that the behaviour of the "no modifier" case is implementation defined
434      in 4.5, but will become "nonmonotonic" in 5.0.
435 
436      Since we're passing a full 32 bit value, we can use a couple of high bits
437      for these flags; out of paranoia we avoid the sign bit.
438 
439      These modifiers can be or-ed into non-static schedules by the compiler to
440      pass the additional information. They will be stripped early in the
441      processing in __kmp_dispatch_init when setting up schedules, so most of the
442      code won't ever see schedules with these bits set.  */
443   kmp_sch_modifier_monotonic =
444       (1 << 29), /**< Set if the monotonic schedule modifier was present */
445   kmp_sch_modifier_nonmonotonic =
446       (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
447 
448 #define SCHEDULE_WITHOUT_MODIFIERS(s)                                          \
449   (enum sched_type)(                                                           \
450       (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
451 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
452 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
453 #define SCHEDULE_HAS_NO_MODIFIERS(s)                                           \
454   (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
455 #define SCHEDULE_GET_MODIFIERS(s)                                              \
456   ((enum sched_type)(                                                          \
457       (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
458 #define SCHEDULE_SET_MODIFIERS(s, m)                                           \
459   (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
460 #define SCHEDULE_NONMONOTONIC 0
461 #define SCHEDULE_MONOTONIC 1
462 
463   kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
464 };
465 
466 // Apply modifiers on internal kind to standard kind
467 static inline void
468 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
469                                enum sched_type internal_kind) {
470   if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
471     *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
472   }
473 }
474 
475 // Apply modifiers on standard kind to internal kind
476 static inline void
477 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
478                                enum sched_type *internal_kind) {
479   if ((int)kind & (int)kmp_sched_monotonic) {
480     *internal_kind = (enum sched_type)((int)*internal_kind |
481                                        (int)kmp_sch_modifier_monotonic);
482   }
483 }
484 
485 // Get standard schedule without modifiers
486 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
487   return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
488 }
489 
490 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
491 typedef union kmp_r_sched {
492   struct {
493     enum sched_type r_sched_type;
494     int chunk;
495   };
496   kmp_int64 sched;
497 } kmp_r_sched_t;
498 
499 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
500 // internal schedule types
501 
502 enum library_type {
503   library_none,
504   library_serial,
505   library_turnaround,
506   library_throughput
507 };
508 
509 #if KMP_OS_LINUX
510 enum clock_function_type {
511   clock_function_gettimeofday,
512   clock_function_clock_gettime
513 };
514 #endif /* KMP_OS_LINUX */
515 
516 #if KMP_MIC_SUPPORTED
517 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
518 #endif
519 
520 /* -- fast reduction stuff ------------------------------------------------ */
521 
522 #undef KMP_FAST_REDUCTION_BARRIER
523 #define KMP_FAST_REDUCTION_BARRIER 1
524 
525 #undef KMP_FAST_REDUCTION_CORE_DUO
526 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
527 #define KMP_FAST_REDUCTION_CORE_DUO 1
528 #endif
529 
530 enum _reduction_method {
531   reduction_method_not_defined = 0,
532   critical_reduce_block = (1 << 8),
533   atomic_reduce_block = (2 << 8),
534   tree_reduce_block = (3 << 8),
535   empty_reduce_block = (4 << 8)
536 };
537 
538 // Description of the packed_reduction_method variable:
539 // The packed_reduction_method variable consists of two enum types variables
540 // that are packed together into 0-th byte and 1-st byte:
541 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
542 // barrier that will be used in fast reduction: bs_plain_barrier or
543 // bs_reduction_barrier
544 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
545 // be used in fast reduction;
546 // Reduction method is of 'enum _reduction_method' type and it's defined the way
547 // so that the bits of 0-th byte are empty, so no need to execute a shift
548 // instruction while packing/unpacking
549 
550 #if KMP_FAST_REDUCTION_BARRIER
551 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
552   ((reduction_method) | (barrier_type))
553 
554 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
555   ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
556 
557 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method)                      \
558   ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
559 #else
560 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
561   (reduction_method)
562 
563 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
564   (packed_reduction_method)
565 
566 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
567 #endif
568 
569 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block)  \
570   ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) ==                       \
571    (which_reduction_block))
572 
573 #if KMP_FAST_REDUCTION_BARRIER
574 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER                               \
575   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
576 
577 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER                                   \
578   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
579 #endif
580 
581 typedef int PACKED_REDUCTION_METHOD_T;
582 
583 /* -- end of fast reduction stuff ----------------------------------------- */
584 
585 #if KMP_OS_WINDOWS
586 #define USE_CBLKDATA
587 #if KMP_MSVC_COMPAT
588 #pragma warning(push)
589 #pragma warning(disable : 271 310)
590 #endif
591 #include <windows.h>
592 #if KMP_MSVC_COMPAT
593 #pragma warning(pop)
594 #endif
595 #endif
596 
597 #if KMP_OS_UNIX
598 #include <dlfcn.h>
599 #include <pthread.h>
600 #endif
601 
602 enum kmp_hw_t : int {
603   KMP_HW_UNKNOWN = -1,
604   KMP_HW_SOCKET = 0,
605   KMP_HW_PROC_GROUP,
606   KMP_HW_NUMA,
607   KMP_HW_DIE,
608   KMP_HW_LLC,
609   KMP_HW_L3,
610   KMP_HW_TILE,
611   KMP_HW_MODULE,
612   KMP_HW_L2,
613   KMP_HW_L1,
614   KMP_HW_CORE,
615   KMP_HW_THREAD,
616   KMP_HW_LAST
617 };
618 
619 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type)                                   \
620   KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
621 #define KMP_ASSERT_VALID_HW_TYPE(type)                                         \
622   KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
623 
624 #define KMP_FOREACH_HW_TYPE(type)                                              \
625   for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST;                        \
626        type = (kmp_hw_t)((int)type + 1))
627 
628 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
629 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
630 
631 /* Only Linux* OS and Windows* OS support thread affinity. */
632 #if KMP_AFFINITY_SUPPORTED
633 
634 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
635 #if KMP_OS_WINDOWS
636 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
637 typedef struct GROUP_AFFINITY {
638   KAFFINITY Mask;
639   WORD Group;
640   WORD Reserved[3];
641 } GROUP_AFFINITY;
642 #endif /* _MSC_VER < 1600 */
643 #if KMP_GROUP_AFFINITY
644 extern int __kmp_num_proc_groups;
645 #else
646 static const int __kmp_num_proc_groups = 1;
647 #endif /* KMP_GROUP_AFFINITY */
648 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
649 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
650 
651 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
652 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
653 
654 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
655 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
656 
657 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
658                                              GROUP_AFFINITY *);
659 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
660 #endif /* KMP_OS_WINDOWS */
661 
662 #if KMP_USE_HWLOC
663 extern hwloc_topology_t __kmp_hwloc_topology;
664 extern int __kmp_hwloc_error;
665 #endif
666 
667 extern size_t __kmp_affin_mask_size;
668 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
669 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
670 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
671 #define KMP_CPU_SET_ITERATE(i, mask)                                           \
672   for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
673 #define KMP_CPU_SET(i, mask) (mask)->set(i)
674 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
675 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
676 #define KMP_CPU_ZERO(mask) (mask)->zero()
677 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
678 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
679 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
680 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
681 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
682 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
683 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
684 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
685 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
686 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
687 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
688 #define KMP_CPU_ALLOC_ARRAY(arr, n)                                            \
689   (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
690 #define KMP_CPU_FREE_ARRAY(arr, n)                                             \
691   __kmp_affinity_dispatch->deallocate_mask_array(arr)
692 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
693 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
694 #define __kmp_get_system_affinity(mask, abort_bool)                            \
695   (mask)->get_system_affinity(abort_bool)
696 #define __kmp_set_system_affinity(mask, abort_bool)                            \
697   (mask)->set_system_affinity(abort_bool)
698 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
699 
700 class KMPAffinity {
701 public:
702   class Mask {
703   public:
704     void *operator new(size_t n);
705     void operator delete(void *p);
706     void *operator new[](size_t n);
707     void operator delete[](void *p);
708     virtual ~Mask() {}
709     // Set bit i to 1
710     virtual void set(int i) {}
711     // Return bit i
712     virtual bool is_set(int i) const { return false; }
713     // Set bit i to 0
714     virtual void clear(int i) {}
715     // Zero out entire mask
716     virtual void zero() {}
717     // Copy src into this mask
718     virtual void copy(const Mask *src) {}
719     // this &= rhs
720     virtual void bitwise_and(const Mask *rhs) {}
721     // this |= rhs
722     virtual void bitwise_or(const Mask *rhs) {}
723     // this = ~this
724     virtual void bitwise_not() {}
725     // API for iterating over an affinity mask
726     // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
727     virtual int begin() const { return 0; }
728     virtual int end() const { return 0; }
729     virtual int next(int previous) const { return 0; }
730 #if KMP_OS_WINDOWS
731     virtual int set_process_affinity(bool abort_on_error) const { return -1; }
732 #endif
733     // Set the system's affinity to this affinity mask's value
734     virtual int set_system_affinity(bool abort_on_error) const { return -1; }
735     // Set this affinity mask to the current system affinity
736     virtual int get_system_affinity(bool abort_on_error) { return -1; }
737     // Only 1 DWORD in the mask should have any procs set.
738     // Return the appropriate index, or -1 for an invalid mask.
739     virtual int get_proc_group() const { return -1; }
740   };
741   void *operator new(size_t n);
742   void operator delete(void *p);
743   // Need virtual destructor
744   virtual ~KMPAffinity() = default;
745   // Determine if affinity is capable
746   virtual void determine_capable(const char *env_var) {}
747   // Bind the current thread to os proc
748   virtual void bind_thread(int proc) {}
749   // Factory functions to allocate/deallocate a mask
750   virtual Mask *allocate_mask() { return nullptr; }
751   virtual void deallocate_mask(Mask *m) {}
752   virtual Mask *allocate_mask_array(int num) { return nullptr; }
753   virtual void deallocate_mask_array(Mask *m) {}
754   virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
755   static void pick_api();
756   static void destroy_api();
757   enum api_type {
758     NATIVE_OS
759 #if KMP_USE_HWLOC
760     ,
761     HWLOC
762 #endif
763   };
764   virtual api_type get_api_type() const {
765     KMP_ASSERT(0);
766     return NATIVE_OS;
767   }
768 
769 private:
770   static bool picked_api;
771 };
772 
773 typedef KMPAffinity::Mask kmp_affin_mask_t;
774 extern KMPAffinity *__kmp_affinity_dispatch;
775 
776 // Declare local char buffers with this size for printing debug and info
777 // messages, using __kmp_affinity_print_mask().
778 #define KMP_AFFIN_MASK_PRINT_LEN 1024
779 
780 enum affinity_type {
781   affinity_none = 0,
782   affinity_physical,
783   affinity_logical,
784   affinity_compact,
785   affinity_scatter,
786   affinity_explicit,
787   affinity_balanced,
788   affinity_disabled, // not used outsize the env var parser
789   affinity_default
790 };
791 
792 enum affinity_top_method {
793   affinity_top_method_all = 0, // try all (supported) methods, in order
794 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
795   affinity_top_method_apicid,
796   affinity_top_method_x2apicid,
797   affinity_top_method_x2apicid_1f,
798 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
799   affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
800 #if KMP_GROUP_AFFINITY
801   affinity_top_method_group,
802 #endif /* KMP_GROUP_AFFINITY */
803   affinity_top_method_flat,
804 #if KMP_USE_HWLOC
805   affinity_top_method_hwloc,
806 #endif
807   affinity_top_method_default
808 };
809 
810 #define affinity_respect_mask_default (-1)
811 
812 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
813 extern kmp_hw_t __kmp_affinity_gran; /* Affinity granularity */
814 extern int __kmp_affinity_gran_levels; /* corresponding int value */
815 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
816 extern enum affinity_top_method __kmp_affinity_top_method;
817 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
818 extern int __kmp_affinity_offset; /* Affinity offset value  */
819 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
820 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
821 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
822 extern char *__kmp_affinity_proclist; /* proc ID list */
823 extern kmp_affin_mask_t *__kmp_affinity_masks;
824 extern unsigned __kmp_affinity_num_masks;
825 extern void __kmp_affinity_bind_thread(int which);
826 
827 extern kmp_affin_mask_t *__kmp_affin_fullMask;
828 extern char *__kmp_cpuinfo_file;
829 
830 #endif /* KMP_AFFINITY_SUPPORTED */
831 
832 // This needs to be kept in sync with the values in omp.h !!!
833 typedef enum kmp_proc_bind_t {
834   proc_bind_false = 0,
835   proc_bind_true,
836   proc_bind_primary,
837   proc_bind_close,
838   proc_bind_spread,
839   proc_bind_intel, // use KMP_AFFINITY interface
840   proc_bind_default
841 } kmp_proc_bind_t;
842 
843 typedef struct kmp_nested_proc_bind_t {
844   kmp_proc_bind_t *bind_types;
845   int size;
846   int used;
847 } kmp_nested_proc_bind_t;
848 
849 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
850 
851 extern int __kmp_display_affinity;
852 extern char *__kmp_affinity_format;
853 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
854 #if OMPT_SUPPORT
855 extern int __kmp_tool;
856 extern char *__kmp_tool_libraries;
857 #endif // OMPT_SUPPORT
858 
859 #if KMP_AFFINITY_SUPPORTED
860 #define KMP_PLACE_ALL (-1)
861 #define KMP_PLACE_UNDEFINED (-2)
862 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
863 #define KMP_AFFINITY_NON_PROC_BIND                                             \
864   ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false ||                 \
865     __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) &&                \
866    (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
867 #endif /* KMP_AFFINITY_SUPPORTED */
868 
869 extern int __kmp_affinity_num_places;
870 
871 typedef enum kmp_cancel_kind_t {
872   cancel_noreq = 0,
873   cancel_parallel = 1,
874   cancel_loop = 2,
875   cancel_sections = 3,
876   cancel_taskgroup = 4
877 } kmp_cancel_kind_t;
878 
879 // KMP_HW_SUBSET support:
880 typedef struct kmp_hws_item {
881   int num;
882   int offset;
883 } kmp_hws_item_t;
884 
885 extern kmp_hws_item_t __kmp_hws_socket;
886 extern kmp_hws_item_t __kmp_hws_die;
887 extern kmp_hws_item_t __kmp_hws_node;
888 extern kmp_hws_item_t __kmp_hws_tile;
889 extern kmp_hws_item_t __kmp_hws_core;
890 extern kmp_hws_item_t __kmp_hws_proc;
891 extern int __kmp_hws_requested;
892 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
893 
894 /* ------------------------------------------------------------------------ */
895 
896 #define KMP_PAD(type, sz)                                                      \
897   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
898 
899 // We need to avoid using -1 as a GTID as +1 is added to the gtid
900 // when storing it in a lock, and the value 0 is reserved.
