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_master,
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 
962 typedef struct {
963   omp_alloctrait_key_t key;
964   omp_uintptr_t value;
965 } omp_alloctrait_t;
966 
967 typedef void *omp_allocator_handle_t;
968 extern omp_allocator_handle_t const omp_null_allocator;
969 extern omp_allocator_handle_t const omp_default_mem_alloc;
970 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
971 extern omp_allocator_handle_t const omp_const_mem_alloc;
972 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
973 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
974 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
975 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
976 extern omp_allocator_handle_t const omp_thread_mem_alloc;
977 extern omp_allocator_handle_t const kmp_max_mem_alloc;
978 extern omp_allocator_handle_t __kmp_def_allocator;
979 
980 // end of duplicate type definitions from omp.h
981 #endif
982 
983 extern int __kmp_memkind_available;
984 
985 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
986 
987 typedef struct kmp_allocator_t {
988   omp_memspace_handle_t memspace;
989   void **memkind; // pointer to memkind
990   int alignment;
991   omp_alloctrait_value_t fb;
992   kmp_allocator_t *fb_data;
993   kmp_uint64 pool_size;
994   kmp_uint64 pool_used;
995 } kmp_allocator_t;
996 
997 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
998                                                     omp_memspace_handle_t,
999                                                     int ntraits,
1000                                                     omp_alloctrait_t traits[]);
1001 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1002 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1003 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1004 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1005 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1006                            omp_allocator_handle_t al);
1007 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1008                             omp_allocator_handle_t al,
1009                             omp_allocator_handle_t free_al);
1010 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1011 
1012 extern void __kmp_init_memkind();
1013 extern void __kmp_fini_memkind();
1014 
1015 /* ------------------------------------------------------------------------ */
1016 
1017 #define KMP_UINT64_MAX                                                         \
1018   (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1019 
1020 #define KMP_MIN_NTH 1
1021 
1022 #ifndef KMP_MAX_NTH
1023 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1024 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1025 #else
1026 #define KMP_MAX_NTH INT_MAX
1027 #endif
1028 #endif /* KMP_MAX_NTH */
1029 
1030 #ifdef PTHREAD_STACK_MIN
1031 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1032 #else
1033 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1034 #endif
1035 
1036 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1037 
1038 #if KMP_ARCH_X86
1039 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1040 #elif KMP_ARCH_X86_64
1041 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1042 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1043 #else
1044 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1045 #endif
1046 
1047 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1048 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1049 #define KMP_MAX_MALLOC_POOL_INCR                                               \
1050   (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1051 
1052 #define KMP_MIN_STKOFFSET (0)
1053 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1054 #if KMP_OS_DARWIN
1055 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1056 #else
1057 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1058 #endif
1059 
1060 #define KMP_MIN_STKPADDING (0)
1061 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1062 
1063 #define KMP_BLOCKTIME_MULTIPLIER                                               \
1064   (1000) /* number of blocktime units per second */
1065 #define KMP_MIN_BLOCKTIME (0)
1066 #define KMP_MAX_BLOCKTIME                                                      \
1067   (INT_MAX) /* Must be this for "infinite" setting the work */
1068 #define KMP_DEFAULT_BLOCKTIME (200) /*  __kmp_blocktime is in milliseconds  */
1069 
1070 #if KMP_USE_MONITOR
1071 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1072 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1073 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1074 
1075 /* Calculate new number of monitor wakeups for a specific block time based on
1076    previous monitor_wakeups. Only allow increasing number of wakeups */
1077 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups)                 \
1078   (((blocktime) == KMP_MAX_BLOCKTIME)                                          \
1079        ? (monitor_wakeups)                                                     \
1080        : ((blocktime) == KMP_MIN_BLOCKTIME)                                    \
1081              ? KMP_MAX_MONITOR_WAKEUPS                                         \
1082              : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime)))  \
1083                    ? (monitor_wakeups)                                         \
1084                    : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1085 
1086 /* Calculate number of intervals for a specific block time based on
1087    monitor_wakeups */
1088 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups)               \
1089   (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) /        \
1090    (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1091 #else
1092 #define KMP_BLOCKTIME(team, tid)                                               \
1093   (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1094 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1095 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1096 extern kmp_uint64 __kmp_ticks_per_msec;
1097 #if KMP_COMPILER_ICC
1098 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1099 #else
1100 #define KMP_NOW() __kmp_hardware_timestamp()
1101 #endif
1102 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1103 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1104   (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1105 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1106 #else
1107 // System time is retrieved sporadically while blocking.
1108 extern kmp_uint64 __kmp_now_nsec();
1109 #define KMP_NOW() __kmp_now_nsec()
1110 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1111 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1112   (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1113 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1114 #endif
1115 #endif // KMP_USE_MONITOR
1116 
1117 #define KMP_MIN_STATSCOLS 40
1118 #define KMP_MAX_STATSCOLS 4096
1119 #define KMP_DEFAULT_STATSCOLS 80
1120 
1121 #define KMP_MIN_INTERVAL 0
1122 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1123 #define KMP_DEFAULT_INTERVAL 0
1124 
1125 #define KMP_MIN_CHUNK 1
1126 #define KMP_MAX_CHUNK (INT_MAX - 1)
1127 #define KMP_DEFAULT_CHUNK 1
1128 
1129 #define KMP_DFLT_DISP_NUM_BUFF 7
1130 #define KMP_MAX_ORDERED 8
1131 
1132 #define KMP_MAX_FIELDS 32
1133 
1134 #define KMP_MAX_BRANCH_BITS 31
1135 
1136 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1137 
1138 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1139 
1140 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1141 
1142 /* Minimum number of threads before switch to TLS gtid (experimentally
1143    determined) */
1144 /* josh TODO: what about OS X* tuning? */
1145 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1146 #define KMP_TLS_GTID_MIN 5
1147 #else
1148 #define KMP_TLS_GTID_MIN INT_MAX
1149 #endif
1150 
1151 #define KMP_MASTER_TID(tid) (0 == (tid))
1152 #define KMP_WORKER_TID(tid) (0 != (tid))
1153 
1154 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1155 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1156 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1157 
1158 #ifndef TRUE
1159 #define FALSE 0
1160 #define TRUE (!FALSE)
1161 #endif
1162 
1163 /* NOTE: all of the following constants must be even */
1164 
1165 #if KMP_OS_WINDOWS
1166 #define KMP_INIT_WAIT 64U /* initial number of spin-tests   */
1167 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1168 #elif KMP_OS_LINUX
1169 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1170 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1171 #elif KMP_OS_DARWIN
1172 /* TODO: tune for KMP_OS_DARWIN */
1173 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1174 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1175 #elif KMP_OS_DRAGONFLY
1176 /* TODO: tune for KMP_OS_DRAGONFLY */
1177 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1178 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1179 #elif KMP_OS_FREEBSD
1180 /* TODO: tune for KMP_OS_FREEBSD */
1181 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1182 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1183 #elif KMP_OS_NETBSD
1184 /* TODO: tune for KMP_OS_NETBSD */
1185 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1186 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1187 #elif KMP_OS_HURD
1188 /* TODO: tune for KMP_OS_HURD */
1189 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1190 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1191 #elif KMP_OS_OPENBSD
1192 /* TODO: tune for KMP_OS_OPENBSD */
1193 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1194 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1195 #endif
1196 
1197 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1198 typedef struct kmp_cpuid {
1199   kmp_uint32 eax;
1200   kmp_uint32 ebx;
1201   kmp_uint32 ecx;
1202   kmp_uint32 edx;
1203 } kmp_cpuid_t;
1204 
1205 typedef struct kmp_cpuinfo {
1206   int initialized; // If 0, other fields are not initialized.
1207   int signature; // CPUID(1).EAX
1208   int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1209   int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1210   // Model << 4 ) + Model)
1211   int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1212   int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1213   int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1214   int cpu_stackoffset;
1215   int apic_id;
1216   int physical_id;
1217   int logical_id;
1218   kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1219   char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1220 } kmp_cpuinfo_t;
1221 
1222 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1223 
1224 #if KMP_OS_UNIX
1225 // subleaf is only needed for cache and topology discovery and can be set to
1226 // zero in most cases
1227 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1228   __asm__ __volatile__("cpuid"
1229                        : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1230                        : "a"(leaf), "c"(subleaf));
1231 }
1232 // Load p into FPU control word
1233 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1234   __asm__ __volatile__("fldcw %0" : : "m"(*p));
1235 }
1236 // Store FPU control word into p
1237 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1238   __asm__ __volatile__("fstcw %0" : "=m"(*p));
1239 }
1240 static inline void __kmp_clear_x87_fpu_status_word() {
1241 #if KMP_MIC
1242   // 32-bit protected mode x87 FPU state
1243   struct x87_fpu_state {
1244     unsigned cw;
1245     unsigned sw;
1246     unsigned tw;
1247     unsigned fip;
1248     unsigned fips;
1249     unsigned fdp;
1250     unsigned fds;
1251   };
1252   struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1253   __asm__ __volatile__("fstenv %0\n\t" // store FP env
1254                        "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1255                        "fldenv %0\n\t" // load FP env back
1256                        : "+m"(fpu_state), "+m"(fpu_state.sw));
1257 #else
1258   __asm__ __volatile__("fnclex");
1259 #endif // KMP_MIC
1260 }
1261 #if __SSE__
1262 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1263 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1264 #else
1265 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1266 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1267 #endif
1268 #else
1269 // Windows still has these as external functions in assembly file
1270 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1271 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1272 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1273 extern void __kmp_clear_x87_fpu_status_word();
1274 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1275 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1276 #endif // KMP_OS_UNIX
1277 
1278 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1279 
1280 #if KMP_ARCH_X86
1281 extern void __kmp_x86_pause(void);
1282 #elif KMP_MIC
1283 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1284 // regression after removal of extra PAUSE from spin loops. Changing
1285 // the delay from 100 to 300 showed even better performance than double PAUSE
1286 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1287 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1288 #else
1289 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1290 #endif
1291 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1292 #elif KMP_ARCH_PPC64
1293 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1294 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1295 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1296 #define KMP_CPU_PAUSE()                                                        \
1297   do {                                                                         \
1298     KMP_PPC64_PRI_LOW();                                                       \
1299     KMP_PPC64_PRI_MED();                                                       \
1300     KMP_PPC64_PRI_LOC_MB();                                                    \
1301   } while (0)
1302 #else
1303 #define KMP_CPU_PAUSE() /* nothing to do */
1304 #endif
1305 
1306 #define KMP_INIT_YIELD(count)                                                  \
1307   { (count) = __kmp_yield_init; }
1308 
1309 #define KMP_OVERSUBSCRIBED                                                     \
1310   (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1311 
1312 #define KMP_TRY_YIELD                                                          \
1313   ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1314 
1315 #define KMP_TRY_YIELD_OVERSUB                                                  \
1316   ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1317 
1318 #define KMP_YIELD(cond)                                                        \
1319   {                                                                            \
1320     KMP_CPU_PAUSE();                                                           \
1321     if ((cond) && (KMP_TRY_YIELD))                                             \
1322       __kmp_yield();                                                           \
1323   }
1324 
1325 #define KMP_YIELD_OVERSUB()                                                    \
1326   {                                                                            \
1327     KMP_CPU_PAUSE();                                                           \
1328     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1329       __kmp_yield();                                                           \
1330   }
1331 
1332 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1333 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1334 #define KMP_YIELD_SPIN(count)                                                  \
1335   {                                                                            \
1336     KMP_CPU_PAUSE();                                                           \
1337     if (KMP_TRY_YIELD) {                                                       \
1338       (count) -= 2;                                                            \
1339       if (!(count)) {                                                          \
1340         __kmp_yield();                                                         \
1341         (count) = __kmp_yield_next;                                            \
1342       }                                                                        \
1343     }                                                                          \
1344   }
1345 
1346 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count)                                     \
1347   {                                                                            \
1348     KMP_CPU_PAUSE();                                                           \
1349     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1350       __kmp_yield();                                                           \
1351     else if (__kmp_use_yield == 1) {                                           \
1352       (count) -= 2;                                                            \
1353       if (!(count)) {                                                          \
1354         __kmp_yield();                                                         \
1355         (count) = __kmp_yield_next;                                            \
1356       }                                                                        \
1357     }                                                                          \
1358   }
1359 
1360 // User-level Monitor/Mwait
1361 #if KMP_HAVE_UMWAIT
1362 // We always try for UMWAIT first
1363 #if KMP_HAVE_WAITPKG_INTRINSICS
1364 #if KMP_HAVE_IMMINTRIN_H
1365 #include <immintrin.h>
1366 #elif KMP_HAVE_INTRIN_H
1367 #include <intrin.h>
1368 #endif
1369 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1370 KMP_ATTRIBUTE_TARGET_WAITPKG
1371 static inline int
1372 __kmp_tpause(uint32_t hint, uint64_t counter) {
1373 #if !KMP_HAVE_WAITPKG_INTRINSICS
1374   uint32_t timeHi = uint32_t(counter >> 32);
1375   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1376   char flag;
1377   __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1378                    "setb   %0"
1379                    : "=r"(flag)
1380                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1381                    :);
1382   return flag;
1383 #else
1384   return _tpause(hint, counter);
1385 #endif
1386 }
1387 KMP_ATTRIBUTE_TARGET_WAITPKG
1388 static inline void
1389 __kmp_umonitor(void *cacheline) {
1390 #if !KMP_HAVE_WAITPKG_INTRINSICS
1391   __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1392                    :
1393                    : "a"(cacheline)
1394                    :);
1395 #else
1396   _umonitor(cacheline);
1397 #endif
1398 }
1399 KMP_ATTRIBUTE_TARGET_WAITPKG
1400 static inline int
1401 __kmp_umwait(uint32_t hint, uint64_t counter) {
1402 #if !KMP_HAVE_WAITPKG_INTRINSICS
1403   uint32_t timeHi = uint32_t(counter >> 32);
1404   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1405   char flag;
1406   __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1407                    "setb   %0"
1408                    : "=r"(flag)
1409                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1410                    :);
1411   return flag;
1412 #else
1413   return _umwait(hint, counter);
1414 #endif
1415 }
1416 #elif KMP_HAVE_MWAIT
1417 #if KMP_OS_UNIX
1418 #include <pmmintrin.h>
1419 #else
1420 #include <intrin.h>
1421 #endif
1422 #if KMP_OS_UNIX
1423 __attribute__((target("sse3")))
1424 #endif
1425 static inline void
1426 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1427   _mm_monitor(cacheline, extensions, hints);
1428 }
1429 #if KMP_OS_UNIX
1430 __attribute__((target("sse3")))
1431 #endif
1432 static inline void
1433 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1434   _mm_mwait(extensions, hints);
1435 }
1436 #endif // KMP_HAVE_UMWAIT
1437 
1438 /* ------------------------------------------------------------------------ */
1439 /* Support datatypes for the orphaned construct nesting checks.             */
1440 /* ------------------------------------------------------------------------ */
1441 
1442 enum cons_type {
1443   ct_none,
1444   ct_parallel,
1445   ct_pdo,
1446   ct_pdo_ordered,
1447   ct_psections,
1448   ct_psingle,
1449   ct_critical,
1450   ct_ordered_in_parallel,
1451   ct_ordered_in_pdo,
1452   ct_master,
1453   ct_reduce,
1454   ct_barrier
1455 };
1456 
1457 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1458 
1459 struct cons_data {
1460   ident_t const *ident;
1461   enum cons_type type;
1462   int prev;
1463   kmp_user_lock_p
1464       name; /* address exclusively for critical section name comparison */
1465 };
1466 
1467 struct cons_header {
1468   int p_top, w_top, s_top;
1469   int stack_size, stack_top;
1470   struct cons_data *stack_data;
1471 };
1472 
1473 struct kmp_region_info {
1474   char *text;
1475   int offset[KMP_MAX_FIELDS];
1476   int length[KMP_MAX_FIELDS];
1477 };
1478 
1479 /* ---------------------------------------------------------------------- */
1480 /* ---------------------------------------------------------------------- */
1481 
1482 #if KMP_OS_WINDOWS
1483 typedef HANDLE kmp_thread_t;
1484 typedef DWORD kmp_key_t;
1485 #endif /* KMP_OS_WINDOWS */
1486 
1487 #if KMP_OS_UNIX
1488 typedef pthread_t kmp_thread_t;
1489 typedef pthread_key_t kmp_key_t;
1490 #endif
1491 
1492 extern kmp_key_t __kmp_gtid_threadprivate_key;
1493 
1494 typedef struct kmp_sys_info {
1495   long maxrss; /* the maximum resident set size utilized (in kilobytes)     */
1496   long minflt; /* the number of page faults serviced without any I/O        */
1497   long majflt; /* the number of page faults serviced that required I/O      */
1498   long nswap; /* the number of times a process was "swapped" out of memory */
1499   long inblock; /* the number of times the file system had to perform input  */
1500   long oublock; /* the number of times the file system had to perform output */
1501   long nvcsw; /* the number of times a context switch was voluntarily      */
1502   long nivcsw; /* the number of times a context switch was forced           */
1503 } kmp_sys_info_t;
1504 
1505 #if USE_ITT_BUILD
1506 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1507 // required type here. Later we will check the type meets requirements.
