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