901 #define KMP_GTID_DNE (-2) /* Does not exist */
902 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
903 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
904 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
905 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
906 
907 /* OpenMP 5.0 Memory Management support */
908 
909 #ifndef __OMP_H
910 // Duplicate type definitions from omp.h
911 typedef uintptr_t omp_uintptr_t;
912 
913 typedef enum {
914   omp_atk_sync_hint = 1,
915   omp_atk_alignment = 2,
916   omp_atk_access = 3,
917   omp_atk_pool_size = 4,
918   omp_atk_fallback = 5,
919   omp_atk_fb_data = 6,
920   omp_atk_pinned = 7,
921   omp_atk_partition = 8
922 } omp_alloctrait_key_t;
923 
924 typedef enum {
925   omp_atv_false = 0,
926   omp_atv_true = 1,
927   omp_atv_contended = 3,
928   omp_atv_uncontended = 4,
929   omp_atv_serialized = 5,
930   omp_atv_sequential = omp_atv_serialized, // (deprecated)
931   omp_atv_private = 6,
932   omp_atv_all = 7,
933   omp_atv_thread = 8,
934   omp_atv_pteam = 9,
935   omp_atv_cgroup = 10,
936   omp_atv_default_mem_fb = 11,
937   omp_atv_null_fb = 12,
938   omp_atv_abort_fb = 13,
939   omp_atv_allocator_fb = 14,
940   omp_atv_environment = 15,
941   omp_atv_nearest = 16,
942   omp_atv_blocked = 17,
943   omp_atv_interleaved = 18
944 } omp_alloctrait_value_t;
945 #define omp_atv_default ((omp_uintptr_t)-1)
946 
947 typedef void *omp_memspace_handle_t;
948 extern omp_memspace_handle_t const omp_default_mem_space;
949 extern omp_memspace_handle_t const omp_large_cap_mem_space;
950 extern omp_memspace_handle_t const omp_const_mem_space;
951 extern omp_memspace_handle_t const omp_high_bw_mem_space;
952 extern omp_memspace_handle_t const omp_low_lat_mem_space;
953 // Preview of target memory support
954 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
955 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
956 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
957 
958 typedef struct {
959   omp_alloctrait_key_t key;
960   omp_uintptr_t value;
961 } omp_alloctrait_t;
962 
963 typedef void *omp_allocator_handle_t;
964 extern omp_allocator_handle_t const omp_null_allocator;
965 extern omp_allocator_handle_t const omp_default_mem_alloc;
966 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
967 extern omp_allocator_handle_t const omp_const_mem_alloc;
968 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
969 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
970 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
971 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
972 extern omp_allocator_handle_t const omp_thread_mem_alloc;
973 // Preview of target memory support
974 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
975 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
976 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
977 extern omp_allocator_handle_t const kmp_max_mem_alloc;
978 extern omp_allocator_handle_t __kmp_def_allocator;
979 
980 // end of duplicate type definitions from omp.h
981 #endif
982 
983 extern int __kmp_memkind_available;
984 
985 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
986 
987 typedef struct kmp_allocator_t {
988   omp_memspace_handle_t memspace;
989   void **memkind; // pointer to memkind
990   int alignment;
991   omp_alloctrait_value_t fb;
992   kmp_allocator_t *fb_data;
993   kmp_uint64 pool_size;
994   kmp_uint64 pool_used;
995 } kmp_allocator_t;
996 
997 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
998                                                     omp_memspace_handle_t,
999                                                     int ntraits,
1000                                                     omp_alloctrait_t traits[]);
1001 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1002 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1003 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1004 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1005 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1006                            omp_allocator_handle_t al);
1007 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1008                             omp_allocator_handle_t al,
1009                             omp_allocator_handle_t free_al);
1010 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1011 
1012 extern void __kmp_init_memkind();
1013 extern void __kmp_fini_memkind();
1014 extern void __kmp_init_target_mem();
1015 
1016 /* ------------------------------------------------------------------------ */
1017 
1018 #define KMP_UINT64_MAX                                                         \
1019   (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1020 
1021 #define KMP_MIN_NTH 1
1022 
1023 #ifndef KMP_MAX_NTH
1024 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1025 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1026 #else
1027 #define KMP_MAX_NTH INT_MAX
1028 #endif
1029 #endif /* KMP_MAX_NTH */
1030 
1031 #ifdef PTHREAD_STACK_MIN
1032 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1033 #else
1034 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1035 #endif
1036 
1037 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1038 
1039 #if KMP_ARCH_X86
1040 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1041 #elif KMP_ARCH_X86_64
1042 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1043 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1044 #else
1045 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1046 #endif
1047 
1048 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1049 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1050 #define KMP_MAX_MALLOC_POOL_INCR                                               \
1051   (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1052 
1053 #define KMP_MIN_STKOFFSET (0)
1054 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1055 #if KMP_OS_DARWIN
1056 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1057 #else
1058 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1059 #endif
1060 
1061 #define KMP_MIN_STKPADDING (0)
1062 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1063 
1064 #define KMP_BLOCKTIME_MULTIPLIER                                               \
1065   (1000) /* number of blocktime units per second */
1066 #define KMP_MIN_BLOCKTIME (0)
1067 #define KMP_MAX_BLOCKTIME                                                      \
1068   (INT_MAX) /* Must be this for "infinite" setting the work */
1069 #define KMP_DEFAULT_BLOCKTIME (200) /*  __kmp_blocktime is in milliseconds  */
1070 
1071 #if KMP_USE_MONITOR
1072 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1073 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1074 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1075 
1076 /* Calculate new number of monitor wakeups for a specific block time based on
1077    previous monitor_wakeups. Only allow increasing number of wakeups */
1078 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups)                 \
1079   (((blocktime) == KMP_MAX_BLOCKTIME)   ? (monitor_wakeups)                    \
1080    : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS              \
1081    : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime)))            \
1082        ? (monitor_wakeups)                                                     \
1083        : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1084 
1085 /* Calculate number of intervals for a specific block time based on
1086    monitor_wakeups */
1087 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups)               \
1088   (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) /        \
1089    (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1090 #else
1091 #define KMP_BLOCKTIME(team, tid)                                               \
1092   (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1093 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1094 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1095 extern kmp_uint64 __kmp_ticks_per_msec;
1096 #if KMP_COMPILER_ICC
1097 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1098 #else
1099 #define KMP_NOW() __kmp_hardware_timestamp()
1100 #endif
1101 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1102 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1103   (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1104 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1105 #else
1106 // System time is retrieved sporadically while blocking.
1107 extern kmp_uint64 __kmp_now_nsec();
1108 #define KMP_NOW() __kmp_now_nsec()
1109 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1110 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1111   (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1112 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1113 #endif
1114 #endif // KMP_USE_MONITOR
1115 
1116 #define KMP_MIN_STATSCOLS 40
1117 #define KMP_MAX_STATSCOLS 4096
1118 #define KMP_DEFAULT_STATSCOLS 80
1119 
1120 #define KMP_MIN_INTERVAL 0
1121 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1122 #define KMP_DEFAULT_INTERVAL 0
1123 
1124 #define KMP_MIN_CHUNK 1
1125 #define KMP_MAX_CHUNK (INT_MAX - 1)
1126 #define KMP_DEFAULT_CHUNK 1
1127 
1128 #define KMP_MIN_DISP_NUM_BUFF 1
1129 #define KMP_DFLT_DISP_NUM_BUFF 7
1130 #define KMP_MAX_DISP_NUM_BUFF 4096
1131 
1132 #define KMP_MAX_ORDERED 8
1133 
1134 #define KMP_MAX_FIELDS 32
1135 
1136 #define KMP_MAX_BRANCH_BITS 31
1137 
1138 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1139 
1140 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1141 
1142 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1143 
1144 /* Minimum number of threads before switch to TLS gtid (experimentally
1145    determined) */
1146 /* josh TODO: what about OS X* tuning? */
1147 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1148 #define KMP_TLS_GTID_MIN 5
1149 #else
1150 #define KMP_TLS_GTID_MIN INT_MAX
1151 #endif
1152 
1153 #define KMP_MASTER_TID(tid) (0 == (tid))
1154 #define KMP_WORKER_TID(tid) (0 != (tid))
1155 
1156 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1157 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1158 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1159 
1160 #ifndef TRUE
1161 #define FALSE 0
1162 #define TRUE (!FALSE)
1163 #endif
1164 
1165 /* NOTE: all of the following constants must be even */
1166 
1167 #if KMP_OS_WINDOWS
1168 #define KMP_INIT_WAIT 64U /* initial number of spin-tests   */
1169 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1170 #elif KMP_OS_LINUX
1171 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1172 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1173 #elif KMP_OS_DARWIN
1174 /* TODO: tune for KMP_OS_DARWIN */
1175 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1176 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1177 #elif KMP_OS_DRAGONFLY
1178 /* TODO: tune for KMP_OS_DRAGONFLY */
1179 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1180 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1181 #elif KMP_OS_FREEBSD
1182 /* TODO: tune for KMP_OS_FREEBSD */
1183 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1184 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1185 #elif KMP_OS_NETBSD
1186 /* TODO: tune for KMP_OS_NETBSD */
1187 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1188 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1189 #elif KMP_OS_HURD
1190 /* TODO: tune for KMP_OS_HURD */
1191 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1192 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1193 #elif KMP_OS_OPENBSD
1194 /* TODO: tune for KMP_OS_OPENBSD */
1195 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1196 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1197 #endif
1198 
1199 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1200 typedef struct kmp_cpuid {
1201   kmp_uint32 eax;
1202   kmp_uint32 ebx;
1203   kmp_uint32 ecx;
1204   kmp_uint32 edx;
1205 } kmp_cpuid_t;
1206 
1207 typedef struct kmp_cpuinfo {
1208   int initialized; // If 0, other fields are not initialized.
1209   int signature; // CPUID(1).EAX
1210   int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1211   int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1212   // Model << 4 ) + Model)
1213   int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1214   int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1215   int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1216   int apic_id;
1217   int physical_id;
1218   int logical_id;
1219   kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1220   char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1221 } kmp_cpuinfo_t;
1222 
1223 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1224 
1225 #if KMP_OS_UNIX
1226 // subleaf is only needed for cache and topology discovery and can be set to
1227 // zero in most cases
1228 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1229   __asm__ __volatile__("cpuid"
1230                        : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1231                        : "a"(leaf), "c"(subleaf));
1232 }
1233 // Load p into FPU control word
1234 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1235   __asm__ __volatile__("fldcw %0" : : "m"(*p));
1236 }
1237 // Store FPU control word into p
1238 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1239   __asm__ __volatile__("fstcw %0" : "=m"(*p));
1240 }
1241 static inline void __kmp_clear_x87_fpu_status_word() {
1242 #if KMP_MIC
1243   // 32-bit protected mode x87 FPU state
1244   struct x87_fpu_state {
1245     unsigned cw;
1246     unsigned sw;
1247     unsigned tw;
1248     unsigned fip;
1249     unsigned fips;
1250     unsigned fdp;
1251     unsigned fds;
1252   };
1253   struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1254   __asm__ __volatile__("fstenv %0\n\t" // store FP env
1255                        "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1256                        "fldenv %0\n\t" // load FP env back
1257                        : "+m"(fpu_state), "+m"(fpu_state.sw));
1258 #else
1259   __asm__ __volatile__("fnclex");
1260 #endif // KMP_MIC
1261 }
1262 #if __SSE__
1263 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1264 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1265 #else
1266 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1267 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1268 #endif
1269 #else
1270 // Windows still has these as external functions in assembly file
1271 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1272 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1273 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1274 extern void __kmp_clear_x87_fpu_status_word();
1275 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1276 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1277 #endif // KMP_OS_UNIX
1278 
1279 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1280 
1281 #if KMP_ARCH_X86
1282 extern void __kmp_x86_pause(void);
1283 #elif KMP_MIC
1284 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1285 // regression after removal of extra PAUSE from spin loops. Changing
1286 // the delay from 100 to 300 showed even better performance than double PAUSE
1287 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1288 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1289 #else
1290 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1291 #endif
1292 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1293 #elif KMP_ARCH_PPC64
1294 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1295 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1296 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1297 #define KMP_CPU_PAUSE()                                                        \
1298   do {                                                                         \
1299     KMP_PPC64_PRI_LOW();                                                       \
1300     KMP_PPC64_PRI_MED();                                                       \
1301     KMP_PPC64_PRI_LOC_MB();                                                    \
1302   } while (0)
1303 #else
1304 #define KMP_CPU_PAUSE() /* nothing to do */
1305 #endif
1306 
1307 #define KMP_INIT_YIELD(count)                                                  \
1308   { (count) = __kmp_yield_init; }
1309 
1310 #define KMP_OVERSUBSCRIBED                                                     \
1311   (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1312 
1313 #define KMP_TRY_YIELD                                                          \
1314   ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1315 
1316 #define KMP_TRY_YIELD_OVERSUB                                                  \
1317   ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1318 
1319 #define KMP_YIELD(cond)                                                        \
1320   {                                                                            \
1321     KMP_CPU_PAUSE();                                                           \
1322     if ((cond) && (KMP_TRY_YIELD))                                             \
1323       __kmp_yield();                                                           \
1324   }
1325 
1326 #define KMP_YIELD_OVERSUB()                                                    \
1327   {                                                                            \
1328     KMP_CPU_PAUSE();                                                           \
1329     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1330       __kmp_yield();                                                           \
1331   }
1332 
1333 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1334 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1335 #define KMP_YIELD_SPIN(count)                                                  \
1336   {                                                                            \
1337     KMP_CPU_PAUSE();                                                           \
1338     if (KMP_TRY_YIELD) {                                                       \
1339       (count) -= 2;                                                            \
1340       if (!(count)) {                                                          \
1341         __kmp_yield();                                                         \
1342         (count) = __kmp_yield_next;                                            \
1343       }                                                                        \
1344     }                                                                          \
1345   }
1346 
1347 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count)                                     \
1348   {                                                                            \
1349     KMP_CPU_PAUSE();                                                           \
1350     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1351       __kmp_yield();                                                           \
1352     else if (__kmp_use_yield == 1) {                                           \
1353       (count) -= 2;                                                            \
1354       if (!(count)) {                                                          \
1355         __kmp_yield();                                                         \
1356         (count) = __kmp_yield_next;                                            \
1357       }                                                                        \
1358     }                                                                          \
1359   }
1360 
1361 // User-level Monitor/Mwait
1362 #if KMP_HAVE_UMWAIT
1363 // We always try for UMWAIT first
1364 #if KMP_HAVE_WAITPKG_INTRINSICS
1365 #if KMP_HAVE_IMMINTRIN_H
1366 #include <immintrin.h>
1367 #elif KMP_HAVE_INTRIN_H
1368 #include <intrin.h>
1369 #endif
1370 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1371 KMP_ATTRIBUTE_TARGET_WAITPKG
1372 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1373 #if !KMP_HAVE_WAITPKG_INTRINSICS
1374   uint32_t timeHi = uint32_t(counter >> 32);
1375   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1376   char flag;
1377   __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1378                    "setb   %0"
1379                    : "=r"(flag)
1380                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1381                    :);
1382   return flag;
1383 #else
1384   return _tpause(hint, counter);
1385 #endif
1386 }
1387 KMP_ATTRIBUTE_TARGET_WAITPKG
1388 static inline void __kmp_umonitor(void *cacheline) {
1389 #if !KMP_HAVE_WAITPKG_INTRINSICS
1390   __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1391                    :
1392                    : "a"(cacheline)
1393                    :);
1394 #else
1395   _umonitor(cacheline);
1396 #endif
1397 }
1398 KMP_ATTRIBUTE_TARGET_WAITPKG
1399 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1400 #if !KMP_HAVE_WAITPKG_INTRINSICS
1401   uint32_t timeHi = uint32_t(counter >> 32);
1402   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1403   char flag;
1404   __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1405                    "setb   %0"
1406                    : "=r"(flag)
1407                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1408                    :);
1409   return flag;
1410 #else
1411   return _umwait(hint, counter);
1412 #endif
1413 }
1414 #elif KMP_HAVE_MWAIT
1415 #if KMP_OS_UNIX
1416 #include <pmmintrin.h>
1417 #else
1418 #include <intrin.h>
1419 #endif
1420 #if KMP_OS_UNIX
1421 __attribute__((target("sse3")))
1422 #endif
1423 static inline void
1424 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1425   _mm_monitor(cacheline, extensions, hints);
1426 }
1427 #if KMP_OS_UNIX
1428 __attribute__((target("sse3")))
1429 #endif
1430 static inline void
1431 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1432   _mm_mwait(extensions, hints);
1433 }
1434 #endif // KMP_HAVE_UMWAIT
1435 
1436 /* ------------------------------------------------------------------------ */
1437 /* Support datatypes for the orphaned construct nesting checks.             */
1438 /* ------------------------------------------------------------------------ */
1439 
1440 enum cons_type {
1441   ct_none,
1442   ct_parallel,
1443   ct_pdo,
1444   ct_pdo_ordered,
1445   ct_psections,
1446   ct_psingle,
1447   ct_critical,
1448   ct_ordered_in_parallel,
1449   ct_ordered_in_pdo,
1450   ct_master,
1451   ct_reduce,
1452   ct_barrier,
1453   ct_masked
1454 };
1455 
1456 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1457 
1458 struct cons_data {
1459   ident_t const *ident;
1460   enum cons_type type;
1461   int prev;
1462   kmp_user_lock_p
1463       name; /* address exclusively for critical section name comparison */
1464 };
1465 
1466 struct cons_header {
1467   int p_top, w_top, s_top;
1468   int stack_size, stack_top;
1469   struct cons_data *stack_data;
1470 };
1471 
1472 struct kmp_region_info {
1473   char *text;
1474   int offset[KMP_MAX_FIELDS];
1475   int length[KMP_MAX_FIELDS];
1476 };
1477 
1478 /* ---------------------------------------------------------------------- */
1479 /* ---------------------------------------------------------------------- */
1480 
1481 #if KMP_OS_WINDOWS
1482 typedef HANDLE kmp_thread_t;
1483 typedef DWORD kmp_key_t;
1484 #endif /* KMP_OS_WINDOWS */
1485 
1486 #if KMP_OS_UNIX
1487 typedef pthread_t kmp_thread_t;
1488 typedef pthread_key_t kmp_key_t;
1489 #endif
1490 
1491 extern kmp_key_t __kmp_gtid_threadprivate_key;
1492 
1493 typedef struct kmp_sys_info {
1494   long maxrss; /* the maximum resident set size utilized (in kilobytes)     */
1495   long minflt; /* the number of page faults serviced without any I/O        */
1496   long majflt; /* the number of page faults serviced that required I/O      */
1497   long nswap; /* the number of times a process was "swapped" out of memory */
1498   long inblock; /* the number of times the file system had to perform input  */
1499   long oublock; /* the number of times the file system had to perform output */
1500   long nvcsw; /* the number of times a context switch was voluntarily      */
1501   long nivcsw; /* the number of times a context switch was forced           */
1502 } kmp_sys_info_t;
1503 
1504 #if USE_ITT_BUILD
1505 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1506 // required type here. Later we will check the type meets requirements.