1508 typedef int kmp_itt_mark_t;
1509 #define KMP_ITT_DEBUG 0
1510 #endif /* USE_ITT_BUILD */
1511 
1512 typedef kmp_int32 kmp_critical_name[8];
1513 
1514 /*!
1515 @ingroup PARALLEL
1516 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1517 The arguments to the outlined function are
1518 @param global_tid the global thread identity of the thread executing the
1519 function.
1520 @param bound_tid  the local identity of the thread executing the function
1521 @param ... pointers to shared variables accessed by the function.
1522 */
1523 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1524 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1525                                  ...);
1526 
1527 /*!
1528 @ingroup THREADPRIVATE
1529 @{
1530 */
1531 /* ---------------------------------------------------------------------------
1532  */
1533 /* Threadprivate initialization/finalization function declarations */
1534 
1535 /*  for non-array objects:  __kmpc_threadprivate_register()  */
1536 
1537 /*!
1538  Pointer to the constructor function.
1539  The first argument is the <tt>this</tt> pointer
1540 */
1541 typedef void *(*kmpc_ctor)(void *);
1542 
1543 /*!
1544  Pointer to the destructor function.
1545  The first argument is the <tt>this</tt> pointer
1546 */
1547 typedef void (*kmpc_dtor)(
1548     void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1549                               compiler */
1550 /*!
1551  Pointer to an alternate constructor.
1552  The first argument is the <tt>this</tt> pointer.
1553 */
1554 typedef void *(*kmpc_cctor)(void *, void *);
1555 
1556 /* for array objects: __kmpc_threadprivate_register_vec() */
1557 /* First arg: "this" pointer */
1558 /* Last arg: number of array elements */
1559 /*!
1560  Array constructor.
1561  First argument is the <tt>this</tt> pointer
1562  Second argument the number of array elements.
1563 */
1564 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1565 /*!
1566  Pointer to the array destructor function.
1567  The first argument is the <tt>this</tt> pointer
1568  Second argument the number of array elements.
1569 */
1570 typedef void (*kmpc_dtor_vec)(void *, size_t);
1571 /*!
1572  Array constructor.
1573  First argument is the <tt>this</tt> pointer
1574  Third argument the number of array elements.
1575 */
1576 typedef void *(*kmpc_cctor_vec)(void *, void *,
1577                                 size_t); /* function unused by compiler */
1578 
1579 /*!
1580 @}
1581 */
1582 
1583 /* keeps tracked of threadprivate cache allocations for cleanup later */
1584 typedef struct kmp_cached_addr {
1585   void **addr; /* address of allocated cache */
1586   void ***compiler_cache; /* pointer to compiler's cache */
1587   void *data; /* pointer to global data */
1588   struct kmp_cached_addr *next; /* pointer to next cached address */
1589 } kmp_cached_addr_t;
1590 
1591 struct private_data {
1592   struct private_data *next; /* The next descriptor in the list      */
1593   void *data; /* The data buffer for this descriptor  */
1594   int more; /* The repeat count for this descriptor */
1595   size_t size; /* The data size for this descriptor    */
1596 };
1597 
1598 struct private_common {
1599   struct private_common *next;
1600   struct private_common *link;
1601   void *gbl_addr;
1602   void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1603   size_t cmn_size;
1604 };
1605 
1606 struct shared_common {
1607   struct shared_common *next;
1608   struct private_data *pod_init;
1609   void *obj_init;
1610   void *gbl_addr;
1611   union {
1612     kmpc_ctor ctor;
1613     kmpc_ctor_vec ctorv;
1614   } ct;
1615   union {
1616     kmpc_cctor cctor;
1617     kmpc_cctor_vec cctorv;
1618   } cct;
1619   union {
1620     kmpc_dtor dtor;
1621     kmpc_dtor_vec dtorv;
1622   } dt;
1623   size_t vec_len;
1624   int is_vec;
1625   size_t cmn_size;
1626 };
1627 
1628 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1629 #define KMP_HASH_TABLE_SIZE                                                    \
1630   (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1631 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1632 #define KMP_HASH(x)                                                            \
1633   ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1634 
1635 struct common_table {
1636   struct private_common *data[KMP_HASH_TABLE_SIZE];
1637 };
1638 
1639 struct shared_table {
1640   struct shared_common *data[KMP_HASH_TABLE_SIZE];
1641 };
1642 
1643 /* ------------------------------------------------------------------------ */
1644 
1645 #if KMP_USE_HIER_SCHED
1646 // Shared barrier data that exists inside a single unit of the scheduling
1647 // hierarchy
1648 typedef struct kmp_hier_private_bdata_t {
1649   kmp_int32 num_active;
1650   kmp_uint64 index;
1651   kmp_uint64 wait_val[2];
1652 } kmp_hier_private_bdata_t;
1653 #endif
1654 
1655 typedef struct kmp_sched_flags {
1656   unsigned ordered : 1;
1657   unsigned nomerge : 1;
1658   unsigned contains_last : 1;
1659 #if KMP_USE_HIER_SCHED
1660   unsigned use_hier : 1;
1661   unsigned unused : 28;
1662 #else
1663   unsigned unused : 29;
1664 #endif
1665 } kmp_sched_flags_t;
1666 
1667 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1668 
1669 #if KMP_STATIC_STEAL_ENABLED
1670 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1671   kmp_int32 count;
1672   kmp_int32 ub;
1673   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1674   kmp_int32 lb;
1675   kmp_int32 st;
1676   kmp_int32 tc;
1677   kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1678                                      after ub */
1679   kmp_lock_t *th_steal_lock; // lock used for chunk stealing
1680   // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1681   //    a) parm3 is properly aligned and
1682   //    b) all parm1-4 are in the same cache line.
1683   // Because of parm1-4 are used together, performance seems to be better
1684   // if they are in the same line (not measured though).
1685 
1686   struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1687     kmp_int32 parm1; //     structures in kmp_dispatch.cpp. This should
1688     kmp_int32 parm2; //     make no real change at least while padding is off.
1689     kmp_int32 parm3;
1690     kmp_int32 parm4;
1691   };
1692 
1693   kmp_uint32 ordered_lower;
1694   kmp_uint32 ordered_upper;
1695 #if KMP_OS_WINDOWS
1696   // This var can be placed in the hole between 'tc' and 'parm1', instead of
1697   // 'static_steal_counter'. It would be nice to measure execution times.
1698   // Conditional if/endif can be removed at all.
1699   kmp_int32 last_upper;
1700 #endif /* KMP_OS_WINDOWS */
1701 } dispatch_private_info32_t;
1702 
1703 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1704   kmp_int64 count; // current chunk number for static & static-steal scheduling
1705   kmp_int64 ub; /* upper-bound */
1706   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1707   kmp_int64 lb; /* lower-bound */
1708   kmp_int64 st; /* stride */
1709   kmp_int64 tc; /* trip count (number of iterations) */
1710   kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1711                                      after ub */
1712   kmp_lock_t *th_steal_lock; // lock used for chunk stealing
1713   /* parm[1-4] are used in different ways by different scheduling algorithms */
1714 
1715   // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1716   //    a) parm3 is properly aligned and
1717   //    b) all parm1-4 are in the same cache line.
1718   // Because of parm1-4 are used together, performance seems to be better
1719   // if they are in the same line (not measured though).
1720 
1721   struct KMP_ALIGN(32) {
1722     kmp_int64 parm1;
1723     kmp_int64 parm2;
1724     kmp_int64 parm3;
1725     kmp_int64 parm4;
1726   };
1727 
1728   kmp_uint64 ordered_lower;
1729   kmp_uint64 ordered_upper;
1730 #if KMP_OS_WINDOWS
1731   // This var can be placed in the hole between 'tc' and 'parm1', instead of
1732   // 'static_steal_counter'. It would be nice to measure execution times.
1733   // Conditional if/endif can be removed at all.
1734   kmp_int64 last_upper;
1735 #endif /* KMP_OS_WINDOWS */
1736 } dispatch_private_info64_t;
1737 #else /* KMP_STATIC_STEAL_ENABLED */
1738 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1739   kmp_int32 lb;
1740   kmp_int32 ub;
1741   kmp_int32 st;
1742   kmp_int32 tc;
1743 
1744   kmp_int32 parm1;
1745   kmp_int32 parm2;
1746   kmp_int32 parm3;
1747   kmp_int32 parm4;
1748 
1749   kmp_int32 count;
1750 
1751   kmp_uint32 ordered_lower;
1752   kmp_uint32 ordered_upper;
1753 #if KMP_OS_WINDOWS
1754   kmp_int32 last_upper;
1755 #endif /* KMP_OS_WINDOWS */
1756 } dispatch_private_info32_t;
1757 
1758 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1759   kmp_int64 lb; /* lower-bound */
1760   kmp_int64 ub; /* upper-bound */
1761   kmp_int64 st; /* stride */
1762   kmp_int64 tc; /* trip count (number of iterations) */
1763 
1764   /* parm[1-4] are used in different ways by different scheduling algorithms */
1765   kmp_int64 parm1;
1766   kmp_int64 parm2;
1767   kmp_int64 parm3;
1768   kmp_int64 parm4;
1769 
1770   kmp_int64 count; /* current chunk number for static scheduling */
1771 
1772   kmp_uint64 ordered_lower;
1773   kmp_uint64 ordered_upper;
1774 #if KMP_OS_WINDOWS
1775   kmp_int64 last_upper;
1776 #endif /* KMP_OS_WINDOWS */
1777 } dispatch_private_info64_t;
1778 #endif /* KMP_STATIC_STEAL_ENABLED */
1779 
1780 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1781   union private_info {
1782     dispatch_private_info32_t p32;
1783     dispatch_private_info64_t p64;
1784   } u;
1785   enum sched_type schedule; /* scheduling algorithm */
1786   kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1787   kmp_int32 ordered_bumped;
1788   // To retain the structure size after making ordered_iteration scalar
1789   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1790   // Stack of buffers for nest of serial regions
1791   struct dispatch_private_info *next;
1792   kmp_int32 type_size; /* the size of types in private_info */
1793 #if KMP_USE_HIER_SCHED
1794   kmp_int32 hier_id;
1795   void *parent; /* hierarchical scheduling parent pointer */
1796 #endif
1797   enum cons_type pushed_ws;
1798 } dispatch_private_info_t;
1799 
1800 typedef struct dispatch_shared_info32 {
1801   /* chunk index under dynamic, number of idle threads under static-steal;
1802      iteration index otherwise */
1803   volatile kmp_uint32 iteration;
1804   volatile kmp_uint32 num_done;
1805   volatile kmp_uint32 ordered_iteration;
1806   // Dummy to retain the structure size after making ordered_iteration scalar
1807   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1808 } dispatch_shared_info32_t;
1809 
1810 typedef struct dispatch_shared_info64 {
1811   /* chunk index under dynamic, number of idle threads under static-steal;
1812      iteration index otherwise */
1813   volatile kmp_uint64 iteration;
1814   volatile kmp_uint64 num_done;
1815   volatile kmp_uint64 ordered_iteration;
1816   // Dummy to retain the structure size after making ordered_iteration scalar
1817   kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1818 } dispatch_shared_info64_t;
1819 
1820 typedef struct dispatch_shared_info {
1821   union shared_info {
1822     dispatch_shared_info32_t s32;
1823     dispatch_shared_info64_t s64;
1824   } u;
1825   volatile kmp_uint32 buffer_index;
1826   volatile kmp_int32 doacross_buf_idx; // teamwise index
1827   volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1828   kmp_int32 doacross_num_done; // count finished threads
1829 #if KMP_USE_HIER_SCHED
1830   void *hier;
1831 #endif
1832 #if KMP_USE_HWLOC
1833   // When linking with libhwloc, the ORDERED EPCC test slows down on big
1834   // machines (> 48 cores). Performance analysis showed that a cache thrash
1835   // was occurring and this padding helps alleviate the problem.