1507 typedef int kmp_itt_mark_t;
1508 #define KMP_ITT_DEBUG 0
1509 #endif /* USE_ITT_BUILD */
1510 
1511 typedef kmp_int32 kmp_critical_name[8];
1512 
1513 /*!
1514 @ingroup PARALLEL
1515 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1516 The arguments to the outlined function are
1517 @param global_tid the global thread identity of the thread executing the
1518 function.
1519 @param bound_tid  the local identity of the thread executing the function
1520 @param ... pointers to shared variables accessed by the function.
1521 */
1522 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1523 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1524                                  ...);
1525 
1526 /*!
1527 @ingroup THREADPRIVATE
1528 @{
1529 */
1530 /* ---------------------------------------------------------------------------
1531  */
1532 /* Threadprivate initialization/finalization function declarations */
1533 
1534 /*  for non-array objects:  __kmpc_threadprivate_register()  */
1535 
1536 /*!
1537  Pointer to the constructor function.
1538  The first argument is the <tt>this</tt> pointer
1539 */
1540 typedef void *(*kmpc_ctor)(void *);
1541 
1542 /*!
1543  Pointer to the destructor function.
1544  The first argument is the <tt>this</tt> pointer
1545 */
1546 typedef void (*kmpc_dtor)(
1547     void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1548                               compiler */
1549 /*!
1550  Pointer to an alternate constructor.
1551  The first argument is the <tt>this</tt> pointer.
1552 */
1553 typedef void *(*kmpc_cctor)(void *, void *);
1554 
1555 /* for array objects: __kmpc_threadprivate_register_vec() */
1556 /* First arg: "this" pointer */
1557 /* Last arg: number of array elements */
1558 /*!
1559  Array constructor.
1560  First argument is the <tt>this</tt> pointer
1561  Second argument the number of array elements.
1562 */
1563 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1564 /*!
1565  Pointer to the array destructor function.
1566  The first argument is the <tt>this</tt> pointer
1567  Second argument the number of array elements.
1568 */
1569 typedef void (*kmpc_dtor_vec)(void *, size_t);
1570 /*!
1571  Array constructor.
1572  First argument is the <tt>this</tt> pointer
1573  Third argument the number of array elements.
1574 */
1575 typedef void *(*kmpc_cctor_vec)(void *, void *,
1576                                 size_t); /* function unused by compiler */
1577 
1578 /*!
1579 @}
1580 */
1581 
1582 /* keeps tracked of threadprivate cache allocations for cleanup later */
1583 typedef struct kmp_cached_addr {
1584   void **addr; /* address of allocated cache */
1585   void ***compiler_cache; /* pointer to compiler's cache */
1586   void *data; /* pointer to global data */
1587   struct kmp_cached_addr *next; /* pointer to next cached address */
1588 } kmp_cached_addr_t;
1589 
1590 struct private_data {
1591   struct private_data *next; /* The next descriptor in the list      */
1592   void *data; /* The data buffer for this descriptor  */
1593   int more; /* The repeat count for this descriptor */
1594   size_t size; /* The data size for this descriptor    */
1595 };
1596 
1597 struct private_common {
1598   struct private_common *next;
1599   struct private_common *link;
1600   void *gbl_addr;
1601   void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1602   size_t cmn_size;
1603 };
1604 
1605 struct shared_common {
1606   struct shared_common *next;
1607   struct private_data *pod_init;
1608   void *obj_init;
1609   void *gbl_addr;
1610   union {
1611     kmpc_ctor ctor;
1612     kmpc_ctor_vec ctorv;
1613   } ct;
1614   union {
1615     kmpc_cctor cctor;
1616     kmpc_cctor_vec cctorv;
1617   } cct;
1618   union {
1619     kmpc_dtor dtor;
1620     kmpc_dtor_vec dtorv;
1621   } dt;
1622   size_t vec_len;
1623   int is_vec;
1624   size_t cmn_size;
1625 };
1626 
1627 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1628 #define KMP_HASH_TABLE_SIZE                                                    \
1629   (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1630 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1631 #define KMP_HASH(x)                                                            \
1632   ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1633 
1634 struct common_table {
1635   struct private_common *data[KMP_HASH_TABLE_SIZE];
1636 };
1637 
1638 struct shared_table {
1639   struct shared_common *data[KMP_HASH_TABLE_SIZE];
1640 };
1641 
1642 /* ------------------------------------------------------------------------ */
1643 
1644 #if KMP_USE_HIER_SCHED
1645 // Shared barrier data that exists inside a single unit of the scheduling
1646 // hierarchy
1647 typedef struct kmp_hier_private_bdata_t {
1648   kmp_int32 num_active;
1649   kmp_uint64 index;
1650   kmp_uint64 wait_val[2];
1651 } kmp_hier_private_bdata_t;
1652 #endif
1653 
1654 typedef struct kmp_sched_flags {
1655   unsigned ordered : 1;
1656   unsigned nomerge : 1;
1657   unsigned contains_last : 1;
1658 #if KMP_USE_HIER_SCHED
1659   unsigned use_hier : 1;
1660   unsigned unused : 28;
1661 #else
1662   unsigned unused : 29;
1663 #endif
1664 } kmp_sched_flags_t;
1665 
1666 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1667 
1668 #if KMP_STATIC_STEAL_ENABLED
1669 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1670   kmp_int32 count;
1671   kmp_int32 ub;
1672   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1673   kmp_int32 lb;
1674   kmp_int32 st;
1675   kmp_int32 tc;
1676   kmp_lock_t *steal_lock; // lock used for chunk stealing
1677   // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1678   //    a) parm3 is properly aligned and
1679   //    b) all parm1-4 are on the same cache line.
1680   // Because of parm1-4 are used together, performance seems to be better
1681   // if they are on the same cache line (not measured though).
1682 
1683   struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1684     kmp_int32 parm1; //     structures in kmp_dispatch.cpp. This should
1685     kmp_int32 parm2; //     make no real change at least while padding is off.
1686     kmp_int32 parm3;
1687     kmp_int32 parm4;
1688   };
1689 
1690   kmp_uint32 ordered_lower;
1691   kmp_uint32 ordered_upper;
1692 #if KMP_OS_WINDOWS
1693   kmp_int32 last_upper;
1694 #endif /* KMP_OS_WINDOWS */
1695 } dispatch_private_info32_t;
1696 
1697 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1698   kmp_int64 count; // current chunk number for static & static-steal scheduling
1699   kmp_int64 ub; /* upper-bound */
1700   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1701   kmp_int64 lb; /* lower-bound */
1702   kmp_int64 st; /* stride */
1703   kmp_int64 tc; /* trip count (number of iterations) */
1704   kmp_lock_t *steal_lock; // lock used for chunk stealing
1705   /* parm[1-4] are used in different ways by different scheduling algorithms */
1706 
1707   // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1708   //    a) parm3 is properly aligned and
1709   //    b) all parm1-4 are in the same cache line.
1710   // Because of parm1-4 are used together, performance seems to be better
1711   // if they are in the same line (not measured though).
1712 
1713   struct KMP_ALIGN(32) {
1714     kmp_int64 parm1;
1715     kmp_int64 parm2;
1716     kmp_int64 parm3;
1717     kmp_int64 parm4;
1718   };
1719 
1720   kmp_uint64 ordered_lower;
1721   kmp_uint64 ordered_upper;
1722 #if KMP_OS_WINDOWS
1723   kmp_int64 last_upper;
1724 #endif /* KMP_OS_WINDOWS */
1725 } dispatch_private_info64_t;
1726 #else /* KMP_STATIC_STEAL_ENABLED */
1727 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1728   kmp_int32 lb;
1729   kmp_int32 ub;
1730   kmp_int32 st;
1731   kmp_int32 tc;
1732 
1733   kmp_int32 parm1;
1734   kmp_int32 parm2;
1735   kmp_int32 parm3;
1736   kmp_int32 parm4;
1737 
1738   kmp_int32 count;
1739 
1740   kmp_uint32 ordered_lower;
1741   kmp_uint32 ordered_upper;
1742 #if KMP_OS_WINDOWS
1743   kmp_int32 last_upper;
1744 #endif /* KMP_OS_WINDOWS */
1745 } dispatch_private_info32_t;
1746 
1747 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1748   kmp_int64 lb; /* lower-bound */
1749   kmp_int64 ub; /* upper-bound */
1750   kmp_int64 st; /* stride */
1751   kmp_int64 tc; /* trip count (number of iterations) */
1752 
1753   /* parm[1-4] are used in different ways by different scheduling algorithms */
1754   kmp_int64 parm1;
1755   kmp_int64 parm2;
1756   kmp_int64 parm3;
1757   kmp_int64 parm4;
1758 
1759   kmp_int64 count; /* current chunk number for static scheduling */
1760 
1761   kmp_uint64 ordered_lower;
1762   kmp_uint64 ordered_upper;
1763 #if KMP_OS_WINDOWS
1764   kmp_int64 last_upper;
1765 #endif /* KMP_OS_WINDOWS */
1766 } dispatch_private_info64_t;
1767 #endif /* KMP_STATIC_STEAL_ENABLED */
1768 
1769 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1770   union private_info {
1771     dispatch_private_info32_t p32;
1772     dispatch_private_info64_t p64;
1773   } u;
1774   enum sched_type schedule; /* scheduling algorithm */
1775   kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1776   std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1777   kmp_int32 ordered_bumped;
1778   // Stack of buffers for nest of serial regions
1779   struct dispatch_private_info *next;
1780   kmp_int32 type_size; /* the size of types in private_info */
1781 #if KMP_USE_HIER_SCHED
1782   kmp_int32 hier_id;
1783   void *parent; /* hierarchical scheduling parent pointer */
1784 #endif
1785   enum cons_type pushed_ws;
1786 } dispatch_private_info_t;
1787 
1788 typedef struct dispatch_shared_info32 {
1789   /* chunk index under dynamic, number of idle threads under static-steal;
1790      iteration index otherwise */
1791   volatile kmp_uint32 iteration;
1792   volatile kmp_int32 num_done;
1793   volatile kmp_uint32 ordered_iteration;
1794   // Dummy to retain the structure size after making ordered_iteration scalar
1795   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1796 } dispatch_shared_info32_t;
1797 
1798 typedef struct dispatch_shared_info64 {
1799   /* chunk index under dynamic, number of idle threads under static-steal;
1800      iteration index otherwise */
1801   volatile kmp_uint64 iteration;
1802   volatile kmp_int64 num_done;
1803   volatile kmp_uint64 ordered_iteration;
1804   // Dummy to retain the structure size after making ordered_iteration scalar
1805   kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1806 } dispatch_shared_info64_t;
1807 
1808 typedef struct dispatch_shared_info {
1809   union shared_info {
1810     dispatch_shared_info32_t s32;
1811     dispatch_shared_info64_t s64;
1812   } u;
1813   volatile kmp_uint32 buffer_index;
1814   volatile kmp_int32 doacross_buf_idx; // teamwise index
1815   volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1816   kmp_int32 doacross_num_done; // count finished threads
1817 #if KMP_USE_HIER_SCHED
1818   void *hier;
1819 #endif
1820 #if KMP_USE_HWLOC
1821   // When linking with libhwloc, the ORDERED EPCC test slows down on big
1822   // machines (> 48 cores). Performance analysis showed that a cache thrash
1823   // was occurring and this padding helps alleviate the problem.
1824   char padding[64];
1825 #endif
1826 } dispatch_shared_info_t;
1827 
1828 typedef struct kmp_disp {
1829   /* Vector for ORDERED SECTION */
1830   void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1831   /* Vector for END ORDERED SECTION */
1832   void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1833 
1834   dispatch_shared_info_t *th_dispatch_sh_current;
1835   dispatch_private_info_t *th_dispatch_pr_current;
1836 
1837   dispatch_private_info_t *th_disp_buffer;
1838   kmp_uint32 th_disp_index;
1839   kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1840   volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1841   kmp_int64 *th_doacross_info; // info on loop bounds
1842 #if KMP_USE_INTERNODE_ALIGNMENT
1843   char more_padding[INTERNODE_CACHE_LINE];
1844 #endif
1845 } kmp_disp_t;
1846 
1847 /* ------------------------------------------------------------------------ */
1848 /* Barrier stuff */
1849 
1850 /* constants for barrier state update */
1851 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1852 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1853 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1854 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1855 
1856 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1857 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1858 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1859 
1860 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1861 #error "Barrier sleep bit must be smaller than barrier bump bit"
1862 #endif
1863 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1864 #error "Barrier unused bit must be smaller than barrier bump bit"
1865 #endif
1866 
1867 // Constants for release barrier wait state: currently, hierarchical only
1868 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1869 #define KMP_BARRIER_OWN_FLAG                                                   \
1870   1 // Normal state; worker waiting on own b_go flag in release
1871 #define KMP_BARRIER_PARENT_FLAG                                                \
1872   2 // Special state; worker waiting on parent's b_go flag in release
1873 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG                                         \
1874   3 // Special state; tells worker to shift from parent to own b_go
1875 #define KMP_BARRIER_SWITCHING                                                  \
1876   4 // Special state; worker resets appropriate flag on wake-up
1877 
1878 #define KMP_NOT_SAFE_TO_REAP                                                   \
1879   0 // Thread th_reap_state: not safe to reap (tasking)
1880 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1881 
1882 enum barrier_type {
1883   bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1884                            barriers if enabled) */
1885   bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1886 #if KMP_FAST_REDUCTION_BARRIER
1887   bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1888 #endif // KMP_FAST_REDUCTION_BARRIER
1889   bs_last_barrier /* Just a placeholder to mark the end */
1890 };
1891 
1892 // to work with reduction barriers just like with plain barriers
1893 #if !KMP_FAST_REDUCTION_BARRIER
1894 #define bs_reduction_barrier bs_plain_barrier
1895 #endif // KMP_FAST_REDUCTION_BARRIER
1896 
1897 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1898                            bp_linear_bar =
1899                                0, /* Single level (degenerate) tree */
1900                            bp_tree_bar =
1901                                1, /* Balanced tree with branching factor 2^n */
1902                            bp_hyper_bar = 2, /* Hypercube-embedded tree with min
1903                                                 branching factor 2^n */
1904                            bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1905                            bp_last_bar /* Placeholder to mark the end */
1906 } kmp_bar_pat_e;
1907 
1908 #define KMP_BARRIER_ICV_PUSH 1
1909 
1910 /* Record for holding the values of the internal controls stack records */
1911 typedef struct kmp_internal_control {
1912   int serial_nesting_level; /* corresponds to the value of the
1913                                th_team_serialized field */
1914   kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1915                        thread) */
1916   kmp_int8
1917       bt_set; /* internal control for whether blocktime is explicitly set */
1918   int blocktime; /* internal control for blocktime */
1919 #if KMP_USE_MONITOR
1920   int bt_intervals; /* internal control for blocktime intervals */
1921 #endif
1922   int nproc; /* internal control for #threads for next parallel region (per
1923                 thread) */
1924   int thread_limit; /* internal control for thread-limit-var */
1925   int max_active_levels; /* internal control for max_active_levels */
1926   kmp_r_sched_t
1927       sched; /* internal control for runtime schedule {sched,chunk} pair */
1928   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
1929   kmp_int32 default_device; /* internal control for default device */
1930   struct kmp_internal_control *next;
1931 } kmp_internal_control_t;
1932 
1933 static inline void copy_icvs(kmp_internal_control_t *dst,
1934                              kmp_internal_control_t *src) {
1935   *dst = *src;
1936 }
1937 
1938 /* Thread barrier needs volatile barrier fields */
1939 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1940   // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1941   // uses of it). It is not explicitly aligned below, because we *don't* want
1942   // it to be padded -- instead, we fit b_go into the same cache line with
1943   // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1944   kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1945   // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1946   // same NGO store
1947   volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1948   KMP_ALIGN_CACHE volatile kmp_uint64
1949       b_arrived; // STATE => task reached synch point.