1836   char padding[64];
1837 #endif
1838 } dispatch_shared_info_t;
1839 
1840 typedef struct kmp_disp {
1841   /* Vector for ORDERED SECTION */
1842   void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1843   /* Vector for END ORDERED SECTION */
1844   void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1845 
1846   dispatch_shared_info_t *th_dispatch_sh_current;
1847   dispatch_private_info_t *th_dispatch_pr_current;
1848 
1849   dispatch_private_info_t *th_disp_buffer;
1850   kmp_int32 th_disp_index;
1851   kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1852   volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1853   kmp_int64 *th_doacross_info; // info on loop bounds
1854 #if KMP_USE_INTERNODE_ALIGNMENT
1855   char more_padding[INTERNODE_CACHE_LINE];
1856 #endif
1857 } kmp_disp_t;
1858 
1859 /* ------------------------------------------------------------------------ */
1860 /* Barrier stuff */
1861 
1862 /* constants for barrier state update */
1863 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1864 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1865 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1866 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1867 
1868 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1869 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1870 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1871 
1872 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1873 #error "Barrier sleep bit must be smaller than barrier bump bit"
1874 #endif
1875 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1876 #error "Barrier unused bit must be smaller than barrier bump bit"
1877 #endif
1878 
1879 // Constants for release barrier wait state: currently, hierarchical only
1880 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1881 #define KMP_BARRIER_OWN_FLAG                                                   \
1882   1 // Normal state; worker waiting on own b_go flag in release
1883 #define KMP_BARRIER_PARENT_FLAG                                                \
1884   2 // Special state; worker waiting on parent's b_go flag in release
1885 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG                                         \
1886   3 // Special state; tells worker to shift from parent to own b_go
1887 #define KMP_BARRIER_SWITCHING                                                  \
1888   4 // Special state; worker resets appropriate flag on wake-up
1889 
1890 #define KMP_NOT_SAFE_TO_REAP                                                   \
1891   0 // Thread th_reap_state: not safe to reap (tasking)
1892 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1893 
1894 enum barrier_type {
1895   bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1896                            barriers if enabled) */
1897   bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1898 #if KMP_FAST_REDUCTION_BARRIER
1899   bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1900 #endif // KMP_FAST_REDUCTION_BARRIER
1901   bs_last_barrier /* Just a placeholder to mark the end */
1902 };
1903 
1904 // to work with reduction barriers just like with plain barriers
1905 #if !KMP_FAST_REDUCTION_BARRIER
1906 #define bs_reduction_barrier bs_plain_barrier
1907 #endif // KMP_FAST_REDUCTION_BARRIER
1908 
1909 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1910                            bp_linear_bar =
1911                                0, /* Single level (degenerate) tree */
1912                            bp_tree_bar =
1913                                1, /* Balanced tree with branching factor 2^n */
1914                            bp_hyper_bar =
1915                                2, /* Hypercube-embedded tree with min branching
1916                                      factor 2^n */
1917                            bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1918                            bp_last_bar /* Placeholder to mark the end */
1919 } kmp_bar_pat_e;
1920 
1921 #define KMP_BARRIER_ICV_PUSH 1
1922 
1923 /* Record for holding the values of the internal controls stack records */
1924 typedef struct kmp_internal_control {
1925   int serial_nesting_level; /* corresponds to the value of the
1926                                th_team_serialized field */
1927   kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1928                        thread) */
1929   kmp_int8
1930       bt_set; /* internal control for whether blocktime is explicitly set */
1931   int blocktime; /* internal control for blocktime */
1932 #if KMP_USE_MONITOR
1933   int bt_intervals; /* internal control for blocktime intervals */
1934 #endif
1935   int nproc; /* internal control for #threads for next parallel region (per
1936                 thread) */
1937   int thread_limit; /* internal control for thread-limit-var */
1938   int max_active_levels; /* internal control for max_active_levels */
1939   kmp_r_sched_t
1940       sched; /* internal control for runtime schedule {sched,chunk} pair */
1941   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
1942   kmp_int32 default_device; /* internal control for default device */
1943   struct kmp_internal_control *next;
1944 } kmp_internal_control_t;
1945 
1946 static inline void copy_icvs(kmp_internal_control_t *dst,
1947                              kmp_internal_control_t *src) {
1948   *dst = *src;
1949 }
1950 
1951 /* Thread barrier needs volatile barrier fields */
1952 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1953   // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1954   // uses of it). It is not explicitly aligned below, because we *don't* want
1955   // it to be padded -- instead, we fit b_go into the same cache line with
1956   // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1957   kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1958   // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1959   // same NGO store
1960   volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1961   KMP_ALIGN_CACHE volatile kmp_uint64
1962       b_arrived; // STATE => task reached synch point.
1963   kmp_uint32 *skip_per_level;
1964   kmp_uint32 my_level;
1965   kmp_int32 parent_tid;
1966   kmp_int32 old_tid;
1967   kmp_uint32 depth;
1968   struct kmp_bstate *parent_bar;
1969   kmp_team_t *team;
1970   kmp_uint64 leaf_state;
1971   kmp_uint32 nproc;
1972   kmp_uint8 base_leaf_kids;
1973   kmp_uint8 leaf_kids;
1974   kmp_uint8 offset;
1975   kmp_uint8 wait_flag;
1976   kmp_uint8 use_oncore_barrier;
1977 #if USE_DEBUGGER
1978   // The following field is intended for the debugger solely. Only the worker
1979   // thread itself accesses this field: the worker increases it by 1 when it
1980   // arrives to a barrier.
1981   KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1982 #endif /* USE_DEBUGGER */
1983 } kmp_bstate_t;
1984 
1985 union KMP_ALIGN_CACHE kmp_barrier_union {
1986   double b_align; /* use worst case alignment */
1987   char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1988   kmp_bstate_t bb;
1989 };
1990 
1991 typedef union kmp_barrier_union kmp_balign_t;
1992 
1993 /* Team barrier needs only non-volatile arrived counter */
1994 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1995   double b_align; /* use worst case alignment */
1996   char b_pad[CACHE_LINE];
1997   struct {
1998     kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1999 #if USE_DEBUGGER
2000     // The following two fields are indended for the debugger solely. Only
2001     // master of the team accesses these fields: the first one is increased by
2002     // 1 when master arrives to a barrier, the second one is increased by one
2003     // when all the threads arrived.
2004     kmp_uint b_master_arrived;
2005     kmp_uint b_team_arrived;
2006 #endif
2007   };
2008 };
2009 
2010 typedef union kmp_barrier_team_union kmp_balign_team_t;
2011 
2012 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2013    threads when a condition changes.  This is to workaround an NPTL bug where
2014    padding was added to pthread_cond_t which caused the initialization routine
2015    to write outside of the structure if compiled on pre-NPTL threads.  */
2016 #if KMP_OS_WINDOWS
2017 typedef struct kmp_win32_mutex {
2018   /* The Lock */
2019   CRITICAL_SECTION cs;
2020 } kmp_win32_mutex_t;
2021 
2022 typedef struct kmp_win32_cond {
2023   /* Count of the number of waiters. */
2024   int waiters_count_;
2025 
2026   /* Serialize access to <waiters_count_> */
2027   kmp_win32_mutex_t waiters_count_lock_;
2028 
2029   /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2030   int release_count_;
2031 
2032   /* Keeps track of the current "generation" so that we don't allow */
2033   /* one thread to steal all the "releases" from the broadcast. */
2034   int wait_generation_count_;
2035 
2036   /* A manual-reset event that's used to block and release waiting threads. */
2037   HANDLE event_;
2038 } kmp_win32_cond_t;
2039 #endif
2040 
2041 #if KMP_OS_UNIX
2042 
2043 union KMP_ALIGN_CACHE kmp_cond_union {
2044   double c_align;
2045   char c_pad[CACHE_LINE];
2046   pthread_cond_t c_cond;
2047 };
2048 
2049 typedef union kmp_cond_union kmp_cond_align_t;
2050 
2051 union KMP_ALIGN_CACHE kmp_mutex_union {
2052   double m_align;
2053   char m_pad[CACHE_LINE];
2054   pthread_mutex_t m_mutex;
2055 };
2056 
2057 typedef union kmp_mutex_union kmp_mutex_align_t;
2058 
2059 #endif /* KMP_OS_UNIX */
2060 
2061 typedef struct kmp_desc_base {
2062   void *ds_stackbase;
2063   size_t ds_stacksize;
2064   int ds_stackgrow;
2065   kmp_thread_t ds_thread;
2066   volatile int ds_tid;
2067   int ds_gtid;
2068 #if KMP_OS_WINDOWS
2069   volatile int ds_alive;
2070   DWORD ds_thread_id;
2071 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2072    However, debugger support (libomp_db) cannot work with handles, because they
2073    uncomparable. For example, debugger requests info about thread with handle h.
2074    h is valid within debugger process, and meaningless within debugee process.
2075    Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2076    within debugee process, but it is a *new* handle which does *not* equal to
2077    any other handle in debugee... The only way to compare handles is convert
2078    them to system-wide ids. GetThreadId() function is available only in
2079    Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2080    on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2081    thread id by call to GetCurrentThreadId() from within the thread and save it
2082    to let libomp_db identify threads.  */
2083 #endif /* KMP_OS_WINDOWS */
2084 } kmp_desc_base_t;
2085 
2086 typedef union KMP_ALIGN_CACHE kmp_desc {
2087   double ds_align; /* use worst case alignment */
2088   char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2089   kmp_desc_base_t ds;
2090 } kmp_desc_t;
2091 
2092 typedef struct kmp_local {
2093   volatile int this_construct; /* count of single's encountered by thread */
2094   void *reduce_data;
2095 #if KMP_USE_BGET
2096   void *bget_data;
2097   void *bget_list;
2098 #if !USE_CMP_XCHG_FOR_BGET
2099 #ifdef USE_QUEUING_LOCK_FOR_BGET
2100   kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2101 #else
2102   kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2103 // bootstrap lock so we can use it at library
2104 // shutdown.
2105 #endif /* USE_LOCK_FOR_BGET */
2106 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2107 #endif /* KMP_USE_BGET */
2108 
2109   PACKED_REDUCTION_METHOD_T
2110   packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2111                               __kmpc_end_reduce*() */
2112 
2113 } kmp_local_t;
2114 
2115 #define KMP_CHECK_UPDATE(a, b)                                                 \
2116   if ((a) != (b))                                                              \
2117   (a) = (b)
2118 #define KMP_CHECK_UPDATE_SYNC(a, b)                                            \
2119   if ((a) != (b))                                                              \
2120   TCW_SYNC_PTR((a), (b))
2121 
2122 #define get__blocktime(xteam, xtid)                                            \
2123   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2124 #define get__bt_set(xteam, xtid)                                               \
2125   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2126 #if KMP_USE_MONITOR
2127 #define get__bt_intervals(xteam, xtid)                                         \
2128   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2129 #endif
2130 
2131 #define get__dynamic_2(xteam, xtid)                                            \
2132   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2133 #define get__nproc_2(xteam, xtid)                                              \
2134   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2135 #define get__sched_2(xteam, xtid)                                              \
2136   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2137 
2138 #define set__blocktime_team(xteam, xtid, xval)                                 \
2139   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) =     \
2140        (xval))
2141 
2142 #if KMP_USE_MONITOR
2143 #define set__bt_intervals_team(xteam, xtid, xval)                              \
2144   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) =  \
2145        (xval))
2146 #endif
2147 
2148 #define set__bt_set_team(xteam, xtid, xval)                                    \
2149   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2150 
2151 #define set__dynamic(xthread, xval)                                            \
2152   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2153 #define get__dynamic(xthread)                                                  \
2154   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2155 
2156 #define set__nproc(xthread, xval)                                              \
2157   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2158 
2159 #define set__thread_limit(xthread, xval)                                       \
2160   (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2161 
2162 #define set__max_active_levels(xthread, xval)                                  \
2163   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2164 
2165 #define get__max_active_levels(xthread)                                        \
2166   ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2167 
2168 #define set__sched(xthread, xval)                                              \
2169   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2170 
2171 #define set__proc_bind(xthread, xval)                                          \
2172   (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2173 #define get__proc_bind(xthread)                                                \
2174   ((xthread)->th.th_current_task->td_icvs.proc_bind)
2175 
2176 // OpenMP tasking data structures
2177 
2178 typedef enum kmp_tasking_mode {
2179   tskm_immediate_exec = 0,
2180   tskm_extra_barrier = 1,
2181   tskm_task_teams = 2,
2182   tskm_max = 2
2183 } kmp_tasking_mode_t;
2184 
2185 extern kmp_tasking_mode_t
2186     __kmp_tasking_mode; /* determines how/when to execute tasks */
2187 extern int __kmp_task_stealing_constraint;
2188 extern int __kmp_enable_task_throttling;
2189 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2190 // specified, defaults to 0 otherwise
2191 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2192 extern kmp_int32 __kmp_max_task_priority;
2193 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2194 extern kmp_uint64 __kmp_taskloop_min_tasks;
2195 
2196 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2197    taskdata first */
2198 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2199 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2200 
2201 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2202 // were spawned and queued since the previous barrier release.
2203 #define KMP_TASKING_ENABLED(task_team)                                         \
2204   (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2205 /*!
2206 @ingroup BASIC_TYPES
2207 @{
2208 */
2209 
2210 /*!
2211  */
2212 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2213 
2214 typedef union kmp_cmplrdata {
2215   kmp_int32 priority; /**< priority specified by user for the task */
2216   kmp_routine_entry_t
2217       destructors; /* pointer to function to invoke deconstructors of
2218                       firstprivate C++ objects */
2219   /* future data */
2220 } kmp_cmplrdata_t;
2221 
2222 /*  sizeof_kmp_task_t passed as arg to kmpc_omp_task call  */
2223 /*!
2224  */
2225 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2226   void *shareds; /**< pointer to block of pointers to shared vars   */
2227   kmp_routine_entry_t
2228       routine; /**< pointer to routine to call for executing task */
2229   kmp_int32 part_id; /**< part id for the task                          */
2230   kmp_cmplrdata_t
2231       data1; /* Two known optional additions: destructors and priority */
2232   kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2233   /* future data */
2234   /*  private vars  */
2235 } kmp_task_t;
2236 
2237 /*!
2238 @}
2239 */
2240 
2241 typedef struct kmp_taskgroup {
2242   std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2243   std::atomic<kmp_int32>
2244       cancel_request; // request for cancellation of this taskgroup
2245   struct kmp_taskgroup *parent; // parent taskgroup
2246   // Block of data to perform task reduction
2247   void *reduce_data; // reduction related info
2248   kmp_int32 reduce_num_data; // number of data items to reduce
2249 } kmp_taskgroup_t;
2250 
2251 // forward declarations
2252 typedef union kmp_depnode kmp_depnode_t;
2253 typedef struct kmp_depnode_list kmp_depnode_list_t;
2254 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2255 
2256 // Compiler sends us this info:
2257 typedef struct kmp_depend_info {
2258   kmp_intptr_t base_addr;
2259   size_t len;
2260   struct {
2261     bool in : 1;
2262     bool out : 1;
2263     bool mtx : 1;
2264   } flags;
2265 } kmp_depend_info_t;
2266 
2267 // Internal structures to work with task dependencies:
2268 struct kmp_depnode_list {
2269   kmp_depnode_t *node;
2270   kmp_depnode_list_t *next;
2271 };
2272 
2273 // Max number of mutexinoutset dependencies per node
2274 #define MAX_MTX_DEPS 4
2275 
2276 typedef struct kmp_base_depnode {
2277   kmp_depnode_list_t *successors; /* used under lock */
2278   kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2279   kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2280   kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2281   kmp_lock_t lock; /* guards shared fields: task, successors */
2282 #if KMP_SUPPORT_GRAPH_OUTPUT
2283   kmp_uint32 id;
2284 #endif
2285   std::atomic<kmp_int32> npredecessors;
2286   std::atomic<kmp_int32> nrefs;
2287 } kmp_base_depnode_t;
2288 
2289 union KMP_ALIGN_CACHE kmp_depnode {
2290   double dn_align; /* use worst case alignment */
2291   char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2292   kmp_base_depnode_t dn;
2293 };
2294 
2295 struct kmp_dephash_entry {
2296   kmp_intptr_t addr;
2297   kmp_depnode_t *last_out;
2298   kmp_depnode_list_t *last_ins;
2299   kmp_depnode_list_t *last_mtxs;
2300   kmp_int32 last_flag;
2301   kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2302   kmp_dephash_entry_t *next_in_bucket;
2303 };
2304 
2305 typedef struct kmp_dephash {
2306   kmp_dephash_entry_t **buckets;
2307   size_t size;
2308   size_t generation;
2309   kmp_uint32 nelements;
2310   kmp_uint32 nconflicts;
2311 } kmp_dephash_t;
2312 
2313 typedef struct kmp_task_affinity_info {
2314   kmp_intptr_t base_addr;
2315   size_t len;
2316   struct {
2317     bool flag1 : 1;
2318     bool flag2 : 1;
2319     kmp_int32 reserved : 30;
2320   } flags;
2321 } kmp_task_affinity_info_t;
2322 
2323 typedef enum kmp_event_type_t {
2324   KMP_EVENT_UNINITIALIZED = 0,
2325   KMP_EVENT_ALLOW_COMPLETION = 1
2326 } kmp_event_type_t;
2327 
2328 typedef struct {
2329   kmp_event_type_t type;
2330   kmp_tas_lock_t lock;
2331   union {
2332     kmp_task_t *task;
2333   } ed;
2334 } kmp_event_t;
2335 
2336 #ifdef BUILD_TIED_TASK_STACK
2337 
2338 /* Tied Task stack definitions */
2339 typedef struct kmp_stack_block {
2340   kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2341   struct kmp_stack_block *sb_next;
2342   struct kmp_stack_block *sb_prev;
2343 } kmp_stack_block_t;
2344 
2345 typedef struct kmp_task_stack {
2346   kmp_stack_block_t ts_first_block; // first block of stack entries
2347   kmp_taskdata_t **ts_top; // pointer to the top of stack
2348   kmp_int32 ts_entries; // number of entries on the stack
2349 } kmp_task_stack_t;
2350 
2351 #endif // BUILD_TIED_TASK_STACK
2352 
2353 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2354   /* Compiler flags */ /* Total compiler flags must be 16 bits */
2355   unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2356   unsigned final : 1; /* task is final(1) so execute immediately */
2357   unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2358                               code path */
2359   unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2360                                      invoke destructors from the runtime */
2361   unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2362                          context of the RTL) */
2363   unsigned priority_specified : 1; /* set if the compiler provides priority
2364                                       setting for the task */
2365   unsigned detachable : 1; /* 1 == can detach */
2366   unsigned hidden_helper : 1; /* 1 == hidden helper task */
2367   unsigned reserved : 8; /* reserved for compiler use */
2368 
2369   /* Library flags */ /* Total library flags must be 16 bits */
2370   unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2371   unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2372   unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2373   // (1) or may be deferred (0)
2374   unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2375   // (0) [>= 2 threads]
2376   /* If either team_serial or tasking_ser is set, task team may be NULL */
2377   /* Task State Flags: */
2378   unsigned started : 1; /* 1==started, 0==not started     */
2379   unsigned executing : 1; /* 1==executing, 0==not executing */
2380   unsigned complete : 1; /* 1==complete, 0==not complete   */
2381   unsigned freed : 1; /* 1==freed, 0==allocated        */
2382   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2383   unsigned reserved31 : 7; /* reserved for library use */
2384 
2385 } kmp_tasking_flags_t;
2386 
2387 struct kmp_taskdata { /* aligned during dynamic allocation       */
2388   kmp_int32 td_task_id; /* id, assigned by debugger                */
2389   kmp_tasking_flags_t td_flags; /* task flags                              */
2390   kmp_team_t *td_team; /* team for this task                      */
2391   kmp_info_p *td_alloc_thread; /* thread that allocated data structures   */
2392   /* Currently not used except for perhaps IDB */
2393   kmp_taskdata_t *td_parent; /* parent task                             */
2394   kmp_int32 td_level; /* task nesting level                      */
2395   std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2396   ident_t *td_ident; /* task identifier                         */
2397   // Taskwait data.
2398   ident_t *td_taskwait_ident;
2399   kmp_uint32 td_taskwait_counter;
2400   kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2401   KMP_ALIGN_CACHE kmp_internal_control_t
2402       td_icvs; /* Internal control variables for the task */
2403   KMP_ALIGN_CACHE std::atomic<kmp_int32>
2404       td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2405                                    deallocated */
2406   std::atomic<kmp_int32>
2407       td_incomplete_child_tasks; /* Child tasks not yet complete */
2408   kmp_taskgroup_t
2409       *td_taskgroup; // Each task keeps pointer to its current taskgroup
2410   kmp_dephash_t
2411       *td_dephash; // Dependencies for children tasks are tracked from here
2412   kmp_depnode_t
2413       *td_depnode; // Pointer to graph node if this task has dependencies
2414   kmp_task_team_t *td_task_team;
2415   // The global thread id of the encountering thread. We need it because when a
2416   // regular task depends on a hidden helper task, and the hidden helper task
2417   // is finished on a hidden helper thread, it will call __kmp_release_deps to
2418   // release all dependences. If now the task is a regular task, we need to pass
2419   // the encountering gtid such that the task will be picked up and executed by
2420   // its encountering team instead of hidden helper team.
2421   kmp_int32 encountering_gtid;
2422   size_t td_size_alloc; // Size of task structure, including shareds etc.
2423 #if defined(KMP_GOMP_COMPAT)
2424   // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2425   kmp_int32 td_size_loop_bounds;
2426 #endif
2427   kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2428 #if defined(KMP_GOMP_COMPAT)
2429   // GOMP sends in a copy function for copy constructors
2430   void (*td_copy_func)(void *, void *);
2431 #endif
2432   kmp_event_t td_allow_completion_event;
2433 #if OMPT_SUPPORT
2434   ompt_task_info_t ompt_task_info;
2435 #endif
2436 }; // struct kmp_taskdata
2437 
2438 // Make sure padding above worked
2439 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2440 
2441 // Data for task team but per thread
2442 typedef struct kmp_base_thread_data {
2443   kmp_info_p *td_thr; // Pointer back to thread info
2444   // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2445   // queued?
2446   kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2447   kmp_taskdata_t *
2448       *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2449   kmp_int32 td_deque_size; // Size of deck
2450   kmp_uint32 td_deque_head; // Head of deque (will wrap)
2451   kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2452   kmp_int32 td_deque_ntasks; // Number of tasks in deque
2453   // GEH: shouldn't this be volatile since used in while-spin?
2454   kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2455 #ifdef BUILD_TIED_TASK_STACK
2456   kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2457 // scheduling constraint
2458 #endif // BUILD_TIED_TASK_STACK
2459 } kmp_base_thread_data_t;
2460 
2461 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2462 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2463 
2464 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2465 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2466 
2467 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2468   kmp_base_thread_data_t td;
2469   double td_align; /* use worst case alignment */
2470   char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2471 } kmp_thread_data_t;
2472 
2473 // Data for task teams which are used when tasking is enabled for the team
2474 typedef struct kmp_base_task_team {
2475   kmp_bootstrap_lock_t
2476       tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2477   /* must be bootstrap lock since used at library shutdown*/
2478   kmp_task_team_t *tt_next; /* For linking the task team free list */
2479   kmp_thread_data_t
2480       *tt_threads_data; /* Array of per-thread structures for task team */
2481   /* Data survives task team deallocation */
2482   kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2483                                executing this team? */
2484   /* TRUE means tt_threads_data is set up and initialized */
2485   kmp_int32 tt_nproc; /* #threads in team           */
2486   kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2487   kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2488   kmp_int32 tt_untied_task_encountered;
2489   // There is hidden helper thread encountered in this task team so that we must
2490   // wait when waiting on task team
2491   kmp_int32 tt_hidden_helper_task_encountered;
2492 
2493   KMP_ALIGN_CACHE
2494   std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2495 
2496   KMP_ALIGN_CACHE
2497   volatile kmp_uint32
2498       tt_active; /* is the team still actively executing tasks */
2499 } kmp_base_task_team_t;
2500 
2501 union KMP_ALIGN_CACHE kmp_task_team {
2502   kmp_base_task_team_t tt;
2503   double tt_align; /* use worst case alignment */
2504   char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2505 };
2506 
2507 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2508 // Free lists keep same-size free memory slots for fast memory allocation
2509 // routines
2510 typedef struct kmp_free_list {
2511   void *th_free_list_self; // Self-allocated tasks free list
2512   void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2513   // threads
2514   void *th_free_list_other; // Non-self free list (to be returned to owner's
2515   // sync list)
2516 } kmp_free_list_t;
2517 #endif
2518 #if KMP_NESTED_HOT_TEAMS
2519 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2520 // are not put in teams pool, and they don't put threads in threads pool.
2521 typedef struct kmp_hot_team_ptr {
2522   kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2523   kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2524 } kmp_hot_team_ptr_t;
2525 #endif
2526 typedef struct kmp_teams_size {
2527   kmp_int32 nteams; // number of teams in a league
2528   kmp_int32 nth; // number of threads in each team of the league
2529 } kmp_teams_size_t;
2530 
2531 // This struct stores a thread that acts as a "root" for a contention
2532 // group. Contention groups are rooted at kmp_root threads, but also at
2533 // each master thread of each team created in the teams construct.
2534 // This struct therefore also stores a thread_limit associated with
2535 // that contention group, and a counter to track the number of threads
2536 // active in that contention group. Each thread has a list of these: CG
2537 // root threads have an entry in their list in which cg_root refers to
2538 // the thread itself, whereas other workers in the CG will have a
2539 // single entry where cg_root is same as the entry containing their CG
2540 // root. When a thread encounters a teams construct, it will add a new
2541 // entry to the front of its list, because it now roots a new CG.