1950   kmp_uint32 *skip_per_level;
1951   kmp_uint32 my_level;
1952   kmp_int32 parent_tid;
1953   kmp_int32 old_tid;
1954   kmp_uint32 depth;
1955   struct kmp_bstate *parent_bar;
1956   kmp_team_t *team;
1957   kmp_uint64 leaf_state;
1958   kmp_uint32 nproc;
1959   kmp_uint8 base_leaf_kids;
1960   kmp_uint8 leaf_kids;
1961   kmp_uint8 offset;
1962   kmp_uint8 wait_flag;
1963   kmp_uint8 use_oncore_barrier;
1964 #if USE_DEBUGGER
1965   // The following field is intended for the debugger solely. Only the worker
1966   // thread itself accesses this field: the worker increases it by 1 when it
1967   // arrives to a barrier.
1968   KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1969 #endif /* USE_DEBUGGER */
1970 } kmp_bstate_t;
1971 
1972 union KMP_ALIGN_CACHE kmp_barrier_union {
1973   double b_align; /* use worst case alignment */
1974   char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1975   kmp_bstate_t bb;
1976 };
1977 
1978 typedef union kmp_barrier_union kmp_balign_t;
1979 
1980 /* Team barrier needs only non-volatile arrived counter */
1981 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1982   double b_align; /* use worst case alignment */
1983   char b_pad[CACHE_LINE];
1984   struct {
1985     kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1986 #if USE_DEBUGGER
1987     // The following two fields are indended for the debugger solely. Only
1988     // primary thread of the team accesses these fields: the first one is
1989     // increased by 1 when the primary thread arrives to a barrier, the second
1990     // one is increased by one when all the threads arrived.
1991     kmp_uint b_master_arrived;
1992     kmp_uint b_team_arrived;
1993 #endif
1994   };
1995 };
1996 
1997 typedef union kmp_barrier_team_union kmp_balign_team_t;
1998 
1999 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2000    threads when a condition changes.  This is to workaround an NPTL bug where
2001    padding was added to pthread_cond_t which caused the initialization routine
2002    to write outside of the structure if compiled on pre-NPTL threads.  */
2003 #if KMP_OS_WINDOWS
2004 typedef struct kmp_win32_mutex {
2005   /* The Lock */
2006   CRITICAL_SECTION cs;
2007 } kmp_win32_mutex_t;
2008 
2009 typedef struct kmp_win32_cond {
2010   /* Count of the number of waiters. */
2011   int waiters_count_;
2012 
2013   /* Serialize access to <waiters_count_> */
2014   kmp_win32_mutex_t waiters_count_lock_;
2015 
2016   /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2017   int release_count_;
2018 
2019   /* Keeps track of the current "generation" so that we don't allow */
2020   /* one thread to steal all the "releases" from the broadcast. */
2021   int wait_generation_count_;
2022 
2023   /* A manual-reset event that's used to block and release waiting threads. */
2024   HANDLE event_;
2025 } kmp_win32_cond_t;
2026 #endif
2027 
2028 #if KMP_OS_UNIX
2029 
2030 union KMP_ALIGN_CACHE kmp_cond_union {
2031   double c_align;
2032   char c_pad[CACHE_LINE];
2033   pthread_cond_t c_cond;
2034 };
2035 
2036 typedef union kmp_cond_union kmp_cond_align_t;
2037 
2038 union KMP_ALIGN_CACHE kmp_mutex_union {
2039   double m_align;
2040   char m_pad[CACHE_LINE];
2041   pthread_mutex_t m_mutex;
2042 };
2043 
2044 typedef union kmp_mutex_union kmp_mutex_align_t;
2045 
2046 #endif /* KMP_OS_UNIX */
2047 
2048 typedef struct kmp_desc_base {
2049   void *ds_stackbase;
2050   size_t ds_stacksize;
2051   int ds_stackgrow;
2052   kmp_thread_t ds_thread;
2053   volatile int ds_tid;
2054   int ds_gtid;
2055 #if KMP_OS_WINDOWS
2056   volatile int ds_alive;
2057   DWORD ds_thread_id;
2058 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2059    However, debugger support (libomp_db) cannot work with handles, because they
2060    uncomparable. For example, debugger requests info about thread with handle h.
2061    h is valid within debugger process, and meaningless within debugee process.
2062    Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2063    within debugee process, but it is a *new* handle which does *not* equal to
2064    any other handle in debugee... The only way to compare handles is convert
2065    them to system-wide ids. GetThreadId() function is available only in
2066    Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2067    on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2068    thread id by call to GetCurrentThreadId() from within the thread and save it
2069    to let libomp_db identify threads.  */
2070 #endif /* KMP_OS_WINDOWS */
2071 } kmp_desc_base_t;
2072 
2073 typedef union KMP_ALIGN_CACHE kmp_desc {
2074   double ds_align; /* use worst case alignment */
2075   char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2076   kmp_desc_base_t ds;
2077 } kmp_desc_t;
2078 
2079 typedef struct kmp_local {
2080   volatile int this_construct; /* count of single's encountered by thread */
2081   void *reduce_data;
2082 #if KMP_USE_BGET
2083   void *bget_data;
2084   void *bget_list;
2085 #if !USE_CMP_XCHG_FOR_BGET
2086 #ifdef USE_QUEUING_LOCK_FOR_BGET
2087   kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2088 #else
2089   kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2090 // bootstrap lock so we can use it at library
2091 // shutdown.
2092 #endif /* USE_LOCK_FOR_BGET */
2093 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2094 #endif /* KMP_USE_BGET */
2095 
2096   PACKED_REDUCTION_METHOD_T
2097   packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2098                               __kmpc_end_reduce*() */
2099 
2100 } kmp_local_t;
2101 
2102 #define KMP_CHECK_UPDATE(a, b)                                                 \
2103   if ((a) != (b))                                                              \
2104   (a) = (b)
2105 #define KMP_CHECK_UPDATE_SYNC(a, b)                                            \
2106   if ((a) != (b))                                                              \
2107   TCW_SYNC_PTR((a), (b))
2108 
2109 #define get__blocktime(xteam, xtid)                                            \
2110   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2111 #define get__bt_set(xteam, xtid)                                               \
2112   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2113 #if KMP_USE_MONITOR
2114 #define get__bt_intervals(xteam, xtid)                                         \
2115   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2116 #endif
2117 
2118 #define get__dynamic_2(xteam, xtid)                                            \
2119   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2120 #define get__nproc_2(xteam, xtid)                                              \
2121   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2122 #define get__sched_2(xteam, xtid)                                              \
2123   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2124 
2125 #define set__blocktime_team(xteam, xtid, xval)                                 \
2126   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) =     \
2127        (xval))
2128 
2129 #if KMP_USE_MONITOR
2130 #define set__bt_intervals_team(xteam, xtid, xval)                              \
2131   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) =  \
2132        (xval))
2133 #endif
2134 
2135 #define set__bt_set_team(xteam, xtid, xval)                                    \
2136   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2137 
2138 #define set__dynamic(xthread, xval)                                            \
2139   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2140 #define get__dynamic(xthread)                                                  \
2141   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2142 
2143 #define set__nproc(xthread, xval)                                              \
2144   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2145 
2146 #define set__thread_limit(xthread, xval)                                       \
2147   (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2148 
2149 #define set__max_active_levels(xthread, xval)                                  \
2150   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2151 
2152 #define get__max_active_levels(xthread)                                        \
2153   ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2154 
2155 #define set__sched(xthread, xval)                                              \
2156   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2157 
2158 #define set__proc_bind(xthread, xval)                                          \
2159   (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2160 #define get__proc_bind(xthread)                                                \
2161   ((xthread)->th.th_current_task->td_icvs.proc_bind)
2162 
2163 // OpenMP tasking data structures
2164 
2165 typedef enum kmp_tasking_mode {
2166   tskm_immediate_exec = 0,
2167   tskm_extra_barrier = 1,
2168   tskm_task_teams = 2,
2169   tskm_max = 2
2170 } kmp_tasking_mode_t;
2171 
2172 extern kmp_tasking_mode_t
2173     __kmp_tasking_mode; /* determines how/when to execute tasks */
2174 extern int __kmp_task_stealing_constraint;
2175 extern int __kmp_enable_task_throttling;
2176 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2177 // specified, defaults to 0 otherwise
2178 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2179 extern kmp_int32 __kmp_max_task_priority;
2180 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2181 extern kmp_uint64 __kmp_taskloop_min_tasks;
2182 
2183 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2184    taskdata first */
2185 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2186 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2187 
2188 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2189 // were spawned and queued since the previous barrier release.
2190 #define KMP_TASKING_ENABLED(task_team)                                         \
2191   (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2192 /*!
2193 @ingroup BASIC_TYPES
2194 @{
2195 */
2196 
2197 /*!
2198  */
2199 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2200 
2201 typedef union kmp_cmplrdata {
2202   kmp_int32 priority; /**< priority specified by user for the task */
2203   kmp_routine_entry_t
2204       destructors; /* pointer to function to invoke deconstructors of
2205                       firstprivate C++ objects */
2206   /* future data */
2207 } kmp_cmplrdata_t;
2208 
2209 /*  sizeof_kmp_task_t passed as arg to kmpc_omp_task call  */
2210 /*!
2211  */
2212 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2213   void *shareds; /**< pointer to block of pointers to shared vars   */
2214   kmp_routine_entry_t
2215       routine; /**< pointer to routine to call for executing task */
2216   kmp_int32 part_id; /**< part id for the task                          */
2217   kmp_cmplrdata_t
2218       data1; /* Two known optional additions: destructors and priority */
2219   kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2220   /* future data */
2221   /*  private vars  */
2222 } kmp_task_t;
2223 
2224 /*!
2225 @}
2226 */
2227 
2228 typedef struct kmp_taskgroup {
2229   std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2230   std::atomic<kmp_int32>
2231       cancel_request; // request for cancellation of this taskgroup
2232   struct kmp_taskgroup *parent; // parent taskgroup
2233   // Block of data to perform task reduction
2234   void *reduce_data; // reduction related info
2235   kmp_int32 reduce_num_data; // number of data items to reduce
2236   uintptr_t *gomp_data; // gomp reduction data
2237 } kmp_taskgroup_t;
2238 
2239 // forward declarations
2240 typedef union kmp_depnode kmp_depnode_t;
2241 typedef struct kmp_depnode_list kmp_depnode_list_t;
2242 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2243 
2244 #define KMP_DEP_IN 0x1
2245 #define KMP_DEP_OUT 0x2
2246 #define KMP_DEP_INOUT 0x3
2247 #define KMP_DEP_MTX 0x4
2248 #define KMP_DEP_SET 0x8
2249 // Compiler sends us this info:
2250 typedef struct kmp_depend_info {
2251   kmp_intptr_t base_addr;
2252   size_t len;
2253   union {
2254     kmp_uint8 flag;
2255     struct {
2256       unsigned in : 1;
2257       unsigned out : 1;
2258       unsigned mtx : 1;
2259       unsigned set : 1;
2260     } flags;
2261   };
2262 } kmp_depend_info_t;
2263 
2264 // Internal structures to work with task dependencies:
2265 struct kmp_depnode_list {
2266   kmp_depnode_t *node;
2267   kmp_depnode_list_t *next;
2268 };
2269 
2270 // Max number of mutexinoutset dependencies per node
2271 #define MAX_MTX_DEPS 4
2272 
2273 typedef struct kmp_base_depnode {
2274   kmp_depnode_list_t *successors; /* used under lock */
2275   kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2276   kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2277   kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2278   kmp_lock_t lock; /* guards shared fields: task, successors */
2279 #if KMP_SUPPORT_GRAPH_OUTPUT
2280   kmp_uint32 id;
2281 #endif
2282   std::atomic<kmp_int32> npredecessors;
2283   std::atomic<kmp_int32> nrefs;
2284 } kmp_base_depnode_t;
2285 
2286 union KMP_ALIGN_CACHE kmp_depnode {
2287   double dn_align; /* use worst case alignment */
2288   char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2289   kmp_base_depnode_t dn;
2290 };
2291 
2292 struct kmp_dephash_entry {
2293   kmp_intptr_t addr;
2294   kmp_depnode_t *last_out;
2295   kmp_depnode_list_t *last_set;
2296   kmp_depnode_list_t *prev_set;
2297   kmp_uint8 last_flag;
2298   kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2299   kmp_dephash_entry_t *next_in_bucket;
2300 };
2301 
2302 typedef struct kmp_dephash {
2303   kmp_dephash_entry_t **buckets;
2304   size_t size;
2305   size_t generation;
2306   kmp_uint32 nelements;
2307   kmp_uint32 nconflicts;
2308 } kmp_dephash_t;
2309 
2310 typedef struct kmp_task_affinity_info {
2311   kmp_intptr_t base_addr;
2312   size_t len;
2313   struct {
2314     bool flag1 : 1;
2315     bool flag2 : 1;
2316     kmp_int32 reserved : 30;
2317   } flags;
2318 } kmp_task_affinity_info_t;
2319 
2320 typedef enum kmp_event_type_t {
2321   KMP_EVENT_UNINITIALIZED = 0,
2322   KMP_EVENT_ALLOW_COMPLETION = 1
2323 } kmp_event_type_t;
2324 
2325 typedef struct {
2326   kmp_event_type_t type;
2327   kmp_tas_lock_t lock;
2328   union {
2329     kmp_task_t *task;
2330   } ed;
2331 } kmp_event_t;
2332 
2333 #ifdef BUILD_TIED_TASK_STACK
2334 
2335 /* Tied Task stack definitions */
2336 typedef struct kmp_stack_block {
2337   kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2338   struct kmp_stack_block *sb_next;
2339   struct kmp_stack_block *sb_prev;
2340 } kmp_stack_block_t;
2341 
2342 typedef struct kmp_task_stack {
2343   kmp_stack_block_t ts_first_block; // first block of stack entries
2344   kmp_taskdata_t **ts_top; // pointer to the top of stack
2345   kmp_int32 ts_entries; // number of entries on the stack
2346 } kmp_task_stack_t;
2347 
2348 #endif // BUILD_TIED_TASK_STACK
2349 
2350 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2351   /* Compiler flags */ /* Total compiler flags must be 16 bits */
2352   unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2353   unsigned final : 1; /* task is final(1) so execute immediately */
2354   unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2355                               code path */
2356   unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2357                                      invoke destructors from the runtime */
2358   unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2359                          context of the RTL) */
2360   unsigned priority_specified : 1; /* set if the compiler provides priority
2361                                       setting for the task */
2362   unsigned detachable : 1; /* 1 == can detach */
2363   unsigned hidden_helper : 1; /* 1 == hidden helper task */
2364   unsigned reserved : 8; /* reserved for compiler use */
2365 
2366   /* Library flags */ /* Total library flags must be 16 bits */
2367   unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2368   unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2369   unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2370   // (1) or may be deferred (0)
2371   unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2372   // (0) [>= 2 threads]
2373   /* If either team_serial or tasking_ser is set, task team may be NULL */
2374   /* Task State Flags: */
2375   unsigned started : 1; /* 1==started, 0==not started     */
2376   unsigned executing : 1; /* 1==executing, 0==not executing */
2377   unsigned complete : 1; /* 1==complete, 0==not complete   */
2378   unsigned freed : 1; /* 1==freed, 0==allocated        */
2379   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2380   unsigned reserved31 : 7; /* reserved for library use */
2381 
2382 } kmp_tasking_flags_t;
2383 
2384 struct kmp_taskdata { /* aligned during dynamic allocation       */
2385   kmp_int32 td_task_id; /* id, assigned by debugger                */
2386   kmp_tasking_flags_t td_flags; /* task flags                              */
2387   kmp_team_t *td_team; /* team for this task                      */
2388   kmp_info_p *td_alloc_thread; /* thread that allocated data structures   */
2389   /* Currently not used except for perhaps IDB */
2390   kmp_taskdata_t *td_parent; /* parent task                             */
2391   kmp_int32 td_level; /* task nesting level                      */
2392   std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2393   ident_t *td_ident; /* task identifier                         */
2394   // Taskwait data.
2395   ident_t *td_taskwait_ident;
2396   kmp_uint32 td_taskwait_counter;
2397   kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2398   KMP_ALIGN_CACHE kmp_internal_control_t
2399       td_icvs; /* Internal control variables for the task */
2400   KMP_ALIGN_CACHE std::atomic<kmp_int32>
2401       td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2402                                    deallocated */
2403   std::atomic<kmp_int32>
2404       td_incomplete_child_tasks; /* Child tasks not yet complete */
2405   kmp_taskgroup_t
2406       *td_taskgroup; // Each task keeps pointer to its current taskgroup
2407   kmp_dephash_t
2408       *td_dephash; // Dependencies for children tasks are tracked from here
2409   kmp_depnode_t
2410       *td_depnode; // Pointer to graph node if this task has dependencies
2411   kmp_task_team_t *td_task_team;
2412   // The global thread id of the encountering thread. We need it because when a
2413   // regular task depends on a hidden helper task, and the hidden helper task
2414   // is finished on a hidden helper thread, it will call __kmp_release_deps to
2415   // release all dependences. If now the task is a regular task, we need to pass
2416   // the encountering gtid such that the task will be picked up and executed by
2417   // its encountering team instead of hidden helper team.
2418   kmp_int32 encountering_gtid;
2419   size_t td_size_alloc; // Size of task structure, including shareds etc.
2420 #if defined(KMP_GOMP_COMPAT)
2421   // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2422   kmp_int32 td_size_loop_bounds;
2423 #endif
2424   kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2425 #if defined(KMP_GOMP_COMPAT)
2426   // GOMP sends in a copy function for copy constructors
2427   void (*td_copy_func)(void *, void *);
2428 #endif
2429   kmp_event_t td_allow_completion_event;
2430 #if OMPT_SUPPORT
2431   ompt_task_info_t ompt_task_info;
2432 #endif
2433 }; // struct kmp_taskdata
2434 
2435 // Make sure padding above worked
2436 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2437 
2438 // Data for task team but per thread
2439 typedef struct kmp_base_thread_data {
2440   kmp_info_p *td_thr; // Pointer back to thread info
2441   // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2442   // queued?
2443   kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2444   kmp_taskdata_t *
2445       *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2446   kmp_int32 td_deque_size; // Size of deck
2447   kmp_uint32 td_deque_head; // Head of deque (will wrap)
2448   kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2449   kmp_int32 td_deque_ntasks; // Number of tasks in deque
2450   // GEH: shouldn't this be volatile since used in while-spin?
2451   kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2452 #ifdef BUILD_TIED_TASK_STACK
2453   kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2454 // scheduling constraint
2455 #endif // BUILD_TIED_TASK_STACK
2456 } kmp_base_thread_data_t;
2457 
2458 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2459 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2460 
2461 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2462 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2463 
2464 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2465   kmp_base_thread_data_t td;
2466   double td_align; /* use worst case alignment */
2467   char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2468 } kmp_thread_data_t;
2469 
2470 // Data for task teams which are used when tasking is enabled for the team
2471 typedef struct kmp_base_task_team {
2472   kmp_bootstrap_lock_t
2473       tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2474   /* must be bootstrap lock since used at library shutdown*/
2475   kmp_task_team_t *tt_next; /* For linking the task team free list */
2476   kmp_thread_data_t
2477       *tt_threads_data; /* Array of per-thread structures for task team */
2478   /* Data survives task team deallocation */
2479   kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2480                                executing this team? */
2481   /* TRUE means tt_threads_data is set up and initialized */
2482   kmp_int32 tt_nproc; /* #threads in team           */
2483   kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2484   kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2485   kmp_int32 tt_untied_task_encountered;
2486   // There is hidden helper thread encountered in this task team so that we must
2487   // wait when waiting on task team
2488   kmp_int32 tt_hidden_helper_task_encountered;
2489 
2490   KMP_ALIGN_CACHE
2491   std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2492 
2493   KMP_ALIGN_CACHE
2494   volatile kmp_uint32
2495       tt_active; /* is the team still actively executing tasks */
2496 } kmp_base_task_team_t;
2497 
2498 union KMP_ALIGN_CACHE kmp_task_team {
2499   kmp_base_task_team_t tt;
2500   double tt_align; /* use worst case alignment */
2501   char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2502 };
2503 
2504 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2505 // Free lists keep same-size free memory slots for fast memory allocation
2506 // routines
2507 typedef struct kmp_free_list {
2508   void *th_free_list_self; // Self-allocated tasks free list
2509   void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2510   // threads
2511   void *th_free_list_other; // Non-self free list (to be returned to owner's
2512   // sync list)
2513 } kmp_free_list_t;
2514 #endif
2515 #if KMP_NESTED_HOT_TEAMS
2516 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2517 // are not put in teams pool, and they don't put threads in threads pool.
2518 typedef struct kmp_hot_team_ptr {
2519   kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2520   kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2521 } kmp_hot_team_ptr_t;
2522 #endif
2523 typedef struct kmp_teams_size {
2524   kmp_int32 nteams; // number of teams in a league
2525   kmp_int32 nth; // number of threads in each team of the league
2526 } kmp_teams_size_t;
2527 
2528 // This struct stores a thread that acts as a "root" for a contention
2529 // group. Contention groups are rooted at kmp_root threads, but also at
2530 // each primary thread of each team created in the teams construct.
2531 // This struct therefore also stores a thread_limit associated with
2532 // that contention group, and a counter to track the number of threads
2533 // active in that contention group. Each thread has a list of these: CG
2534 // root threads have an entry in their list in which cg_root refers to
2535 // the thread itself, whereas other workers in the CG will have a
2536 // single entry where cg_root is same as the entry containing their CG
2537 // root. When a thread encounters a teams construct, it will add a new
2538 // entry to the front of its list, because it now roots a new CG.
2539 typedef struct kmp_cg_root {
2540   kmp_info_p *cg_root; // "root" thread for a contention group
2541   // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2542   // thread_limit clause for teams primary threads
2543   kmp_int32 cg_thread_limit;
2544   kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2545   struct kmp_cg_root *up; // pointer to higher level CG root in list
2546 } kmp_cg_root_t;
2547 
2548 // OpenMP thread data structures
2549 
2550 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2551   /* Start with the readonly data which is cache aligned and padded. This is
2552      written before the thread starts working by the primary thread. Uber
2553      masters may update themselves later. Usage does not consider serialized
2554      regions.  */
2555   kmp_desc_t th_info;
2556   kmp_team_p *th_team; /* team we belong to */
2557   kmp_root_p *th_root; /* pointer to root of task hierarchy */
2558   kmp_info_p *th_next_pool; /* next available thread in the pool */
2559   kmp_disp_t *th_dispatch; /* thread's dispatch data */
2560   int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2561 
2562   /* The following are cached from the team info structure */
2563   /* TODO use these in more places as determined to be needed via profiling */
2564   int th_team_nproc; /* number of threads in a team */
2565   kmp_info_p *th_team_master; /* the team's primary thread */
2566   int th_team_serialized; /* team is serialized */
2567   microtask_t th_teams_microtask; /* save entry address for teams construct */
2568   int th_teams_level; /* save initial level of teams construct */
2569 /* it is 0 on device but may be any on host */
2570 
2571 /* The blocktime info is copied from the team struct to the thread struct */
2572 /* at the start of a barrier, and the values stored in the team are used  */
2573 /* at points in the code where the team struct is no longer guaranteed    */
2574 /* to exist (from the POV of worker threads).                             */
2575 #if KMP_USE_MONITOR
2576   int th_team_bt_intervals;
2577   int th_team_bt_set;
2578 #else
2579   kmp_uint64 th_team_bt_intervals;
2580 #endif
2581 
2582 #if KMP_AFFINITY_SUPPORTED
2583   kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2584 #endif
2585   omp_allocator_handle_t th_def_allocator; /* default allocator */
2586   /* The data set by the primary thread at reinit, then R/W by the worker */
2587   KMP_ALIGN_CACHE int
2588       th_set_nproc; /* if > 0, then only use this request for the next fork */
2589 #if KMP_NESTED_HOT_TEAMS
2590   kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2591 #endif
2592   kmp_proc_bind_t
2593       th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2594   kmp_teams_size_t
2595       th_teams_size; /* number of teams/threads in teams construct */
2596 #if KMP_AFFINITY_SUPPORTED
2597   int th_current_place; /* place currently bound to */
2598   int th_new_place; /* place to bind to in par reg */
2599   int th_first_place; /* first place in partition */
2600   int th_last_place; /* last place in partition */
2601 #endif
2602   int th_prev_level; /* previous level for affinity format */
2603   int th_prev_num_threads; /* previous num_threads for affinity format */
2604 #if USE_ITT_BUILD
2605   kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2606   kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2607   kmp_uint64 th_frame_time; /* frame timestamp */
2608 #endif /* USE_ITT_BUILD */
2609   kmp_local_t th_local;
2610   struct private_common *th_pri_head;
2611 
2612   /* Now the data only used by the worker (after initial allocation) */
2613   /* TODO the first serial team should actually be stored in the info_t
2614      structure.  this will help reduce initial allocation overhead */
2615   KMP_ALIGN_CACHE kmp_team_p
2616       *th_serial_team; /*serialized team held in reserve*/
2617 
2618 #if OMPT_SUPPORT
2619   ompt_thread_info_t ompt_thread_info;
2620 #endif
2621 
2622   /* The following are also read by the primary thread during reinit */
2623   struct common_table *th_pri_common;
2624 
2625   volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2626   /* while awaiting queuing lock acquire */
2627 
2628   volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2629 
2630   ident_t *th_ident;
2631   unsigned th_x; // Random number generator data
2632   unsigned th_a; // Random number generator data
2633 
2634   /* Tasking-related data for the thread */
2635   kmp_task_team_t *th_task_team; // Task team struct
2636   kmp_taskdata_t *th_current_task; // Innermost Task being executed
2637   kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2638   kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2639   // at nested levels
2640   kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2641   kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2642   kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2643   // tasking, thus safe to reap
2644 
2645   /* More stuff for keeping track of active/sleeping threads (this part is
2646      written by the worker thread) */
2647   kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2648   int th_active; // ! sleeping; 32 bits for TCR/TCW
2649   struct cons_header *th_cons; // used for consistency check
2650 #if KMP_USE_HIER_SCHED
2651   // used for hierarchical scheduling
2652   kmp_hier_private_bdata_t *th_hier_bar_data;
2653 #endif
2654 
2655   /* Add the syncronizing data which is cache aligned and padded. */
2656   KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2657 
2658   KMP_ALIGN_CACHE volatile kmp_int32
2659       th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2660 
2661 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2662 #define NUM_LISTS 4
2663   kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2664 // allocation routines
2665 #endif
2666 
2667 #if KMP_OS_WINDOWS
2668   kmp_win32_cond_t th_suspend_cv;
2669   kmp_win32_mutex_t th_suspend_mx;
2670   std::atomic<int> th_suspend_init;
2671 #endif
2672 #if KMP_OS_UNIX
2673   kmp_cond_align_t th_suspend_cv;
2674   kmp_mutex_align_t th_suspend_mx;
2675   std::atomic<int> th_suspend_init_count;
2676 #endif
2677 
2678 #if USE_ITT_BUILD
2679   kmp_itt_mark_t th_itt_mark_single;
2680 // alignment ???
2681 #endif /* USE_ITT_BUILD */
2682 #if KMP_STATS_ENABLED
2683   kmp_stats_list *th_stats;
2684 #endif
2685 #if KMP_OS_UNIX
2686   std::atomic<bool> th_blocking;
2687 #endif
2688   kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2689 } kmp_base_info_t;
2690 
2691 typedef union KMP_ALIGN_CACHE kmp_info {
2692   double th_align; /* use worst case alignment */
2693   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2694   kmp_base_info_t th;
2695 } kmp_info_t;
2696 
2697 // OpenMP thread team data structures
2698 
2699 typedef struct kmp_base_data {
2700   volatile kmp_uint32 t_value;
2701 } kmp_base_data_t;
2702 
2703 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2704   double dt_align; /* use worst case alignment */
2705   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2706   kmp_base_data_t dt;
2707 } kmp_sleep_team_t;
2708 
2709 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2710   double dt_align; /* use worst case alignment */
2711   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2712   kmp_base_data_t dt;
2713 } kmp_ordered_team_t;
2714 
2715 typedef int (*launch_t)(int gtid);
2716 
2717 /* Minimum number of ARGV entries to malloc if necessary */
2718 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2719 
2720 // Set up how many argv pointers will fit in cache lines containing
2721 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2722 // larger value for more space between the primary write/worker read section and
2723 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2724 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2725 #define KMP_INLINE_ARGV_BYTES                                                  \
2726   (4 * CACHE_LINE -                                                            \
2727    ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) +               \
2728      sizeof(kmp_int16) + sizeof(kmp_uint32)) %                                 \
2729     CACHE_LINE))
2730 #else
2731 #define KMP_INLINE_ARGV_BYTES                                                  \
2732   (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2733 #endif
2734 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2735 
2736 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2737   // Synchronization Data
2738   // ---------------------------------------------------------------------------
2739   KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2740   kmp_balign_team_t t_bar[bs_last_barrier];
2741   std::atomic<int> t_construct; // count of single directive encountered by team
2742   char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2743 
2744   // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2745   std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2746   std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2747 
2748   // Primary thread only
2749   // ---------------------------------------------------------------------------
2750   KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2751   int t_master_this_cons; // "this_construct" single counter of primary thread
2752   // in parent team
2753   ident_t *t_ident; // if volatile, have to change too much other crud to
2754   // volatile too
2755   kmp_team_p *t_parent; // parent team
2756   kmp_team_p *t_next_pool; // next free team in the team pool
2757   kmp_disp_t *t_dispatch; // thread's dispatch data
2758   kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2759   kmp_proc_bind_t t_proc_bind; // bind type for par region
2760 #if USE_ITT_BUILD
2761   kmp_uint64 t_region_time; // region begin timestamp
2762 #endif /* USE_ITT_BUILD */
2763 
2764   // Primary thread write, workers read
2765   // --------------------------------------------------------------------------
2766   KMP_ALIGN_CACHE void **t_argv;
2767   int t_argc;
2768   int t_nproc; // number of threads in team
2769   microtask_t t_pkfn;
2770   launch_t t_invoke; // procedure to launch the microtask
2771 
2772 #if OMPT_SUPPORT
2773   ompt_team_info_t ompt_team_info;
2774   ompt_lw_taskteam_t *ompt_serialized_team_info;
2775 #endif
2776 
2777 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2778   kmp_int8 t_fp_control_saved;
2779   kmp_int8 t_pad2b;
2780   kmp_int16 t_x87_fpu_control_word; // FP control regs
2781   kmp_uint32 t_mxcsr;
2782 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2783 
2784   void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2785 
2786   KMP_ALIGN_CACHE kmp_info_t **t_threads;
2787   kmp_taskdata_t
2788       *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2789   int t_level; // nested parallel level
2790 
2791   KMP_ALIGN_CACHE int t_max_argc;
2792   int t_max_nproc; // max threads this team can handle (dynamically expandable)
2793   int t_serialized; // levels deep of serialized teams
2794   dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2795   int t_id; // team's id, assigned by debugger.