2542 typedef struct kmp_cg_root {
2543   kmp_info_p *cg_root; // "root" thread for a contention group
2544   // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2545   // thread_limit clause for teams masters
2546   kmp_int32 cg_thread_limit;
2547   kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2548   struct kmp_cg_root *up; // pointer to higher level CG root in list
2549 } kmp_cg_root_t;
2550 
2551 // OpenMP thread data structures
2552 
2553 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2554   /* Start with the readonly data which is cache aligned and padded. This is
2555      written before the thread starts working by the master. Uber masters may
2556      update themselves later. Usage does not consider serialized regions.  */
2557   kmp_desc_t th_info;
2558   kmp_team_p *th_team; /* team we belong to */
2559   kmp_root_p *th_root; /* pointer to root of task hierarchy */
2560   kmp_info_p *th_next_pool; /* next available thread in the pool */
2561   kmp_disp_t *th_dispatch; /* thread's dispatch data */
2562   int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2563 
2564   /* The following are cached from the team info structure */
2565   /* TODO use these in more places as determined to be needed via profiling */
2566   int th_team_nproc; /* number of threads in a team */
2567   kmp_info_p *th_team_master; /* the team's master thread */
2568   int th_team_serialized; /* team is serialized */
2569   microtask_t th_teams_microtask; /* save entry address for teams construct */
2570   int th_teams_level; /* save initial level of teams construct */
2571 /* it is 0 on device but may be any on host */
2572 
2573 /* The blocktime info is copied from the team struct to the thread struct */
2574 /* at the start of a barrier, and the values stored in the team are used  */
2575 /* at points in the code where the team struct is no longer guaranteed    */
2576 /* to exist (from the POV of worker threads).                             */
2577 #if KMP_USE_MONITOR
2578   int th_team_bt_intervals;
2579   int th_team_bt_set;
2580 #else
2581   kmp_uint64 th_team_bt_intervals;
2582 #endif
2583 
2584 #if KMP_AFFINITY_SUPPORTED
2585   kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2586 #endif
2587   omp_allocator_handle_t th_def_allocator; /* default allocator */
2588   /* The data set by the master at reinit, then R/W by the worker */
2589   KMP_ALIGN_CACHE int
2590       th_set_nproc; /* if > 0, then only use this request for the next fork */
2591 #if KMP_NESTED_HOT_TEAMS
2592   kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2593 #endif
2594   kmp_proc_bind_t
2595       th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2596   kmp_teams_size_t
2597       th_teams_size; /* number of teams/threads in teams construct */
2598 #if KMP_AFFINITY_SUPPORTED
2599   int th_current_place; /* place currently bound to */
2600   int th_new_place; /* place to bind to in par reg */
2601   int th_first_place; /* first place in partition */
2602   int th_last_place; /* last place in partition */
2603 #endif
2604   int th_prev_level; /* previous level for affinity format */
2605   int th_prev_num_threads; /* previous num_threads for affinity format */
2606 #if USE_ITT_BUILD
2607   kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2608   kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2609   kmp_uint64 th_frame_time; /* frame timestamp */
2610 #endif /* USE_ITT_BUILD */
2611   kmp_local_t th_local;
2612   struct private_common *th_pri_head;
2613 
2614   /* Now the data only used by the worker (after initial allocation) */
2615   /* TODO the first serial team should actually be stored in the info_t
2616      structure.  this will help reduce initial allocation overhead */
2617   KMP_ALIGN_CACHE kmp_team_p
2618       *th_serial_team; /*serialized team held in reserve*/
2619 
2620 #if OMPT_SUPPORT
2621   ompt_thread_info_t ompt_thread_info;
2622 #endif
2623 
2624   /* The following are also read by the master during reinit */
2625   struct common_table *th_pri_common;
2626 
2627   volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2628   /* while awaiting queuing lock acquire */
2629 
2630   volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2631 
2632   ident_t *th_ident;
2633   unsigned th_x; // Random number generator data
2634   unsigned th_a; // Random number generator data
2635 
2636   /* Tasking-related data for the thread */
2637   kmp_task_team_t *th_task_team; // Task team struct
2638   kmp_taskdata_t *th_current_task; // Innermost Task being executed
2639   kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2640   kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2641   // at nested levels
2642   kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2643   kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2644   kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2645   // tasking, thus safe to reap
2646 
2647   /* More stuff for keeping track of active/sleeping threads (this part is
2648      written by the worker thread) */
2649   kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2650   int th_active; // ! sleeping; 32 bits for TCR/TCW
2651   struct cons_header *th_cons; // used for consistency check
2652 #if KMP_USE_HIER_SCHED
2653   // used for hierarchical scheduling
2654   kmp_hier_private_bdata_t *th_hier_bar_data;
2655 #endif
2656 
2657   /* Add the syncronizing data which is cache aligned and padded. */
2658   KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2659 
2660   KMP_ALIGN_CACHE volatile kmp_int32
2661       th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2662 
2663 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2664 #define NUM_LISTS 4
2665   kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2666 // allocation routines
2667 #endif
2668 
2669 #if KMP_OS_WINDOWS
2670   kmp_win32_cond_t th_suspend_cv;
2671   kmp_win32_mutex_t th_suspend_mx;
2672   std::atomic<int> th_suspend_init;
2673 #endif
2674 #if KMP_OS_UNIX
2675   kmp_cond_align_t th_suspend_cv;
2676   kmp_mutex_align_t th_suspend_mx;
2677   std::atomic<int> th_suspend_init_count;
2678 #endif
2679 
2680 #if USE_ITT_BUILD
2681   kmp_itt_mark_t th_itt_mark_single;
2682 // alignment ???
2683 #endif /* USE_ITT_BUILD */
2684 #if KMP_STATS_ENABLED
2685   kmp_stats_list *th_stats;
2686 #endif
2687 #if KMP_OS_UNIX
2688   std::atomic<bool> th_blocking;
2689 #endif
2690   kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2691 } kmp_base_info_t;
2692 
2693 typedef union KMP_ALIGN_CACHE kmp_info {
2694   double th_align; /* use worst case alignment */
2695   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2696   kmp_base_info_t th;
2697 } kmp_info_t;
2698 
2699 // OpenMP thread team data structures
2700 
2701 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2702 
2703 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2704   double dt_align; /* use worst case alignment */
2705   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2706   kmp_base_data_t dt;
2707 } kmp_sleep_team_t;
2708 
2709 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2710   double dt_align; /* use worst case alignment */
2711   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2712   kmp_base_data_t dt;
2713 } kmp_ordered_team_t;
2714 
2715 typedef int (*launch_t)(int gtid);
2716 
2717 /* Minimum number of ARGV entries to malloc if necessary */
2718 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2719 
2720 // Set up how many argv pointers will fit in cache lines containing
2721 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2722 // larger value for more space between the master write/worker read section and
2723 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2724 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2725 #define KMP_INLINE_ARGV_BYTES                                                  \
2726   (4 * CACHE_LINE -                                                            \
2727    ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) +               \
2728      sizeof(kmp_int16) + sizeof(kmp_uint32)) %                                 \
2729     CACHE_LINE))
2730 #else
2731 #define KMP_INLINE_ARGV_BYTES                                                  \
2732   (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2733 #endif
2734 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2735 
2736 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2737   // Synchronization Data
2738   // ---------------------------------------------------------------------------
2739   KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2740   kmp_balign_team_t t_bar[bs_last_barrier];
2741   std::atomic<int> t_construct; // count of single directive encountered by team
2742   char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2743 
2744   // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2745   std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2746   std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2747 
2748   // Master only
2749   // ---------------------------------------------------------------------------
2750   KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2751   int t_master_this_cons; // "this_construct" single counter of master in parent
2752   // team
2753   ident_t *t_ident; // if volatile, have to change too much other crud to
2754   // volatile too
2755   kmp_team_p *t_parent; // parent team
2756   kmp_team_p *t_next_pool; // next free team in the team pool
2757   kmp_disp_t *t_dispatch; // thread's dispatch data
2758   kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2759   kmp_proc_bind_t t_proc_bind; // bind type for par region
2760 #if USE_ITT_BUILD
2761   kmp_uint64 t_region_time; // region begin timestamp
2762 #endif /* USE_ITT_BUILD */
2763 
2764   // Master write, workers read
2765   // --------------------------------------------------------------------------
2766   KMP_ALIGN_CACHE void **t_argv;
2767   int t_argc;
2768   int t_nproc; // number of threads in team
2769   microtask_t t_pkfn;
2770   launch_t t_invoke; // procedure to launch the microtask
2771 
2772 #if OMPT_SUPPORT
2773   ompt_team_info_t ompt_team_info;
2774   ompt_lw_taskteam_t *ompt_serialized_team_info;
2775 #endif
2776 
2777 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2778   kmp_int8 t_fp_control_saved;
2779   kmp_int8 t_pad2b;
2780   kmp_int16 t_x87_fpu_control_word; // FP control regs
2781   kmp_uint32 t_mxcsr;
2782 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2783 
2784   void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2785 
2786   KMP_ALIGN_CACHE kmp_info_t **t_threads;
2787   kmp_taskdata_t
2788       *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2789   int t_level; // nested parallel level
2790 
2791   KMP_ALIGN_CACHE int t_max_argc;
2792   int t_max_nproc; // max threads this team can handle (dynamically expandable)
2793   int t_serialized; // levels deep of serialized teams
2794   dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2795   int t_id; // team's id, assigned by debugger.
2796   int t_active_level; // nested active parallel level
2797   kmp_r_sched_t t_sched; // run-time schedule for the team
2798 #if KMP_AFFINITY_SUPPORTED
2799   int t_first_place; // first & last place in parent thread's partition.
2800   int t_last_place; // Restore these values to master after par region.
2801 #endif // KMP_AFFINITY_SUPPORTED
2802   int t_display_affinity;
2803   int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2804   // omp_set_num_threads() call
2805   omp_allocator_handle_t t_def_allocator; /* default allocator */
2806 
2807 // Read/write by workers as well
2808 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2809   // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2810   // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2811   // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2812   // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2813   char dummy_padding[1024];
2814 #endif
2815   // Internal control stack for additional nested teams.
2816   KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2817   // for SERIALIZED teams nested 2 or more levels deep
2818   // typed flag to store request state of cancellation
2819   std::atomic<kmp_int32> t_cancel_request;
2820   int t_master_active; // save on fork, restore on join
2821   void *t_copypriv_data; // team specific pointer to copyprivate data array
2822 #if KMP_OS_WINDOWS
2823   std::atomic<kmp_uint32> t_copyin_counter;
2824 #endif
2825 #if USE_ITT_BUILD
2826   void *t_stack_id; // team specific stack stitching id (for ittnotify)
2827 #endif /* USE_ITT_BUILD */
2828 } kmp_base_team_t;
2829 
2830 union KMP_ALIGN_CACHE kmp_team {
2831   kmp_base_team_t t;
2832   double t_align; /* use worst case alignment */
2833   char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2834 };
2835 
2836 typedef union KMP_ALIGN_CACHE kmp_time_global {
2837   double dt_align; /* use worst case alignment */
2838   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2839   kmp_base_data_t dt;
2840 } kmp_time_global_t;
2841 
2842 typedef struct kmp_base_global {
2843   /* cache-aligned */
2844   kmp_time_global_t g_time;
2845 
2846   /* non cache-aligned */
2847   volatile int g_abort;
2848   volatile int g_done;
2849 
2850   int g_dynamic;
2851   enum dynamic_mode g_dynamic_mode;
2852 } kmp_base_global_t;
2853 
2854 typedef union KMP_ALIGN_CACHE kmp_global {
2855   kmp_base_global_t g;
2856   double g_align; /* use worst case alignment */
2857   char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2858 } kmp_global_t;
2859 
2860 typedef struct kmp_base_root {
2861   // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2862   // (r_in_parallel>= 0)
2863   // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2864   // the synch overhead or keeping r_active
2865   volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2866   // keeps a count of active parallel regions per root
2867   std::atomic<int> r_in_parallel;
2868   // GEH: This is misnamed, should be r_active_levels
2869   kmp_team_t *r_root_team;
2870   kmp_team_t *r_hot_team;
2871   kmp_info_t *r_uber_thread;
2872   kmp_lock_t r_begin_lock;
2873   volatile int r_begin;
2874   int r_blocktime; /* blocktime for this root and descendants */
2875 } kmp_base_root_t;
2876 
2877 typedef union KMP_ALIGN_CACHE kmp_root {
2878   kmp_base_root_t r;
2879   double r_align; /* use worst case alignment */
2880   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2881 } kmp_root_t;
2882 
2883 struct fortran_inx_info {
2884   kmp_int32 data;
2885 };
2886 
2887 /* ------------------------------------------------------------------------ */
2888 
2889 extern int __kmp_settings;
2890 extern int __kmp_duplicate_library_ok;
2891 #if USE_ITT_BUILD
2892 extern int __kmp_forkjoin_frames;
2893 extern int __kmp_forkjoin_frames_mode;
2894 #endif
2895 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2896 extern int __kmp_determ_red;
2897 
2898 #ifdef KMP_DEBUG
2899 extern int kmp_a_debug;
2900 extern int kmp_b_debug;
2901 extern int kmp_c_debug;
2902 extern int kmp_d_debug;
2903 extern int kmp_e_debug;
2904 extern int kmp_f_debug;
2905 #endif /* KMP_DEBUG */
2906 
2907 /* For debug information logging using rotating buffer */
2908 #define KMP_DEBUG_BUF_LINES_INIT 512
2909 #define KMP_DEBUG_BUF_LINES_MIN 1
2910 
2911 #define KMP_DEBUG_BUF_CHARS_INIT 128
2912 #define KMP_DEBUG_BUF_CHARS_MIN 2
2913 
2914 extern int
2915     __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2916 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2917 extern int
2918     __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2919 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2920                                       entry pointer */
2921 
2922 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2923 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2924                                               printed in buffer so far */