2796   int t_active_level; // nested active parallel level
2797   kmp_r_sched_t t_sched; // run-time schedule for the team
2798 #if KMP_AFFINITY_SUPPORTED
2799   int t_first_place; // first & last place in parent thread's partition.
2800   int t_last_place; // Restore these values to primary thread after par region.
2801 #endif // KMP_AFFINITY_SUPPORTED
2802   int t_display_affinity;
2803   int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2804   // omp_set_num_threads() call
2805   omp_allocator_handle_t t_def_allocator; /* default allocator */
2806 
2807 // Read/write by workers as well
2808 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2809   // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2810   // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2811   // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2812   // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2813   char dummy_padding[1024];
2814 #endif
2815   // Internal control stack for additional nested teams.
2816   KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2817   // for SERIALIZED teams nested 2 or more levels deep
2818   // typed flag to store request state of cancellation
2819   std::atomic<kmp_int32> t_cancel_request;
2820   int t_master_active; // save on fork, restore on join
2821   void *t_copypriv_data; // team specific pointer to copyprivate data array
2822 #if KMP_OS_WINDOWS
2823   std::atomic<kmp_uint32> t_copyin_counter;
2824 #endif
2825 #if USE_ITT_BUILD
2826   void *t_stack_id; // team specific stack stitching id (for ittnotify)
2827 #endif /* USE_ITT_BUILD */
2828 } kmp_base_team_t;
2829 
2830 union KMP_ALIGN_CACHE kmp_team {
2831   kmp_base_team_t t;
2832   double t_align; /* use worst case alignment */
2833   char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2834 };
2835 
2836 typedef union KMP_ALIGN_CACHE kmp_time_global {
2837   double dt_align; /* use worst case alignment */
2838   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2839   kmp_base_data_t dt;
2840 } kmp_time_global_t;
2841 
2842 typedef struct kmp_base_global {
2843   /* cache-aligned */
2844   kmp_time_global_t g_time;
2845 
2846   /* non cache-aligned */
2847   volatile int g_abort;
2848   volatile int g_done;
2849 
2850   int g_dynamic;
2851   enum dynamic_mode g_dynamic_mode;
2852 } kmp_base_global_t;
2853 
2854 typedef union KMP_ALIGN_CACHE kmp_global {
2855   kmp_base_global_t g;
2856   double g_align; /* use worst case alignment */
2857   char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2858 } kmp_global_t;
2859 
2860 typedef struct kmp_base_root {
2861   // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2862   // (r_in_parallel>= 0)
2863   // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2864   // the synch overhead or keeping r_active
2865   volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2866   // keeps a count of active parallel regions per root
2867   std::atomic<int> r_in_parallel;
2868   // GEH: This is misnamed, should be r_active_levels
2869   kmp_team_t *r_root_team;
2870   kmp_team_t *r_hot_team;
2871   kmp_info_t *r_uber_thread;
2872   kmp_lock_t r_begin_lock;
2873   volatile int r_begin;
2874   int r_blocktime; /* blocktime for this root and descendants */
2875 #if KMP_AFFINITY_SUPPORTED
2876   int r_affinity_assigned;
2877 #endif // KMP_AFFINITY_SUPPORTED
2878 } kmp_base_root_t;
2879 
2880 typedef union KMP_ALIGN_CACHE kmp_root {
2881   kmp_base_root_t r;
2882   double r_align; /* use worst case alignment */
2883   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2884 } kmp_root_t;
2885 
2886 struct fortran_inx_info {
2887   kmp_int32 data;
2888 };
2889 
2890 /* ------------------------------------------------------------------------ */
2891 
2892 extern int __kmp_settings;
2893 extern int __kmp_duplicate_library_ok;
2894 #if USE_ITT_BUILD
2895 extern int __kmp_forkjoin_frames;
2896 extern int __kmp_forkjoin_frames_mode;
2897 #endif
2898 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2899 extern int __kmp_determ_red;
2900 
2901 #ifdef KMP_DEBUG
2902 extern int kmp_a_debug;
2903 extern int kmp_b_debug;
2904 extern int kmp_c_debug;
2905 extern int kmp_d_debug;
2906 extern int kmp_e_debug;
2907 extern int kmp_f_debug;
2908 #endif /* KMP_DEBUG */
2909 
2910 /* For debug information logging using rotating buffer */
2911 #define KMP_DEBUG_BUF_LINES_INIT 512
2912 #define KMP_DEBUG_BUF_LINES_MIN 1
2913 
2914 #define KMP_DEBUG_BUF_CHARS_INIT 128
2915 #define KMP_DEBUG_BUF_CHARS_MIN 2
2916 
2917 extern int
2918     __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2919 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2920 extern int
2921     __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2922 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2923                                       entry pointer */
2924 
2925 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2926 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2927                                               printed in buffer so far */
2928 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2929                                           recommended in warnings */
2930 /* end rotating debug buffer */
2931 
2932 #ifdef KMP_DEBUG
2933 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2934 
2935 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2936 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2937 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2938 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2939 extern int __kmp_par_range_lb;
2940 extern int __kmp_par_range_ub;
2941 #endif
2942 
2943 /* For printing out dynamic storage map for threads and teams */
2944 extern int
2945     __kmp_storage_map; /* True means print storage map for threads and teams */
2946 extern int __kmp_storage_map_verbose; /* True means storage map includes
2947                                          placement info */
2948 extern int __kmp_storage_map_verbose_specified;
2949 
2950 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2951 extern kmp_cpuinfo_t __kmp_cpuinfo;
2952 #endif
2953 
2954 extern volatile int __kmp_init_serial;
2955 extern volatile int __kmp_init_gtid;
2956 extern volatile int __kmp_init_common;
2957 extern volatile int __kmp_init_middle;
2958 extern volatile int __kmp_init_parallel;
2959 #if KMP_USE_MONITOR
2960 extern volatile int __kmp_init_monitor;
2961 #endif
2962 extern volatile int __kmp_init_user_locks;
2963 extern volatile int __kmp_init_hidden_helper_threads;
2964 extern int __kmp_init_counter;
2965 extern int __kmp_root_counter;
2966 extern int __kmp_version;
2967 
2968 /* list of address of allocated caches for commons */
2969 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2970 
2971 /* Barrier algorithm types and options */
2972 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2973 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2974 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2975 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2976 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2977 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2978 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2979 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2980 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2981 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2982 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2983 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2984 
2985 /* Global Locks */
2986 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2987 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2988 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2989 extern kmp_bootstrap_lock_t
2990     __kmp_exit_lock; /* exit() is not always thread-safe */
2991 #if KMP_USE_MONITOR
2992 extern kmp_bootstrap_lock_t
2993     __kmp_monitor_lock; /* control monitor thread creation */
2994 #endif
2995 extern kmp_bootstrap_lock_t
2996     __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2997                              __kmp_threads expansion to co-exist */
2998 
2999 extern kmp_lock_t __kmp_global_lock; /* control OS/global access  */
3000 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access  */
3001 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3002 
3003 extern enum library_type __kmp_library;
3004 
3005 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3006 extern enum sched_type __kmp_static; /* default static scheduling method */
3007 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3008 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3009 extern int __kmp_chunk; /* default runtime chunk size */
3010 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3011 
3012 extern size_t __kmp_stksize; /* stack size per thread         */
3013 #if KMP_USE_MONITOR
3014 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3015 #endif
3016 extern size_t __kmp_stkoffset; /* stack offset per thread       */
3017 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3018 
3019 extern size_t
3020     __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3021 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3022 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3023 extern int __kmp_env_checks; /* was KMP_CHECKS specified?    */
3024 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3025 extern int __kmp_generate_warnings; /* should we issue warnings? */
3026 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3027 
3028 #ifdef DEBUG_SUSPEND
3029 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3030 #endif
3031 
3032 extern kmp_int32 __kmp_use_yield;
3033 extern kmp_int32 __kmp_use_yield_exp_set;
3034 extern kmp_uint32 __kmp_yield_init;
3035 extern kmp_uint32 __kmp_yield_next;
3036 
3037 /* ------------------------------------------------------------------------- */
3038 extern int __kmp_allThreadsSpecified;
3039 
3040 extern size_t __kmp_align_alloc;
3041 /* following data protected by initialization routines */
3042 extern int __kmp_xproc; /* number of processors in the system */
3043 extern int __kmp_avail_proc; /* number of processors available to the process */
3044 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3045 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3046 // maximum total number of concurrently-existing threads on device
3047 extern int __kmp_max_nth;
3048 // maximum total number of concurrently-existing threads in a contention group
3049 extern int __kmp_cg_max_nth;
3050 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3051 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3052                                       __kmp_root */
3053 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3054                                    region a la OMP_NUM_THREADS */
3055 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3056                                       initialization */
3057 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3058                                  used (fixed) */
3059 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3060                                (__kmpc_threadprivate_cached()) */
3061 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3062                                     blocking (env setting) */
3063 #if KMP_USE_MONITOR
3064 extern int
3065     __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3066 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3067                                   blocking */
3068 #endif
3069 #ifdef KMP_ADJUST_BLOCKTIME
3070 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3071 #endif /* KMP_ADJUST_BLOCKTIME */
3072 #ifdef KMP_DFLT_NTH_CORES
3073 extern int __kmp_ncores; /* Total number of cores for threads placement */
3074 #endif
3075 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3076 extern int __kmp_abort_delay;
3077 
3078 extern int __kmp_need_register_atfork_specified;
3079 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3080                                           to install fork handler */
3081 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3082                                0 - not set, will be set at runtime
3083                                1 - using stack search
3084                                2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3085                                    X*) or TlsGetValue(Windows* OS))
3086                                3 - static TLS (__declspec(thread) __kmp_gtid),
3087                                    Linux* OS .so only.  */
3088 extern int
3089     __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3090 #ifdef KMP_TDATA_GTID
3091 extern KMP_THREAD_LOCAL int __kmp_gtid;
3092 #endif
3093 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3094 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3095 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3096 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3097 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3098 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3099 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3100 
3101 // max_active_levels for nested parallelism enabled by default via
3102 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3103 extern int __kmp_dflt_max_active_levels;
3104 // Indicates whether value of __kmp_dflt_max_active_levels was already
3105 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3106 extern bool __kmp_dflt_max_active_levels_set;
3107 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3108                                           concurrent execution per team */
3109 #if KMP_NESTED_HOT_TEAMS
3110 extern int __kmp_hot_teams_mode;
3111 extern int __kmp_hot_teams_max_level;
3112 #endif
3113 
3114 #if KMP_OS_LINUX
3115 extern enum clock_function_type __kmp_clock_function;
3116 extern int __kmp_clock_function_param;
3117 #endif /* KMP_OS_LINUX */
3118 
3119 #if KMP_MIC_SUPPORTED
3120 extern enum mic_type __kmp_mic_type;
3121 #endif
3122 
3123 #ifdef USE_LOAD_BALANCE
3124 extern double __kmp_load_balance_interval; // load balance algorithm interval
3125 #endif /* USE_LOAD_BALANCE */
3126 
3127 // OpenMP 3.1 - Nested num threads array
3128 typedef struct kmp_nested_nthreads_t {
3129   int *nth;
3130   int size;
3131   int used;
3132 } kmp_nested_nthreads_t;
3133 
3134 extern kmp_nested_nthreads_t __kmp_nested_nth;
3135 
3136 #if KMP_USE_ADAPTIVE_LOCKS
3137 
3138 // Parameters for the speculative lock backoff system.
3139 struct kmp_adaptive_backoff_params_t {
3140   // Number of soft retries before it counts as a hard retry.
3141   kmp_uint32 max_soft_retries;
3142   // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3143   // the right
3144   kmp_uint32 max_badness;
3145 };
3146 
3147 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3148 
3149 #if KMP_DEBUG_ADAPTIVE_LOCKS
3150 extern const char *__kmp_speculative_statsfile;
3151 #endif
3152 
3153 #endif // KMP_USE_ADAPTIVE_LOCKS
3154 
3155 extern int __kmp_display_env; /* TRUE or FALSE */
3156 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3157 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3158 extern int __kmp_nteams;
3159 extern int __kmp_teams_thread_limit;
3160 
3161 /* ------------------------------------------------------------------------- */
3162 
3163 /* the following are protected by the fork/join lock */
3164 /* write: lock  read: anytime */
3165 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3166 /* read/write: lock */
3167 extern volatile kmp_team_t *__kmp_team_pool;
3168 extern volatile kmp_info_t *__kmp_thread_pool;
3169 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3170 
3171 // total num threads reachable from some root thread including all root threads
3172 extern volatile int __kmp_nth;
3173 /* total number of threads reachable from some root thread including all root
3174    threads, and those in the thread pool */
3175 extern volatile int __kmp_all_nth;
3176 extern std::atomic<int> __kmp_thread_pool_active_nth;
3177 
3178 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3179 /* end data protected by fork/join lock */
3180 /* ------------------------------------------------------------------------- */
3181 
3182 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3183 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3184 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3185 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3186 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3187 
3188 // AT: Which way is correct?
3189 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3190 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3191 #define __kmp_get_team_num_threads(gtid)                                       \
3192   (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3193 
3194 static inline bool KMP_UBER_GTID(int gtid) {
3195   KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3196   KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3197   return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3198           __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3199 }
3200 
3201 static inline int __kmp_tid_from_gtid(int gtid) {
3202   KMP_DEBUG_ASSERT(gtid >= 0);
3203   return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3204 }
3205 
3206 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3207   KMP_DEBUG_ASSERT(tid >= 0 && team);
3208   return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3209 }
3210 
3211 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3212   KMP_DEBUG_ASSERT(thr);
3213   return thr->th.th_info.ds.ds_gtid;
3214 }
3215 
3216 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3217   KMP_DEBUG_ASSERT(gtid >= 0);
3218   return __kmp_threads[gtid];
3219 }
3220 
3221 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3222   KMP_DEBUG_ASSERT(gtid >= 0);
3223   return __kmp_threads[gtid]->th.th_team;
3224 }
3225 
3226 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3227   if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3228     KMP_FATAL(ThreadIdentInvalid);
3229 }
3230 
3231 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3232 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3233 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3234 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3235 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3236 #endif
3237 
3238 /* ------------------------------------------------------------------------- */
3239 
3240 extern kmp_global_t __kmp_global; /* global status */
3241 
3242 extern kmp_info_t __kmp_monitor;
3243 // For Debugging Support Library
3244 extern std::atomic<kmp_int32> __kmp_team_counter;
3245 // For Debugging Support Library
3246 extern std::atomic<kmp_int32> __kmp_task_counter;
3247 
3248 #if USE_DEBUGGER
3249 #define _KMP_GEN_ID(counter)                                                   \
3250   (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3251 #else
3252 #define _KMP_GEN_ID(counter) (~0)
3253 #endif /* USE_DEBUGGER */
3254 
3255 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3256 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3257 
3258 /* ------------------------------------------------------------------------ */
3259 
3260 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3261                                          size_t size, char const *format, ...);
3262 
3263 extern void __kmp_serial_initialize(void);
3264 extern void __kmp_middle_initialize(void);
3265 extern void __kmp_parallel_initialize(void);
3266 
3267 extern void __kmp_internal_begin(void);
3268 extern void __kmp_internal_end_library(int gtid);
3269 extern void __kmp_internal_end_thread(int gtid);
3270 extern void __kmp_internal_end_atexit(void);
3271 extern void __kmp_internal_end_dtor(void);
3272 extern void __kmp_internal_end_dest(void *);
3273 
3274 extern int __kmp_register_root(int initial_thread);
3275 extern void __kmp_unregister_root(int gtid);
3276 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3277 
3278 extern int __kmp_ignore_mppbeg(void);
3279 extern int __kmp_ignore_mppend(void);
3280 
3281 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3282 extern void __kmp_exit_single(int gtid);
3283 
3284 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3285 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3286 
3287 #ifdef USE_LOAD_BALANCE
3288 extern int __kmp_get_load_balance(int);
3289 #endif
3290 
3291 extern int __kmp_get_global_thread_id(void);
3292 extern int __kmp_get_global_thread_id_reg(void);
3293 extern void __kmp_exit_thread(int exit_status);
3294 extern void __kmp_abort(char const *format, ...);
3295 extern void __kmp_abort_thread(void);
3296 KMP_NORETURN extern void __kmp_abort_process(void);
3297 extern void __kmp_warn(char const *format, ...);
3298 
3299 extern void __kmp_set_num_threads(int new_nth, int gtid);
3300 
3301 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3302 // registered.