2925 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2926                                           recommended in warnings */
2927 /* end rotating debug buffer */
2928 
2929 #ifdef KMP_DEBUG
2930 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2931 
2932 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2933 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2934 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2935 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2936 extern int __kmp_par_range_lb;
2937 extern int __kmp_par_range_ub;
2938 #endif
2939 
2940 /* For printing out dynamic storage map for threads and teams */
2941 extern int
2942     __kmp_storage_map; /* True means print storage map for threads and teams */
2943 extern int __kmp_storage_map_verbose; /* True means storage map includes
2944                                          placement info */
2945 extern int __kmp_storage_map_verbose_specified;
2946 
2947 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2948 extern kmp_cpuinfo_t __kmp_cpuinfo;
2949 #endif
2950 
2951 extern volatile int __kmp_init_serial;
2952 extern volatile int __kmp_init_gtid;
2953 extern volatile int __kmp_init_common;
2954 extern volatile int __kmp_init_middle;
2955 extern volatile int __kmp_init_parallel;
2956 #if KMP_USE_MONITOR
2957 extern volatile int __kmp_init_monitor;
2958 #endif
2959 extern volatile int __kmp_init_user_locks;
2960 extern volatile int __kmp_init_hidden_helper_threads;
2961 extern int __kmp_init_counter;
2962 extern int __kmp_root_counter;
2963 extern int __kmp_version;
2964 
2965 /* list of address of allocated caches for commons */
2966 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2967 
2968 /* Barrier algorithm types and options */
2969 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2970 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2971 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2972 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2973 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2974 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2975 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2976 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2977 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2978 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2979 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2980 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2981 
2982 /* Global Locks */
2983 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2984 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2985 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2986 extern kmp_bootstrap_lock_t
2987     __kmp_exit_lock; /* exit() is not always thread-safe */
2988 #if KMP_USE_MONITOR
2989 extern kmp_bootstrap_lock_t
2990     __kmp_monitor_lock; /* control monitor thread creation */
2991 #endif
2992 extern kmp_bootstrap_lock_t
2993     __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2994                              __kmp_threads expansion to co-exist */
2995 
2996 extern kmp_lock_t __kmp_global_lock; /* control OS/global access  */
2997 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access  */
2998 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2999 
3000 extern enum library_type __kmp_library;
3001 
3002 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3003 extern enum sched_type __kmp_static; /* default static scheduling method */
3004 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3005 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3006 extern int __kmp_chunk; /* default runtime chunk size */
3007 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3008 
3009 extern size_t __kmp_stksize; /* stack size per thread         */
3010 #if KMP_USE_MONITOR
3011 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3012 #endif
3013 extern size_t __kmp_stkoffset; /* stack offset per thread       */
3014 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3015 
3016 extern size_t
3017     __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3018 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3019 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3020 extern int __kmp_env_checks; /* was KMP_CHECKS specified?    */
3021 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3022 extern int __kmp_generate_warnings; /* should we issue warnings? */
3023 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3024 
3025 #ifdef DEBUG_SUSPEND
3026 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3027 #endif
3028 
3029 extern kmp_int32 __kmp_use_yield;
3030 extern kmp_int32 __kmp_use_yield_exp_set;
3031 extern kmp_uint32 __kmp_yield_init;
3032 extern kmp_uint32 __kmp_yield_next;
3033 
3034 /* ------------------------------------------------------------------------- */
3035 extern int __kmp_allThreadsSpecified;
3036 
3037 extern size_t __kmp_align_alloc;
3038 /* following data protected by initialization routines */
3039 extern int __kmp_xproc; /* number of processors in the system */
3040 extern int __kmp_avail_proc; /* number of processors available to the process */
3041 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3042 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3043 // maximum total number of concurrently-existing threads on device
3044 extern int __kmp_max_nth;
3045 // maximum total number of concurrently-existing threads in a contention group
3046 extern int __kmp_cg_max_nth;
3047 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3048 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3049                                       __kmp_root */
3050 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3051                                    region a la OMP_NUM_THREADS */
3052 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3053                                       initialization */
3054 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3055                                  used (fixed) */
3056 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3057                                (__kmpc_threadprivate_cached()) */
3058 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3059                                     blocking (env setting) */
3060 #if KMP_USE_MONITOR
3061 extern int
3062     __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3063 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3064                                   blocking */
3065 #endif
3066 #ifdef KMP_ADJUST_BLOCKTIME
3067 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3068 #endif /* KMP_ADJUST_BLOCKTIME */
3069 #ifdef KMP_DFLT_NTH_CORES
3070 extern int __kmp_ncores; /* Total number of cores for threads placement */
3071 #endif
3072 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3073 extern int __kmp_abort_delay;
3074 
3075 extern int __kmp_need_register_atfork_specified;
3076 extern int
3077     __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
3078                                    install fork handler */
3079 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3080                                0 - not set, will be set at runtime
3081                                1 - using stack search
3082                                2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3083                                    X*) or TlsGetValue(Windows* OS))
3084                                3 - static TLS (__declspec(thread) __kmp_gtid),
3085                                    Linux* OS .so only.  */
3086 extern int
3087     __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3088 #ifdef KMP_TDATA_GTID
3089 extern KMP_THREAD_LOCAL int __kmp_gtid;
3090 #endif
3091 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3092 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3093 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3094 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3095 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3096 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3097 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3098 
3099 // max_active_levels for nested parallelism enabled by default via
3100 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3101 extern int __kmp_dflt_max_active_levels;
3102 // Indicates whether value of __kmp_dflt_max_active_levels was already
3103 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3104 extern bool __kmp_dflt_max_active_levels_set;
3105 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3106                                           concurrent execution per team */
3107 #if KMP_NESTED_HOT_TEAMS
3108 extern int __kmp_hot_teams_mode;
3109 extern int __kmp_hot_teams_max_level;
3110 #endif
3111 
3112 #if KMP_OS_LINUX
3113 extern enum clock_function_type __kmp_clock_function;
3114 extern int __kmp_clock_function_param;
3115 #endif /* KMP_OS_LINUX */
3116 
3117 #if KMP_MIC_SUPPORTED
3118 extern enum mic_type __kmp_mic_type;
3119 #endif
3120 
3121 #ifdef USE_LOAD_BALANCE
3122 extern double __kmp_load_balance_interval; // load balance algorithm interval
3123 #endif /* USE_LOAD_BALANCE */
3124 
3125 // OpenMP 3.1 - Nested num threads array
3126 typedef struct kmp_nested_nthreads_t {
3127   int *nth;
3128   int size;
3129   int used;
3130 } kmp_nested_nthreads_t;
3131 
3132 extern kmp_nested_nthreads_t __kmp_nested_nth;
3133 
3134 #if KMP_USE_ADAPTIVE_LOCKS
3135 
3136 // Parameters for the speculative lock backoff system.
3137 struct kmp_adaptive_backoff_params_t {
3138   // Number of soft retries before it counts as a hard retry.
3139   kmp_uint32 max_soft_retries;
3140   // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3141   // the right
3142   kmp_uint32 max_badness;
3143 };
3144 
3145 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3146 
3147 #if KMP_DEBUG_ADAPTIVE_LOCKS
3148 extern const char *__kmp_speculative_statsfile;
3149 #endif
3150 
3151 #endif // KMP_USE_ADAPTIVE_LOCKS
3152 
3153 extern int __kmp_display_env; /* TRUE or FALSE */
3154 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3155 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3156 extern int __kmp_nteams;
3157 extern int __kmp_teams_thread_limit;
3158 
3159 /* ------------------------------------------------------------------------- */
3160 
3161 /* the following are protected by the fork/join lock */
3162 /* write: lock  read: anytime */
3163 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3164 /* read/write: lock */
3165 extern volatile kmp_team_t *__kmp_team_pool;
3166 extern volatile kmp_info_t *__kmp_thread_pool;
3167 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3168 
3169 // total num threads reachable from some root thread including all root threads
3170 extern volatile int __kmp_nth;
3171 /* total number of threads reachable from some root thread including all root
3172    threads, and those in the thread pool */
3173 extern volatile int __kmp_all_nth;
3174 extern std::atomic<int> __kmp_thread_pool_active_nth;
3175 
3176 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3177 /* end data protected by fork/join lock */
3178 /* ------------------------------------------------------------------------- */
3179 
3180 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3181 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3182 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3183 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3184 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3185 
3186 // AT: Which way is correct?
3187 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3188 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3189 #define __kmp_get_team_num_threads(gtid)                                       \
3190   (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3191 
3192 static inline bool KMP_UBER_GTID(int gtid) {
3193   KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3194   KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3195   return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3196           __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3197 }
3198 
3199 static inline int __kmp_tid_from_gtid(int gtid) {
3200   KMP_DEBUG_ASSERT(gtid >= 0);
3201   return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3202 }
3203 
3204 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3205   KMP_DEBUG_ASSERT(tid >= 0 && team);
3206   return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3207 }
3208 
3209 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3210   KMP_DEBUG_ASSERT(thr);
3211   return thr->th.th_info.ds.ds_gtid;
3212 }
3213 
3214 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3215   KMP_DEBUG_ASSERT(gtid >= 0);
3216   return __kmp_threads[gtid];
3217 }
3218 
3219 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3220   KMP_DEBUG_ASSERT(gtid >= 0);
3221   return __kmp_threads[gtid]->th.th_team;
3222 }
3223 
3224 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3225   if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3226     KMP_FATAL(ThreadIdentInvalid);
3227 }
3228 
3229 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3230 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3231 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3232 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3233 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3234 #endif
3235 
3236 /* ------------------------------------------------------------------------- */
3237 
3238 extern kmp_global_t __kmp_global; /* global status */
3239 
3240 extern kmp_info_t __kmp_monitor;
3241 // For Debugging Support Library
3242 extern std::atomic<kmp_int32> __kmp_team_counter;
3243 // For Debugging Support Library
3244 extern std::atomic<kmp_int32> __kmp_task_counter;
3245 
3246 #if USE_DEBUGGER
3247 #define _KMP_GEN_ID(counter)                                                   \
3248   (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3249 #else
3250 #define _KMP_GEN_ID(counter) (~0)
3251 #endif /* USE_DEBUGGER */
3252 
3253 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3254 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3255 
3256 /* ------------------------------------------------------------------------ */
3257 
3258 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3259                                          size_t size, char const *format, ...);
3260 
3261 extern void __kmp_serial_initialize(void);
3262 extern void __kmp_middle_initialize(void);
3263 extern void __kmp_parallel_initialize(void);
3264 
3265 extern void __kmp_internal_begin(void);
3266 extern void __kmp_internal_end_library(int gtid);
3267 extern void __kmp_internal_end_thread(int gtid);
3268 extern void __kmp_internal_end_atexit(void);
3269 extern void __kmp_internal_end_dtor(void);
3270 extern void __kmp_internal_end_dest(void *);
3271 
3272 extern int __kmp_register_root(int initial_thread);
3273 extern void __kmp_unregister_root(int gtid);
3274 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3275 
3276 extern int __kmp_ignore_mppbeg(void);
3277 extern int __kmp_ignore_mppend(void);
3278 
3279 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3280 extern void __kmp_exit_single(int gtid);
3281 
3282 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3283 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3284 
3285 #ifdef USE_LOAD_BALANCE
3286 extern int __kmp_get_load_balance(int);
3287 #endif
3288 
3289 extern int __kmp_get_global_thread_id(void);
3290 extern int __kmp_get_global_thread_id_reg(void);
3291 extern void __kmp_exit_thread(int exit_status);
3292 extern void __kmp_abort(char const *format, ...);
3293 extern void __kmp_abort_thread(void);
3294 KMP_NORETURN extern void __kmp_abort_process(void);
3295 extern void __kmp_warn(char const *format, ...);
3296 
3297 extern void __kmp_set_num_threads(int new_nth, int gtid);
3298 
3299 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3300 // registered.