3303 static inline kmp_info_t *__kmp_entry_thread() {
3304   int gtid = __kmp_entry_gtid();
3305 
3306   return __kmp_threads[gtid];
3307 }
3308 
3309 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3310 extern int __kmp_get_max_active_levels(int gtid);
3311 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3312 extern int __kmp_get_team_size(int gtid, int level);
3313 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3314 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3315 
3316 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3317 extern void __kmp_init_random(kmp_info_t *thread);
3318 
3319 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3320 extern void __kmp_adjust_num_threads(int new_nproc);
3321 extern void __kmp_check_stksize(size_t *val);
3322 
3323 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3324 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3325 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3326 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3327 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3328 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3329 
3330 #if USE_FAST_MEMORY
3331 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3332                                   size_t size KMP_SRC_LOC_DECL);
3333 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3334 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3335 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3336 #define __kmp_fast_allocate(this_thr, size)                                    \
3337   ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3338 #define __kmp_fast_free(this_thr, ptr)                                         \
3339   ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3340 #endif
3341 
3342 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3343 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3344                                   size_t elsize KMP_SRC_LOC_DECL);
3345 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3346                                    size_t size KMP_SRC_LOC_DECL);
3347 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3348 #define __kmp_thread_malloc(th, size)                                          \
3349   ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3350 #define __kmp_thread_calloc(th, nelem, elsize)                                 \
3351   ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3352 #define __kmp_thread_realloc(th, ptr, size)                                    \
3353   ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3354 #define __kmp_thread_free(th, ptr)                                             \
3355   ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3356 
3357 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3358 #define KMP_INTERNAL_FREE(p) free(p)
3359 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3360 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3361 
3362 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3363 
3364 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3365                                  kmp_proc_bind_t proc_bind);
3366 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3367                                  int num_threads);
3368 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3369                                     int num_teams_ub, int num_threads);
3370 
3371 extern void __kmp_yield();
3372 
3373 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3374                                    enum sched_type schedule, kmp_int32 lb,
3375                                    kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3376 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3377                                     enum sched_type schedule, kmp_uint32 lb,
3378                                     kmp_uint32 ub, kmp_int32 st,
3379                                     kmp_int32 chunk);
3380 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3381                                    enum sched_type schedule, kmp_int64 lb,
3382                                    kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3383 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3384                                     enum sched_type schedule, kmp_uint64 lb,
3385                                     kmp_uint64 ub, kmp_int64 st,
3386                                     kmp_int64 chunk);
3387 
3388 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3389                                   kmp_int32 *p_last, kmp_int32 *p_lb,
3390                                   kmp_int32 *p_ub, kmp_int32 *p_st);
3391 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3392                                    kmp_int32 *p_last, kmp_uint32 *p_lb,
3393                                    kmp_uint32 *p_ub, kmp_int32 *p_st);
3394 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3395                                   kmp_int32 *p_last, kmp_int64 *p_lb,
3396                                   kmp_int64 *p_ub, kmp_int64 *p_st);
3397 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3398                                    kmp_int32 *p_last, kmp_uint64 *p_lb,
3399                                    kmp_uint64 *p_ub, kmp_int64 *p_st);
3400 
3401 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3402 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3403 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3404 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3405 
3406 #ifdef KMP_GOMP_COMPAT
3407 
3408 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3409                                       enum sched_type schedule, kmp_int32 lb,
3410                                       kmp_int32 ub, kmp_int32 st,
3411                                       kmp_int32 chunk, int push_ws);
3412 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3413                                        enum sched_type schedule, kmp_uint32 lb,
3414                                        kmp_uint32 ub, kmp_int32 st,
3415                                        kmp_int32 chunk, int push_ws);
3416 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3417                                       enum sched_type schedule, kmp_int64 lb,
3418                                       kmp_int64 ub, kmp_int64 st,
3419                                       kmp_int64 chunk, int push_ws);
3420 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3421                                        enum sched_type schedule, kmp_uint64 lb,
3422                                        kmp_uint64 ub, kmp_int64 st,
3423                                        kmp_int64 chunk, int push_ws);
3424 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3425 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3426 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3427 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3428 
3429 #endif /* KMP_GOMP_COMPAT */
3430 
3431 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3432 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3433 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3434 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3435 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3436 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3437                                kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3438                                void *obj);
3439 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3440                              kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3441 
3442 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3443                           int final_spin
3444 #if USE_ITT_BUILD
3445                           ,
3446                           void *itt_sync_obj
3447 #endif
3448 );
3449 extern void __kmp_release_64(kmp_flag_64<> *flag);
3450 
3451 extern void __kmp_infinite_loop(void);
3452 
3453 extern void __kmp_cleanup(void);
3454 
3455 #if KMP_HANDLE_SIGNALS
3456 extern int __kmp_handle_signals;
3457 extern void __kmp_install_signals(int parallel_init);
3458 extern void __kmp_remove_signals(void);
3459 #endif
3460 
3461 extern void __kmp_clear_system_time(void);
3462 extern void __kmp_read_system_time(double *delta);
3463 
3464 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3465 
3466 extern void __kmp_expand_host_name(char *buffer, size_t size);
3467 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3468 
3469 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64)
3470 extern void
3471 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3472 #endif
3473 
3474 extern void
3475 __kmp_runtime_initialize(void); /* machine specific initialization */
3476 extern void __kmp_runtime_destroy(void);
3477 
3478 #if KMP_AFFINITY_SUPPORTED
3479 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3480                                        kmp_affin_mask_t *mask);
3481 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3482                                                   kmp_affin_mask_t *mask);
3483 extern void __kmp_affinity_initialize(void);
3484 extern void __kmp_affinity_uninitialize(void);
3485 extern void __kmp_affinity_set_init_mask(
3486     int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3487 extern void __kmp_affinity_set_place(int gtid);
3488 extern void __kmp_affinity_determine_capable(const char *env_var);
3489 extern int __kmp_aux_set_affinity(void **mask);
3490 extern int __kmp_aux_get_affinity(void **mask);
3491 extern int __kmp_aux_get_affinity_max_proc();
3492 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3493 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3494 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3495 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3496 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3497 extern int kmp_set_thread_affinity_mask_initial(void);
3498 #endif
3499 static inline void __kmp_assign_root_init_mask() {
3500   int gtid = __kmp_entry_gtid();
3501   kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3502   if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3503     __kmp_affinity_set_init_mask(gtid, TRUE);
3504     r->r.r_affinity_assigned = TRUE;
3505   }
3506 }
3507 #else /* KMP_AFFINITY_SUPPORTED */
3508 #define __kmp_assign_root_init_mask() /* Nothing */
3509 #endif /* KMP_AFFINITY_SUPPORTED */
3510 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3511 // format string is for affinity, so platforms that do not support
3512 // affinity can still use the other fields, e.g., %n for num_threads
3513 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3514                                          kmp_str_buf_t *buffer);
3515 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3516 
3517 extern void __kmp_cleanup_hierarchy();
3518 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3519 
3520 #if KMP_USE_FUTEX
3521 
3522 extern int __kmp_futex_determine_capable(void);
3523 
3524 #endif // KMP_USE_FUTEX
3525 
3526 extern void __kmp_gtid_set_specific(int gtid);
3527 extern int __kmp_gtid_get_specific(void);
3528 
3529 extern double __kmp_read_cpu_time(void);
3530 
3531 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3532 
3533 #if KMP_USE_MONITOR
3534 extern void __kmp_create_monitor(kmp_info_t *th);
3535 #endif
3536 
3537 extern void *__kmp_launch_thread(kmp_info_t *thr);
3538 
3539 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3540 
3541 #if KMP_OS_WINDOWS
3542 extern int __kmp_still_running(kmp_info_t *th);
3543 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3544 extern void __kmp_free_handle(kmp_thread_t tHandle);
3545 #endif
3546 
3547 #if KMP_USE_MONITOR
3548 extern void __kmp_reap_monitor(kmp_info_t *th);
3549 #endif
3550 extern void __kmp_reap_worker(kmp_info_t *th);
3551 extern void __kmp_terminate_thread(int gtid);
3552 
3553 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3554 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3555 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3556 
3557 extern void __kmp_elapsed(double *);
3558 extern void __kmp_elapsed_tick(double *);
3559 
3560 extern void __kmp_enable(int old_state);
3561 extern void __kmp_disable(int *old_state);
3562 
3563 extern void __kmp_thread_sleep(int millis);
3564 
3565 extern void __kmp_common_initialize(void);
3566 extern void __kmp_common_destroy(void);
3567 extern void __kmp_common_destroy_gtid(int gtid);
3568 
3569 #if KMP_OS_UNIX
3570 extern void __kmp_register_atfork(void);
3571 #endif
3572 extern void __kmp_suspend_initialize(void);
3573 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3574 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3575 
3576 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3577                                          int tid);
3578 extern kmp_team_t *
3579 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3580 #if OMPT_SUPPORT
3581                     ompt_data_t ompt_parallel_data,
3582 #endif
3583                     kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3584                     int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3585 extern void __kmp_free_thread(kmp_info_t *);
3586 extern void __kmp_free_team(kmp_root_t *,
3587                             kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3588 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3589 
3590 /* ------------------------------------------------------------------------ */
3591 
3592 extern void __kmp_initialize_bget(kmp_info_t *th);
3593 extern void __kmp_finalize_bget(kmp_info_t *th);
3594 
3595 KMP_EXPORT void *kmpc_malloc(size_t size);
3596 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3597 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3598 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3599 KMP_EXPORT void kmpc_free(void *ptr);
3600 
3601 /* declarations for internal use */
3602 
3603 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3604                          size_t reduce_size, void *reduce_data,
3605                          void (*reduce)(void *, void *));
3606 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3607 extern int __kmp_barrier_gomp_cancel(int gtid);
3608 
3609 /*!
3610  * Tell the fork call which compiler generated the fork call, and therefore how
3611  * to deal with the call.
3612  */
3613 enum fork_context_e {
3614   fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3615                        microtask internally. */
3616   fork_context_intel, /**< Called from Intel generated code.  */
3617   fork_context_last
3618 };
3619 extern int __kmp_fork_call(ident_t *loc, int gtid,
3620                            enum fork_context_e fork_context, kmp_int32 argc,
3621                            microtask_t microtask, launch_t invoker,
3622                            kmp_va_list ap);
3623 
3624 extern void __kmp_join_call(ident_t *loc, int gtid
3625 #if OMPT_SUPPORT
3626                             ,
3627                             enum fork_context_e fork_context
3628 #endif
3629                             ,
3630                             int exit_teams = 0);
3631 
3632 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3633 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3634 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3635 extern int __kmp_invoke_task_func(int gtid);
3636 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3637                                           kmp_info_t *this_thr,
3638                                           kmp_team_t *team);
3639 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3640                                          kmp_info_t *this_thr,
3641                                          kmp_team_t *team);
3642 
3643 // should never have been exported
3644 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3645 extern int __kmp_invoke_teams_master(int gtid);
3646 extern void __kmp_teams_master(int gtid);
3647 extern int __kmp_aux_get_team_num();
3648 extern int __kmp_aux_get_num_teams();
3649 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3650 extern void __kmp_user_set_library(enum library_type arg);
3651 extern void __kmp_aux_set_library(enum library_type arg);
3652 extern void __kmp_aux_set_stacksize(size_t arg);
3653 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3654 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3655 
3656 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3657 void kmpc_set_blocktime(int arg);
3658 void ompc_set_nested(int flag);
3659 void ompc_set_dynamic(int flag);
3660 void ompc_set_num_threads(int arg);
3661 
3662 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3663                                               kmp_team_t *team, int tid);
3664 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3665 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3666                                     kmp_tasking_flags_t *flags,
3667                                     size_t sizeof_kmp_task_t,
3668                                     size_t sizeof_shareds,
3669                                     kmp_routine_entry_t task_entry);
3670 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3671                                      kmp_team_t *team, int tid,
3672                                      int set_curr_task);
3673 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3674 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3675 
3676 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3677                                                        int gtid,
3678                                                        kmp_task_t *task);
3679 extern void __kmp_fulfill_event(kmp_event_t *event);
3680 
3681 extern void __kmp_free_task_team(kmp_info_t *thread,
3682                                  kmp_task_team_t *task_team);
3683 extern void __kmp_reap_task_teams(void);
3684 extern void __kmp_wait_to_unref_task_teams(void);
3685 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3686                                   int always);
3687 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3688 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3689 #if USE_ITT_BUILD
3690                                  ,
3691                                  void *itt_sync_obj
3692 #endif /* USE_ITT_BUILD */
3693                                  ,
3694                                  int wait = 1);
3695 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3696                                   int gtid);
3697 
3698 extern int __kmp_is_address_mapped(void *addr);
3699 extern kmp_uint64 __kmp_hardware_timestamp(void);
3700 
3701 #if KMP_OS_UNIX
3702 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3703 #endif
3704 
3705 /* ------------------------------------------------------------------------ */
3706 //
3707 // Assembly routines that have no compiler intrinsic replacement
3708 //
3709 
3710 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3711                                   void *argv[]
3712 #if OMPT_SUPPORT
3713                                   ,
3714                                   void **exit_frame_ptr
3715 #endif
3716 );
3717 
3718 /* ------------------------------------------------------------------------ */
3719 
3720 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3721 KMP_EXPORT void __kmpc_end(ident_t *);
3722 
3723 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3724                                                   kmpc_ctor_vec ctor,
3725                                                   kmpc_cctor_vec cctor,
3726                                                   kmpc_dtor_vec dtor,
3727                                                   size_t vector_length);
3728 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3729                                               kmpc_ctor ctor, kmpc_cctor cctor,
3730                                               kmpc_dtor dtor);
3731 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3732                                       void *data, size_t size);
3733 
3734 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3735 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3736 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3737 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3738 
3739 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3740 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3741                                  kmpc_micro microtask, ...);
3742 
3743 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3744 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3745 
3746 KMP_EXPORT void __kmpc_flush(ident_t *);
3747 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3748 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3749 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3750 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3751                                    kmp_int32 filter);
3752 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3753 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3754 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3755 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3756                                 kmp_critical_name *);
3757 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3758                                     kmp_critical_name *);
3759 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3760                                           kmp_critical_name *, uint32_t hint);
3761 
3762 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3763 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3764 
3765 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3766                                                   kmp_int32 global_tid);
3767 
3768 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3769 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3770 
3771 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3772                                      kmp_int32 schedtype, kmp_int32 *plastiter,
3773                                      kmp_int *plower, kmp_int *pupper,
3774                                      kmp_int *pstride, kmp_int incr,
3775                                      kmp_int chunk);
3776 
3777 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3778 
3779 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3780                                    size_t cpy_size, void *cpy_data,
3781                                    void (*cpy_func)(void *, void *),
3782                                    kmp_int32 didit);
3783 
3784 extern void KMPC_SET_NUM_THREADS(int arg);
3785 extern void KMPC_SET_DYNAMIC(int flag);
3786 extern void KMPC_SET_NESTED(int flag);
3787 
3788 /* OMP 3.0 tasking interface routines */
3789 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3790                                      kmp_task_t *new_task);
3791 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3792                                              kmp_int32 flags,
3793                                              size_t sizeof_kmp_task_t,
3794                                              size_t sizeof_shareds,
3795                                              kmp_routine_entry_t task_entry);
3796 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
3797     ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
3798     size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
3799 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3800                                           kmp_task_t *task);
3801 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3802                                              kmp_task_t *task);
3803 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3804                                            kmp_task_t *new_task);
3805 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3806 
3807 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3808                                           int end_part);
3809 
3810 #if TASK_UNUSED
3811 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3812 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3813                               kmp_task_t *task);
3814 #endif // TASK_UNUSED
3815 
3816 /* ------------------------------------------------------------------------ */
3817 
3818 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3819 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3820 
3821 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3822     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3823     kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3824     kmp_depend_info_t *noalias_dep_list);
3825 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3826                                      kmp_int32 ndeps,
3827                                      kmp_depend_info_t *dep_list,
3828                                      kmp_int32 ndeps_noalias,
3829                                      kmp_depend_info_t *noalias_dep_list);
3830 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3831                                 bool serialize_immediate);
3832 
3833 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3834                                    kmp_int32 cncl_kind);
3835 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3836                                               kmp_int32 cncl_kind);
3837 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3838 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3839 
3840 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3841 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3842 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3843                                 kmp_int32 if_val, kmp_uint64 *lb,
3844                                 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3845                                 kmp_int32 sched, kmp_uint64 grainsize,
3846                                 void *task_dup);
3847 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3848                                   kmp_task_t *task, kmp_int32 if_val,
3849                                   kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3850                                   kmp_int32 nogroup, kmp_int32 sched,
3851                                   kmp_uint64 grainsize, kmp_int32 modifier,
3852                                   void *task_dup);
3853 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3854 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3855 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3856 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3857                                                      int is_ws, int num,
3858                                                      void *data);
3859 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3860                                               int num, void *data);
3861 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3862                                                     int is_ws);
3863 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3864     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3865     kmp_task_affinity_info_t *affin_list);
3866 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
3867 KMP_EXPORT int __kmp_get_max_teams(void);
3868 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
3869 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
3870 
3871 /* Lock interface routines (fast versions with gtid passed in) */
3872 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3873                                  void **user_lock);
3874 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3875                                       void **user_lock);
3876 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3877                                     void **user_lock);
3878 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3879                                          void **user_lock);
3880 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3881 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3882                                      void **user_lock);
3883 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3884                                   void **user_lock);
3885 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3886                                        void **user_lock);
3887 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3888 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3889                                      void **user_lock);
3890 
3891 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3892                                            void **user_lock, uintptr_t hint);
3893 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3894                                                 void **user_lock,
3895                                                 uintptr_t hint);
3896 
3897 /* Interface to fast scalable reduce methods routines */
3898 
3899 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3900     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3901     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3902     kmp_critical_name *lck);
3903 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3904                                          kmp_critical_name *lck);
3905 KMP_EXPORT kmp_int32 __kmpc_reduce(
3906     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3907     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3908     kmp_critical_name *lck);
3909 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3910                                   kmp_critical_name *lck);
3911 
3912 /* Internal fast reduction routines */
3913 
3914 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3915     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3916     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3917     kmp_critical_name *lck);
3918 
3919 // this function is for testing set/get/determine reduce method
3920 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3921 
3922 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3923 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3924 
3925 // C++ port
3926 // missing 'extern "C"' declarations
3927 
3928 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3929 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3930 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3931                                         kmp_int32 num_threads);
3932 
3933 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3934                                       int proc_bind);
3935 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3936                                       kmp_int32 num_teams,
3937                                       kmp_int32 num_threads);
3938 /* Function for OpenMP 5.1 num_teams clause */
3939 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
3940                                          kmp_int32 num_teams_lb,
3941                                          kmp_int32 num_teams_ub,
3942                                          kmp_int32 num_threads);
3943 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3944                                   kmpc_micro microtask, ...);
3945 struct kmp_dim { // loop bounds info casted to kmp_int64
3946   kmp_int64 lo; // lower
3947   kmp_int64 up; // upper
3948   kmp_int64 st; // stride
3949 };
3950 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3951                                      kmp_int32 num_dims,
3952                                      const struct kmp_dim *dims);
3953 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3954                                      const kmp_int64 *vec);
3955 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3956                                      const kmp_int64 *vec);
3957 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3958 
3959 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3960                                              void *data, size_t size,
3961                                              void ***cache);
3962 
3963 // Symbols for MS mutual detection.