3301 static inline kmp_info_t *__kmp_entry_thread() {
3302   int gtid = __kmp_entry_gtid();
3303 
3304   return __kmp_threads[gtid];
3305 }
3306 
3307 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3308 extern int __kmp_get_max_active_levels(int gtid);
3309 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3310 extern int __kmp_get_team_size(int gtid, int level);
3311 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3312 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3313 
3314 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3315 extern void __kmp_init_random(kmp_info_t *thread);
3316 
3317 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3318 extern void __kmp_adjust_num_threads(int new_nproc);
3319 extern void __kmp_check_stksize(size_t *val);
3320 
3321 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3322 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3323 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3324 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3325 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3326 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3327 
3328 #if USE_FAST_MEMORY
3329 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3330                                   size_t size KMP_SRC_LOC_DECL);
3331 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3332 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3333 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3334 #define __kmp_fast_allocate(this_thr, size)                                    \
3335   ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3336 #define __kmp_fast_free(this_thr, ptr)                                         \
3337   ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3338 #endif
3339 
3340 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3341 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3342                                   size_t elsize KMP_SRC_LOC_DECL);
3343 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3344                                    size_t size KMP_SRC_LOC_DECL);
3345 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3346 #define __kmp_thread_malloc(th, size)                                          \
3347   ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3348 #define __kmp_thread_calloc(th, nelem, elsize)                                 \
3349   ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3350 #define __kmp_thread_realloc(th, ptr, size)                                    \
3351   ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3352 #define __kmp_thread_free(th, ptr)                                             \
3353   ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3354 
3355 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3356 #define KMP_INTERNAL_FREE(p) free(p)
3357 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3358 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3359 
3360 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3361 
3362 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3363                                  kmp_proc_bind_t proc_bind);
3364 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3365                                  int num_threads);
3366 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3367                                     int num_teams_ub, int num_threads);
3368 
3369 extern void __kmp_yield();
3370 
3371 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3372                                    enum sched_type schedule, kmp_int32 lb,
3373                                    kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3374 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3375                                     enum sched_type schedule, kmp_uint32 lb,
3376                                     kmp_uint32 ub, kmp_int32 st,
3377                                     kmp_int32 chunk);
3378 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3379                                    enum sched_type schedule, kmp_int64 lb,
3380                                    kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3381 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3382                                     enum sched_type schedule, kmp_uint64 lb,
3383                                     kmp_uint64 ub, kmp_int64 st,
3384                                     kmp_int64 chunk);
3385 
3386 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3387                                   kmp_int32 *p_last, kmp_int32 *p_lb,
3388                                   kmp_int32 *p_ub, kmp_int32 *p_st);
3389 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3390                                    kmp_int32 *p_last, kmp_uint32 *p_lb,
3391                                    kmp_uint32 *p_ub, kmp_int32 *p_st);
3392 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3393                                   kmp_int32 *p_last, kmp_int64 *p_lb,
3394                                   kmp_int64 *p_ub, kmp_int64 *p_st);
3395 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3396                                    kmp_int32 *p_last, kmp_uint64 *p_lb,
3397                                    kmp_uint64 *p_ub, kmp_int64 *p_st);
3398 
3399 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3400 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3401 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3402 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3403 
3404 #ifdef KMP_GOMP_COMPAT
3405 
3406 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3407                                       enum sched_type schedule, kmp_int32 lb,
3408                                       kmp_int32 ub, kmp_int32 st,
3409                                       kmp_int32 chunk, int push_ws);
3410 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3411                                        enum sched_type schedule, kmp_uint32 lb,
3412                                        kmp_uint32 ub, kmp_int32 st,
3413                                        kmp_int32 chunk, int push_ws);
3414 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3415                                       enum sched_type schedule, kmp_int64 lb,
3416                                       kmp_int64 ub, kmp_int64 st,
3417                                       kmp_int64 chunk, int push_ws);
3418 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3419                                        enum sched_type schedule, kmp_uint64 lb,
3420                                        kmp_uint64 ub, kmp_int64 st,
3421                                        kmp_int64 chunk, int push_ws);
3422 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3423 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3424 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3425 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3426 
3427 #endif /* KMP_GOMP_COMPAT */
3428 
3429 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3430 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3431 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3432 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3433 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3434 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3435                                kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3436                                void *obj);
3437 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3438                              kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3439 
3440 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3441                           int final_spin
3442 #if USE_ITT_BUILD
3443                           ,
3444                           void *itt_sync_obj
3445 #endif
3446                           );
3447 extern void __kmp_release_64(kmp_flag_64<> *flag);
3448 
3449 extern void __kmp_infinite_loop(void);
3450 
3451 extern void __kmp_cleanup(void);
3452 
3453 #if KMP_HANDLE_SIGNALS
3454 extern int __kmp_handle_signals;
3455 extern void __kmp_install_signals(int parallel_init);
3456 extern void __kmp_remove_signals(void);
3457 #endif
3458 
3459 extern void __kmp_clear_system_time(void);
3460 extern void __kmp_read_system_time(double *delta);
3461 
3462 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3463 
3464 extern void __kmp_expand_host_name(char *buffer, size_t size);
3465 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3466 
3467 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3468 extern void
3469 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3470 #endif
3471 
3472 extern void
3473 __kmp_runtime_initialize(void); /* machine specific initialization */
3474 extern void __kmp_runtime_destroy(void);
3475 
3476 #if KMP_AFFINITY_SUPPORTED
3477 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3478                                        kmp_affin_mask_t *mask);
3479 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3480                                                   kmp_affin_mask_t *mask);
3481 extern void __kmp_affinity_initialize(void);
3482 extern void __kmp_affinity_uninitialize(void);
3483 extern void __kmp_affinity_set_init_mask(
3484     int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3485 extern void __kmp_affinity_set_place(int gtid);
3486 extern void __kmp_affinity_determine_capable(const char *env_var);
3487 extern int __kmp_aux_set_affinity(void **mask);
3488 extern int __kmp_aux_get_affinity(void **mask);
3489 extern int __kmp_aux_get_affinity_max_proc();
3490 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3491 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3492 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3493 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3494 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3495 extern int kmp_set_thread_affinity_mask_initial(void);
3496 #endif
3497 #endif /* KMP_AFFINITY_SUPPORTED */
3498 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3499 // format string is for affinity, so platforms that do not support
3500 // affinity can still use the other fields, e.g., %n for num_threads
3501 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3502                                          kmp_str_buf_t *buffer);
3503 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3504 
3505 extern void __kmp_cleanup_hierarchy();
3506 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3507 
3508 #if KMP_USE_FUTEX
3509 
3510 extern int __kmp_futex_determine_capable(void);
3511 
3512 #endif // KMP_USE_FUTEX
3513 
3514 extern void __kmp_gtid_set_specific(int gtid);
3515 extern int __kmp_gtid_get_specific(void);
3516 
3517 extern double __kmp_read_cpu_time(void);
3518 
3519 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3520 
3521 #if KMP_USE_MONITOR
3522 extern void __kmp_create_monitor(kmp_info_t *th);
3523 #endif
3524 
3525 extern void *__kmp_launch_thread(kmp_info_t *thr);
3526 
3527 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3528 
3529 #if KMP_OS_WINDOWS
3530 extern int __kmp_still_running(kmp_info_t *th);
3531 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3532 extern void __kmp_free_handle(kmp_thread_t tHandle);
3533 #endif
3534 
3535 #if KMP_USE_MONITOR
3536 extern void __kmp_reap_monitor(kmp_info_t *th);
3537 #endif
3538 extern void __kmp_reap_worker(kmp_info_t *th);
3539 extern void __kmp_terminate_thread(int gtid);
3540 
3541 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3542 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3543 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3544 
3545 extern void __kmp_elapsed(double *);
3546 extern void __kmp_elapsed_tick(double *);
3547 
3548 extern void __kmp_enable(int old_state);
3549 extern void __kmp_disable(int *old_state);
3550 
3551 extern void __kmp_thread_sleep(int millis);
3552 
3553 extern void __kmp_common_initialize(void);
3554 extern void __kmp_common_destroy(void);
3555 extern void __kmp_common_destroy_gtid(int gtid);
3556 
3557 #if KMP_OS_UNIX
3558 extern void __kmp_register_atfork(void);
3559 #endif
3560 extern void __kmp_suspend_initialize(void);
3561 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3562 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3563 
3564 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3565                                          int tid);
3566 extern kmp_team_t *
3567 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3568 #if OMPT_SUPPORT
3569                     ompt_data_t ompt_parallel_data,
3570 #endif
3571                     kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3572                     int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3573 extern void __kmp_free_thread(kmp_info_t *);
3574 extern void __kmp_free_team(kmp_root_t *,
3575                             kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3576 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3577 
3578 /* ------------------------------------------------------------------------ */
3579 
3580 extern void __kmp_initialize_bget(kmp_info_t *th);
3581 extern void __kmp_finalize_bget(kmp_info_t *th);
3582 
3583 KMP_EXPORT void *kmpc_malloc(size_t size);
3584 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3585 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3586 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3587 KMP_EXPORT void kmpc_free(void *ptr);
3588 
3589 /* declarations for internal use */
3590 
3591 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3592                          size_t reduce_size, void *reduce_data,
3593                          void (*reduce)(void *, void *));
3594 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3595 extern int __kmp_barrier_gomp_cancel(int gtid);
3596 
3597 /*!
3598  * Tell the fork call which compiler generated the fork call, and therefore how
3599  * to deal with the call.
3600  */
3601 enum fork_context_e {
3602   fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3603                        microtask internally. */
3604   fork_context_intel, /**< Called from Intel generated code.  */
3605   fork_context_last
3606 };
3607 extern int __kmp_fork_call(ident_t *loc, int gtid,
3608                            enum fork_context_e fork_context, kmp_int32 argc,
3609                            microtask_t microtask, launch_t invoker,
3610                            kmp_va_list ap);
3611 
3612 extern void __kmp_join_call(ident_t *loc, int gtid
3613 #if OMPT_SUPPORT
3614                             ,
3615                             enum fork_context_e fork_context
3616 #endif
3617                             ,
3618                             int exit_teams = 0);
3619 
3620 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3621 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3622 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3623 extern int __kmp_invoke_task_func(int gtid);
3624 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3625                                           kmp_info_t *this_thr,
3626                                           kmp_team_t *team);
3627 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3628                                          kmp_info_t *this_thr,
3629                                          kmp_team_t *team);
3630 
3631 // should never have been exported
3632 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3633 extern int __kmp_invoke_teams_master(int gtid);
3634 extern void __kmp_teams_master(int gtid);
3635 extern int __kmp_aux_get_team_num();
3636 extern int __kmp_aux_get_num_teams();
3637 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3638 extern void __kmp_user_set_library(enum library_type arg);
3639 extern void __kmp_aux_set_library(enum library_type arg);
3640 extern void __kmp_aux_set_stacksize(size_t arg);
3641 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3642 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3643 
3644 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3645 void kmpc_set_blocktime(int arg);
3646 void ompc_set_nested(int flag);
3647 void ompc_set_dynamic(int flag);
3648 void ompc_set_num_threads(int arg);
3649 
3650 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3651                                               kmp_team_t *team, int tid);
3652 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3653 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3654                                     kmp_tasking_flags_t *flags,
3655                                     size_t sizeof_kmp_task_t,
3656                                     size_t sizeof_shareds,
3657                                     kmp_routine_entry_t task_entry);
3658 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3659                                      kmp_team_t *team, int tid,
3660                                      int set_curr_task);
3661 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3662 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3663 
3664 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3665                                                        int gtid,
3666                                                        kmp_task_t *task);
3667 extern void __kmp_fulfill_event(kmp_event_t *event);
3668 
3669 extern void __kmp_free_task_team(kmp_info_t *thread,
3670                                  kmp_task_team_t *task_team);
3671 extern void __kmp_reap_task_teams(void);
3672 extern void __kmp_wait_to_unref_task_teams(void);
3673 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3674                                   int always);
3675 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3676 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3677 #if USE_ITT_BUILD
3678                                  ,
3679                                  void *itt_sync_obj
3680 #endif /* USE_ITT_BUILD */
3681                                  ,
3682                                  int wait = 1);
3683 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3684                                   int gtid);
3685 
3686 extern int __kmp_is_address_mapped(void *addr);
3687 extern kmp_uint64 __kmp_hardware_timestamp(void);
3688 
3689 #if KMP_OS_UNIX
3690 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3691 #endif
3692 
3693 /* ------------------------------------------------------------------------ */
3694 //
3695 // Assembly routines that have no compiler intrinsic replacement
3696 //
3697 
3698 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3699                                   void *argv[]
3700 #if OMPT_SUPPORT
3701                                   ,
3702                                   void **exit_frame_ptr
3703 #endif
3704                                   );
3705 
3706 /* ------------------------------------------------------------------------ */
3707 
3708 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3709 KMP_EXPORT void __kmpc_end(ident_t *);
3710 
3711 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3712                                                   kmpc_ctor_vec ctor,
3713                                                   kmpc_cctor_vec cctor,
3714                                                   kmpc_dtor_vec dtor,
3715                                                   size_t vector_length);
3716 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3717                                               kmpc_ctor ctor, kmpc_cctor cctor,
3718                                               kmpc_dtor dtor);
3719 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3720                                       void *data, size_t size);
3721 
3722 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3723 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3724 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3725 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3726 
3727 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3728 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3729                                  kmpc_micro microtask, ...);
3730 
3731 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3732 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3733 
3734 KMP_EXPORT void __kmpc_flush(ident_t *);
3735 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3736 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3737 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3738 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3739 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3740 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3741                                 kmp_critical_name *);
3742 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3743                                     kmp_critical_name *);
3744 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3745                                           kmp_critical_name *, uint32_t hint);
3746 
3747 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3748 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3749 
3750 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3751                                                   kmp_int32 global_tid);
3752 
3753 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3754 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3755 
3756 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3757                                      kmp_int32 schedtype, kmp_int32 *plastiter,
3758                                      kmp_int *plower, kmp_int *pupper,
3759                                      kmp_int *pstride, kmp_int incr,
3760                                      kmp_int chunk);
3761 
3762 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3763 
3764 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3765                                    size_t cpy_size, void *cpy_data,
3766                                    void (*cpy_func)(void *, void *),
3767                                    kmp_int32 didit);
3768 
3769 extern void KMPC_SET_NUM_THREADS(int arg);
3770 extern void KMPC_SET_DYNAMIC(int flag);
3771 extern void KMPC_SET_NESTED(int flag);
3772 
3773 /* OMP 3.0 tasking interface routines */
3774 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3775                                      kmp_task_t *new_task);
3776 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3777                                              kmp_int32 flags,
3778                                              size_t sizeof_kmp_task_t,
3779                                              size_t sizeof_shareds,
3780                                              kmp_routine_entry_t task_entry);
3781 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3782                                                     kmp_int32 flags,
3783                                                     size_t sizeof_kmp_task_t,
3784                                                     size_t sizeof_shareds,
3785                                                     kmp_routine_entry_t task_entry,
3786                                                     kmp_int64 device_id);
3787 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3788                                           kmp_task_t *task);
3789 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3790                                              kmp_task_t *task);
3791 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3792                                            kmp_task_t *new_task);
3793 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3794 
3795 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3796                                           int end_part);
3797 
3798 #if TASK_UNUSED
3799 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3800 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3801                               kmp_task_t *task);
3802 #endif // TASK_UNUSED
3803 
3804 /* ------------------------------------------------------------------------ */
3805 
3806 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3807 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3808 
3809 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3810     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3811     kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3812     kmp_depend_info_t *noalias_dep_list);
3813 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3814                                      kmp_int32 ndeps,
3815                                      kmp_depend_info_t *dep_list,
3816                                      kmp_int32 ndeps_noalias,
3817                                      kmp_depend_info_t *noalias_dep_list);
3818 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3819                                 bool serialize_immediate);
3820 
3821 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3822                                    kmp_int32 cncl_kind);
3823 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3824                                               kmp_int32 cncl_kind);
3825 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3826 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3827 
3828 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3829 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3830 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3831                                 kmp_int32 if_val, kmp_uint64 *lb,
3832                                 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3833                                 kmp_int32 sched, kmp_uint64 grainsize,
3834                                 void *task_dup);
3835 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3836                                   kmp_task_t *task, kmp_int32 if_val,
3837                                   kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3838                                   kmp_int32 nogroup, kmp_int32 sched,
3839                                   kmp_uint64 grainsize, kmp_int32 modifier,
3840                                   void *task_dup);
3841 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3842 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3843 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3844 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3845                                                      int is_ws, int num,
3846                                                      void *data);
3847 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3848                                               int num, void *data);
3849 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3850                                                     int is_ws);
3851 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3852     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3853     kmp_task_affinity_info_t *affin_list);
3854 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
3855 KMP_EXPORT int __kmp_get_max_teams(void);
3856 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
3857 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
3858 
3859 /* Lock interface routines (fast versions with gtid passed in) */
3860 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3861                                  void **user_lock);
3862 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3863                                       void **user_lock);
3864 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3865                                     void **user_lock);
3866 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3867                                          void **user_lock);
3868 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3869 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3870                                      void **user_lock);
3871 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3872                                   void **user_lock);
3873 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3874                                        void **user_lock);
3875 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3876 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3877                                      void **user_lock);
3878 
3879 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3880                                            void **user_lock, uintptr_t hint);
3881 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3882                                                 void **user_lock,
3883                                                 uintptr_t hint);
3884 
3885 /* Interface to fast scalable reduce methods routines */
3886 
3887 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3888     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3889     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3890     kmp_critical_name *lck);
3891 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3892                                          kmp_critical_name *lck);
3893 KMP_EXPORT kmp_int32 __kmpc_reduce(
3894     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3895     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3896     kmp_critical_name *lck);
3897 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3898                                   kmp_critical_name *lck);
3899 
3900 /* Internal fast reduction routines */
3901 
3902 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
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 
3907 // this function is for testing set/get/determine reduce method
3908 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3909 
3910 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3911 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3912 
3913 // C++ port
3914 // missing 'extern "C"' declarations
3915 
3916 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3917 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3918 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3919                                         kmp_int32 num_threads);
3920 
3921 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3922                                       int proc_bind);
3923 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3924                                       kmp_int32 num_teams,
3925                                       kmp_int32 num_threads);
3926 /* Function for OpenMP 5.1 num_teams clause */
3927 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
3928                                          kmp_int32 num_teams_lb,
3929                                          kmp_int32 num_teams_ub,
3930                                          kmp_int32 num_threads);
3931 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3932                                   kmpc_micro microtask, ...);
3933 struct kmp_dim { // loop bounds info casted to kmp_int64
3934   kmp_int64 lo; // lower
3935   kmp_int64 up; // upper
3936   kmp_int64 st; // stride
3937 };
3938 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3939                                      kmp_int32 num_dims,
3940                                      const struct kmp_dim *dims);
3941 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3942                                      const kmp_int64 *vec);
3943 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3944                                      const kmp_int64 *vec);
3945 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3946 
3947 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3948                                              void *data, size_t size,
3949                                              void ***cache);
3950 
3951 // Symbols for MS mutual detection.
3952 extern int _You_must_link_with_exactly_one_OpenMP_library;
3953 extern int _You_must_link_with_Intel_OpenMP_library;
3954 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3955 extern int _You_must_link_with_Microsoft_OpenMP_library;
3956 #endif
3957 
3958 // The routines below are not exported.
3959 // Consider making them 'static' in corresponding source files.
3960 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3961                                            void *data_addr, size_t pc_size);
3962 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3963                                                 void *data_addr,
3964                                                 size_t pc_size);
3965 void __kmp_threadprivate_resize_cache(int newCapacity);
3966 void __kmp_cleanup_threadprivate_caches();
3967 
3968 // ompc_, kmpc_ entries moved from omp.h.
3969 #if KMP_OS_WINDOWS
3970 #define KMPC_CONVENTION __cdecl
3971 #else
3972 #define KMPC_CONVENTION
3973 #endif
3974 
3975 #ifndef __OMP_H
3976 typedef enum omp_sched_t {
3977   omp_sched_static = 1,
3978   omp_sched_dynamic = 2,
3979   omp_sched_guided = 3,
3980   omp_sched_auto = 4
3981 } omp_sched_t;
3982 typedef void *kmp_affinity_mask_t;
3983 #endif
3984 
3985 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3986 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3987 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3988 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3989 KMP_EXPORT int KMPC_CONVENTION
3990 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3991 KMP_EXPORT int KMPC_CONVENTION
3992 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3993 KMP_EXPORT int KMPC_CONVENTION
3994 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3995 
3996 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3997 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3998 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3999 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4000 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4001 
4002 enum kmp_target_offload_kind {
4003   tgt_disabled = 0,
4004   tgt_default = 1,
4005   tgt_mandatory = 2
4006 };
4007 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4008 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4009 extern kmp_target_offload_kind_t __kmp_target_offload;
4010 extern int __kmpc_get_target_offload();
4011 
4012 // Constants used in libomptarget
4013 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4014 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4015 
4016 // OMP Pause Resource
4017 
4018 // The following enum is used both to set the status in __kmp_pause_status, and
4019 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4020 typedef enum kmp_pause_status_t {
4021   kmp_not_paused = 0, // status is not paused, or, requesting resume
4022   kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4023   kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4024 } kmp_pause_status_t;
4025 
4026 // This stores the pause state of the runtime
4027 extern kmp_pause_status_t __kmp_pause_status;
4028 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4029 extern int __kmp_pause_resource(kmp_pause_status_t level);
4030 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4031 extern void __kmp_resume_if_soft_paused();
4032 // Hard resume simply resets the status to not paused. Library will appear to
4033 // be uninitialized after hard pause. Let OMP constructs trigger required
4034 // initializations.
4035 static inline void __kmp_resume_if_hard_paused() {
4036   if (__kmp_pause_status == kmp_hard_paused) {
4037     __kmp_pause_status = kmp_not_paused;
4038   }
4039 }
4040 
4041 extern void __kmp_omp_display_env(int verbose);
4042 
4043 // 1: it is initializing hidden helper team
4044 extern volatile int __kmp_init_hidden_helper;
4045 // 1: the hidden helper team is done
4046 extern volatile int __kmp_hidden_helper_team_done;
4047 // 1: enable hidden helper task
4048 extern kmp_int32 __kmp_enable_hidden_helper;
4049 // Main thread of hidden helper team
4050 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4051 // Descriptors for the hidden helper threads
4052 extern kmp_info_t **__kmp_hidden_helper_threads;
4053 // Number of hidden helper threads
4054 extern kmp_int32 __kmp_hidden_helper_threads_num;
4055 // Number of hidden helper tasks that have not been executed yet
4056 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4057 
4058 extern void __kmp_hidden_helper_initialize();
4059 extern void __kmp_hidden_helper_threads_initz_routine();
4060 extern void __kmp_do_initialize_hidden_helper_threads();
4061 extern void __kmp_hidden_helper_threads_initz_wait();
4062 extern void __kmp_hidden_helper_initz_release();
4063 extern void __kmp_hidden_helper_threads_deinitz_wait();
4064 extern void __kmp_hidden_helper_threads_deinitz_release();
4065 extern void __kmp_hidden_helper_main_thread_wait();
4066 extern void __kmp_hidden_helper_worker_thread_wait();
4067 extern void __kmp_hidden_helper_worker_thread_signal();
4068 extern void __kmp_hidden_helper_main_thread_release();
4069 
4070 // Check whether a given thread is a hidden helper thread
4071 #define KMP_HIDDEN_HELPER_THREAD(gtid)                                         \
4072   ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4073 
4074 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)                                  \
4075   ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4076 
4077 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4078 // main thread, is skipped.
4079 #define KMP_GTID_TO_SHADOW_GTID(gtid)                                          \
4080   ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4081 
4082 #ifdef __cplusplus
4083 }
4084 #endif
4085 
4086 template <bool C, bool S>
4087 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4088 template <bool C, bool S>
4089 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4090 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4091 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4092 template <bool C, bool S>
4093 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4094 template <bool C, bool S>
4095 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4096 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4097 #endif
4098 template <bool C, bool S>
4099 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4100 template <bool C, bool S>
4101 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4102 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4103 
4104 template <bool C, bool S>
4105 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4106                            kmp_flag_32<C, S> *flag, int final_spin,
4107                            int *thread_finished,
4108 #if USE_ITT_BUILD
4109                            void *itt_sync_obj,
4110 #endif /* USE_ITT_BUILD */
4111                            kmp_int32 is_constrained);
4112 template <bool C, bool S>
4113 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4114                            kmp_flag_64<C, S> *flag, int final_spin,
4115                            int *thread_finished,
4116 #if USE_ITT_BUILD
4117                            void *itt_sync_obj,
4118 #endif /* USE_ITT_BUILD */
4119                            kmp_int32 is_constrained);
4120 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4121                                kmp_flag_oncore *flag, int final_spin,
4122                                int *thread_finished,
4123 #if USE_ITT_BUILD
4124                                void *itt_sync_obj,
4125 #endif /* USE_ITT_BUILD */
4126                                kmp_int32 is_constrained);
4127 
4128 /// This class safely opens and closes a C-style FILE* object using RAII
4129 /// semantics. There are also methods which allow using stdout or stderr as
4130 /// the underlying FILE* object. With the implicit conversion operator to
4131 /// FILE*, an object with this type can be used in any function which takes
4132 /// a FILE* object e.g., fprintf().
4133 /// No close method is needed at use sites.
4134 class kmp_safe_raii_file_t {
4135   FILE *f;
4136 
4137   void close() {
4138     if (f && f != stdout && f != stderr) {
4139       fclose(f);
4140       f = nullptr;
4141     }
4142   }
4143 
4144 public:
4145   kmp_safe_raii_file_t() : f(nullptr) {}
4146   kmp_safe_raii_file_t(const char *filename, const char *mode,
4147                        const char *env_var = nullptr)
4148       : f(nullptr) {
4149     open(filename, mode, env_var);
4150   }
4151   ~kmp_safe_raii_file_t() { close(); }
4152 
4153   /// Open filename using mode. This is automatically closed in the destructor.
4154   /// The env_var parameter indicates the environment variable the filename
4155   /// came from if != nullptr.
4156   void open(const char *filename, const char *mode,
4157             const char *env_var = nullptr) {
4158     KMP_ASSERT(!f);
4159     f = fopen(filename, mode);
4160     if (!f) {
4161       int code = errno;
4162       if (env_var) {
4163         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4164                     KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4165       } else {
4166         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4167                     __kmp_msg_null);
4168       }
4169     }
4170   }
4171   /// Set the FILE* object to stdout and output there
4172   /// No open call should happen before this call.
4173   void set_stdout() {
4174     KMP_ASSERT(!f);
4175     f = stdout;
4176   }
4177   /// Set the FILE* object to stderr and output there
4178   /// No open call should happen before this call.
4179   void set_stderr() {
4180     KMP_ASSERT(!f);
4181     f = stderr;
4182   }
4183   operator bool() { return bool(f); }
4184   operator FILE *() { return f; }
4185 };
4186 
4187 template <typename SourceType, typename TargetType,
4188           bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4189           bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4190           bool isSourceSigned = std::is_signed<SourceType>::value,
4191           bool isTargetSigned = std::is_signed<TargetType>::value>
4192 struct kmp_convert {};
4193 
4194 // Both types are signed; Source smaller
4195 template <typename SourceType, typename TargetType>
4196 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4197   static TargetType to(SourceType src) { return (TargetType)src; }
4198 };
4199 // Source equal
4200 template <typename SourceType, typename TargetType>
4201 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4202   static TargetType to(SourceType src) { return src; }
4203 };
4204 // Source bigger
4205 template <typename SourceType, typename TargetType>
4206 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4207   static TargetType to(SourceType src) {
4208     KMP_ASSERT(src <= static_cast<SourceType>(
4209                           (std::numeric_limits<TargetType>::max)()));
4210     KMP_ASSERT(src >= static_cast<SourceType>(
4211                           (std::numeric_limits<TargetType>::min)()));
4212     return (TargetType)src;
4213   }
4214 };
4215 
4216 // Source signed, Target unsigned
4217 // Source smaller
4218 template <typename SourceType, typename TargetType>
4219 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4220   static TargetType to(SourceType src) {
4221     KMP_ASSERT(src >= 0);
4222     return (TargetType)src;
4223   }
4224 };
4225 // Source equal
4226 template <typename SourceType, typename TargetType>
4227 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4228   static TargetType to(SourceType src) {
4229     KMP_ASSERT(src >= 0);
4230     return (TargetType)src;
4231   }
4232 };
4233 // Source bigger
4234 template <typename SourceType, typename TargetType>
4235 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4236   static TargetType to(SourceType src) {
4237     KMP_ASSERT(src >= 0);
4238     KMP_ASSERT(src <= static_cast<SourceType>(
4239                           (std::numeric_limits<TargetType>::max)()));
4240     return (TargetType)src;
4241   }
4242 };
4243 
4244 // Source unsigned, Target signed
4245 // Source smaller
4246 template <typename SourceType, typename TargetType>
4247 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4248   static TargetType to(SourceType src) { return (TargetType)src; }
4249 };
4250 // Source equal
4251 template <typename SourceType, typename TargetType>
4252 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4253   static TargetType to(SourceType src) {
4254     KMP_ASSERT(src <= static_cast<SourceType>(
4255                           (std::numeric_limits<TargetType>::max)()));
4256     return (TargetType)src;
4257   }
4258 };
4259 // Source bigger
4260 template <typename SourceType, typename TargetType>
4261 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4262   static TargetType to(SourceType src) {
4263     KMP_ASSERT(src <= static_cast<SourceType>(
4264                           (std::numeric_limits<TargetType>::max)()));
4265     return (TargetType)src;
4266   }
4267 };
4268 
4269 // Source unsigned, Target unsigned
4270 // Source smaller
4271 template <typename SourceType, typename TargetType>
4272 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4273   static TargetType to(SourceType src) { return (TargetType)src; }
4274 };
4275 // Source equal
4276 template <typename SourceType, typename TargetType>
4277 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4278   static TargetType to(SourceType src) { return src; }
4279 };
4280 // Source bigger
4281 template <typename SourceType, typename TargetType>
4282 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4283   static TargetType to(SourceType src) {
4284     KMP_ASSERT(src <= static_cast<SourceType>(
4285                           (std::numeric_limits<TargetType>::max)()));
4286     return (TargetType)src;
4287   }
4288 };
4289 
4290 template <typename T1, typename T2>
4291 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4292   *dest = kmp_convert<T1, T2>::to(src);
4293 }
4294 
4295 #endif /* KMP_H */
4296