3964 extern int _You_must_link_with_exactly_one_OpenMP_library;
3965 extern int _You_must_link_with_Intel_OpenMP_library;
3966 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3967 extern int _You_must_link_with_Microsoft_OpenMP_library;
3968 #endif
3969 
3970 // The routines below are not exported.
3971 // Consider making them 'static' in corresponding source files.
3972 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3973                                            void *data_addr, size_t pc_size);
3974 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3975                                                 void *data_addr,
3976                                                 size_t pc_size);
3977 void __kmp_threadprivate_resize_cache(int newCapacity);
3978 void __kmp_cleanup_threadprivate_caches();
3979 
3980 // ompc_, kmpc_ entries moved from omp.h.
3981 #if KMP_OS_WINDOWS
3982 #define KMPC_CONVENTION __cdecl
3983 #else
3984 #define KMPC_CONVENTION
3985 #endif
3986 
3987 #ifndef __OMP_H
3988 typedef enum omp_sched_t {
3989   omp_sched_static = 1,
3990   omp_sched_dynamic = 2,
3991   omp_sched_guided = 3,
3992   omp_sched_auto = 4
3993 } omp_sched_t;
3994 typedef void *kmp_affinity_mask_t;
3995 #endif
3996 
3997 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3998 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3999 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4000 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4001 KMP_EXPORT int KMPC_CONVENTION
4002 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4003 KMP_EXPORT int KMPC_CONVENTION
4004 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4005 KMP_EXPORT int KMPC_CONVENTION
4006 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4007 
4008 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4009 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4010 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4011 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4012 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4013 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4014 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4015 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4016 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4017                                               char const *format);
4018 
4019 enum kmp_target_offload_kind {
4020   tgt_disabled = 0,
4021   tgt_default = 1,
4022   tgt_mandatory = 2
4023 };
4024 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4025 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4026 extern kmp_target_offload_kind_t __kmp_target_offload;
4027 extern int __kmpc_get_target_offload();
4028 
4029 // Constants used in libomptarget
4030 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4031 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4032 
4033 // OMP Pause Resource
4034 
4035 // The following enum is used both to set the status in __kmp_pause_status, and
4036 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4037 typedef enum kmp_pause_status_t {
4038   kmp_not_paused = 0, // status is not paused, or, requesting resume
4039   kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4040   kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4041 } kmp_pause_status_t;
4042 
4043 // This stores the pause state of the runtime
4044 extern kmp_pause_status_t __kmp_pause_status;
4045 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4046 extern int __kmp_pause_resource(kmp_pause_status_t level);
4047 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4048 extern void __kmp_resume_if_soft_paused();
4049 // Hard resume simply resets the status to not paused. Library will appear to
4050 // be uninitialized after hard pause. Let OMP constructs trigger required
4051 // initializations.
4052 static inline void __kmp_resume_if_hard_paused() {
4053   if (__kmp_pause_status == kmp_hard_paused) {
4054     __kmp_pause_status = kmp_not_paused;
4055   }
4056 }
4057 
4058 extern void __kmp_omp_display_env(int verbose);
4059 
4060 // 1: it is initializing hidden helper team
4061 extern volatile int __kmp_init_hidden_helper;
4062 // 1: the hidden helper team is done
4063 extern volatile int __kmp_hidden_helper_team_done;
4064 // 1: enable hidden helper task
4065 extern kmp_int32 __kmp_enable_hidden_helper;
4066 // Main thread of hidden helper team
4067 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4068 // Descriptors for the hidden helper threads
4069 extern kmp_info_t **__kmp_hidden_helper_threads;
4070 // Number of hidden helper threads
4071 extern kmp_int32 __kmp_hidden_helper_threads_num;
4072 // Number of hidden helper tasks that have not been executed yet
4073 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4074 
4075 extern void __kmp_hidden_helper_initialize();
4076 extern void __kmp_hidden_helper_threads_initz_routine();
4077 extern void __kmp_do_initialize_hidden_helper_threads();
4078 extern void __kmp_hidden_helper_threads_initz_wait();
4079 extern void __kmp_hidden_helper_initz_release();
4080 extern void __kmp_hidden_helper_threads_deinitz_wait();
4081 extern void __kmp_hidden_helper_threads_deinitz_release();
4082 extern void __kmp_hidden_helper_main_thread_wait();
4083 extern void __kmp_hidden_helper_worker_thread_wait();
4084 extern void __kmp_hidden_helper_worker_thread_signal();
4085 extern void __kmp_hidden_helper_main_thread_release();
4086 
4087 // Check whether a given thread is a hidden helper thread
4088 #define KMP_HIDDEN_HELPER_THREAD(gtid)                                         \
4089   ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4090 
4091 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)                                  \
4092   ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4093 
4094 #define KMP_HIDDEN_HELPER_TEAM(team)                                           \
4095   (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4096 
4097 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4098 // main thread, is skipped.
4099 #define KMP_GTID_TO_SHADOW_GTID(gtid)                                          \
4100   ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4101 
4102 // Return the adjusted gtid value by subtracting from gtid the number
4103 // of hidden helper threads. This adjusted value is the gtid the thread would
4104 // have received if there were no hidden helper threads.
4105 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4106   int adjusted_gtid = gtid;
4107   if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4108       gtid - __kmp_hidden_helper_threads_num >= 0) {
4109     adjusted_gtid -= __kmp_hidden_helper_threads_num;
4110   }
4111   return adjusted_gtid;
4112 }
4113 
4114 // Support for error directive
4115 typedef enum kmp_severity_t {
4116   severity_warning = 1,
4117   severity_fatal = 2
4118 } kmp_severity_t;
4119 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4120 
4121 #ifdef __cplusplus
4122 }
4123 #endif
4124 
4125 template <bool C, bool S>
4126 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4127 template <bool C, bool S>
4128 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4129 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4130 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4131 template <bool C, bool S>
4132 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4133 template <bool C, bool S>
4134 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4135 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4136 #endif
4137 template <bool C, bool S>
4138 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4139 template <bool C, bool S>
4140 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4141 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4142 
4143 template <bool C, bool S>
4144 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4145                            kmp_flag_32<C, S> *flag, int final_spin,
4146                            int *thread_finished,
4147 #if USE_ITT_BUILD
4148                            void *itt_sync_obj,
4149 #endif /* USE_ITT_BUILD */
4150                            kmp_int32 is_constrained);
4151 template <bool C, bool S>
4152 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4153                            kmp_flag_64<C, S> *flag, int final_spin,
4154                            int *thread_finished,
4155 #if USE_ITT_BUILD
4156                            void *itt_sync_obj,
4157 #endif /* USE_ITT_BUILD */
4158                            kmp_int32 is_constrained);
4159 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4160                                kmp_flag_oncore *flag, int final_spin,
4161                                int *thread_finished,
4162 #if USE_ITT_BUILD
4163                                void *itt_sync_obj,
4164 #endif /* USE_ITT_BUILD */
4165                                kmp_int32 is_constrained);
4166 
4167 extern int __kmp_nesting_mode;
4168 extern int __kmp_nesting_mode_nlevels;
4169 extern int *__kmp_nesting_nth_level;
4170 extern void __kmp_init_nesting_mode();
4171 extern void __kmp_set_nesting_mode_threads();
4172 
4173 /// This class safely opens and closes a C-style FILE* object using RAII
4174 /// semantics. There are also methods which allow using stdout or stderr as
4175 /// the underlying FILE* object. With the implicit conversion operator to
4176 /// FILE*, an object with this type can be used in any function which takes
4177 /// a FILE* object e.g., fprintf().
4178 /// No close method is needed at use sites.
4179 class kmp_safe_raii_file_t {
4180   FILE *f;
4181 
4182   void close() {
4183     if (f && f != stdout && f != stderr) {
4184       fclose(f);
4185       f = nullptr;
4186     }
4187   }
4188 
4189 public:
4190   kmp_safe_raii_file_t() : f(nullptr) {}
4191   kmp_safe_raii_file_t(const char *filename, const char *mode,
4192                        const char *env_var = nullptr)
4193       : f(nullptr) {
4194     open(filename, mode, env_var);
4195   }
4196   ~kmp_safe_raii_file_t() { close(); }
4197 
4198   /// Open filename using mode. This is automatically closed in the destructor.
4199   /// The env_var parameter indicates the environment variable the filename
4200   /// came from if != nullptr.
4201   void open(const char *filename, const char *mode,
4202             const char *env_var = nullptr) {
4203     KMP_ASSERT(!f);
4204     f = fopen(filename, mode);
4205     if (!f) {
4206       int code = errno;
4207       if (env_var) {
4208         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4209                     KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4210       } else {
4211         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4212                     __kmp_msg_null);
4213       }
4214     }
4215   }
4216   /// Set the FILE* object to stdout and output there
4217   /// No open call should happen before this call.
4218   void set_stdout() {
4219     KMP_ASSERT(!f);
4220     f = stdout;
4221   }
4222   /// Set the FILE* object to stderr and output there
4223   /// No open call should happen before this call.
4224   void set_stderr() {
4225     KMP_ASSERT(!f);
4226     f = stderr;
4227   }
4228   operator bool() { return bool(f); }
4229   operator FILE *() { return f; }
4230 };
4231 
4232 template <typename SourceType, typename TargetType,
4233           bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4234           bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4235           bool isSourceSigned = std::is_signed<SourceType>::value,
4236           bool isTargetSigned = std::is_signed<TargetType>::value>
4237 struct kmp_convert {};
4238 
4239 // Both types are signed; Source smaller
4240 template <typename SourceType, typename TargetType>
4241 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4242   static TargetType to(SourceType src) { return (TargetType)src; }
4243 };
4244 // Source equal
4245 template <typename SourceType, typename TargetType>
4246 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4247   static TargetType to(SourceType src) { return src; }
4248 };
4249 // Source bigger
4250 template <typename SourceType, typename TargetType>
4251 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4252   static TargetType to(SourceType src) {
4253     KMP_ASSERT(src <= static_cast<SourceType>(
4254                           (std::numeric_limits<TargetType>::max)()));
4255     KMP_ASSERT(src >= static_cast<SourceType>(
4256                           (std::numeric_limits<TargetType>::min)()));
4257     return (TargetType)src;
4258   }
4259 };
4260 
4261 // Source signed, Target unsigned
4262 // Source smaller
4263 template <typename SourceType, typename TargetType>
4264 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4265   static TargetType to(SourceType src) {
4266     KMP_ASSERT(src >= 0);
4267     return (TargetType)src;
4268   }
4269 };
4270 // Source equal
4271 template <typename SourceType, typename TargetType>
4272 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4273   static TargetType to(SourceType src) {
4274     KMP_ASSERT(src >= 0);
4275     return (TargetType)src;
4276   }
4277 };
4278 // Source bigger
4279 template <typename SourceType, typename TargetType>
4280 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4281   static TargetType to(SourceType src) {
4282     KMP_ASSERT(src >= 0);
4283     KMP_ASSERT(src <= static_cast<SourceType>(
4284                           (std::numeric_limits<TargetType>::max)()));
4285     return (TargetType)src;
4286   }
4287 };
4288 
4289 // Source unsigned, Target signed
4290 // Source smaller
4291 template <typename SourceType, typename TargetType>
4292 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4293   static TargetType to(SourceType src) { return (TargetType)src; }
4294 };
4295 // Source equal
4296 template <typename SourceType, typename TargetType>
4297 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4298   static TargetType to(SourceType src) {
4299     KMP_ASSERT(src <= static_cast<SourceType>(
4300                           (std::numeric_limits<TargetType>::max)()));
4301     return (TargetType)src;
4302   }
4303 };
4304 // Source bigger
4305 template <typename SourceType, typename TargetType>
4306 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4307   static TargetType to(SourceType src) {
4308     KMP_ASSERT(src <= static_cast<SourceType>(
4309                           (std::numeric_limits<TargetType>::max)()));
4310     return (TargetType)src;
4311   }
4312 };
4313 
4314 // Source unsigned, Target unsigned
4315 // Source smaller
4316 template <typename SourceType, typename TargetType>
4317 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4318   static TargetType to(SourceType src) { return (TargetType)src; }
4319 };
4320 // Source equal
4321 template <typename SourceType, typename TargetType>
4322 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4323   static TargetType to(SourceType src) { return src; }
4324 };
4325 // Source bigger
4326 template <typename SourceType, typename TargetType>
4327 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4328   static TargetType to(SourceType src) {
4329     KMP_ASSERT(src <= static_cast<SourceType>(
4330                           (std::numeric_limits<TargetType>::max)()));
4331     return (TargetType)src;
4332   }
4333 };
4334 
4335 template <typename T1, typename T2>
4336 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4337   *dest = kmp_convert<T1, T2>::to(src);
4338 }
4339 
4340 #endif /* KMP_H */
4341