xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp.h (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
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 #endif
104 
105 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
106 #include <xmmintrin.h>
107 #endif
108 
109 // The below has to be defined before including "kmp_barrier.h".
110 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
111 #define KMP_INTERNAL_FREE(p) free(p)
112 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
113 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
114 
115 #include "kmp_debug.h"
116 #include "kmp_lock.h"
117 #include "kmp_version.h"
118 #include "kmp_barrier.h"
119 #if USE_DEBUGGER
120 #include "kmp_debugger.h"
121 #endif
122 #include "kmp_i18n.h"
123 
124 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
125 
126 #include "kmp_wrapper_malloc.h"
127 #if KMP_OS_UNIX
128 #include <unistd.h>
129 #if !defined NSIG && defined _NSIG
130 #define NSIG _NSIG
131 #endif
132 #endif
133 
134 #if KMP_OS_LINUX
135 #pragma weak clock_gettime
136 #endif
137 
138 #if OMPT_SUPPORT
139 #include "ompt-internal.h"
140 #endif
141 
142 #if OMPD_SUPPORT
143 #include "ompd-specific.h"
144 #endif
145 
146 #ifndef UNLIKELY
147 #define UNLIKELY(x) (x)
148 #endif
149 
150 // Affinity format function
151 #include "kmp_str.h"
152 
153 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
154 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
155 // free lists of limited size.
156 #ifndef USE_FAST_MEMORY
157 #define USE_FAST_MEMORY 3
158 #endif
159 
160 #ifndef KMP_NESTED_HOT_TEAMS
161 #define KMP_NESTED_HOT_TEAMS 0
162 #define USE_NESTED_HOT_ARG(x)
163 #else
164 #if KMP_NESTED_HOT_TEAMS
165 #define USE_NESTED_HOT_ARG(x) , x
166 #else
167 #define USE_NESTED_HOT_ARG(x)
168 #endif
169 #endif
170 
171 // Assume using BGET compare_exchange instruction instead of lock by default.
172 #ifndef USE_CMP_XCHG_FOR_BGET
173 #define USE_CMP_XCHG_FOR_BGET 1
174 #endif
175 
176 // Test to see if queuing lock is better than bootstrap lock for bget
177 // #ifndef USE_QUEUING_LOCK_FOR_BGET
178 // #define USE_QUEUING_LOCK_FOR_BGET
179 // #endif
180 
181 #define KMP_NSEC_PER_SEC 1000000000L
182 #define KMP_USEC_PER_SEC 1000000L
183 
184 /*!
185 @ingroup BASIC_TYPES
186 @{
187 */
188 
189 /*!
190 Values for bit flags used in the ident_t to describe the fields.
191 */
192 enum {
193   /*! Use trampoline for internal microtasks */
194   KMP_IDENT_IMB = 0x01,
195   /*! Use c-style ident structure */
196   KMP_IDENT_KMPC = 0x02,
197   /* 0x04 is no longer used */
198   /*! Entry point generated by auto-parallelization */
199   KMP_IDENT_AUTOPAR = 0x08,
200   /*! Compiler generates atomic reduction option for kmpc_reduce* */
201   KMP_IDENT_ATOMIC_REDUCE = 0x10,
202   /*! To mark a 'barrier' directive in user code */
203   KMP_IDENT_BARRIER_EXPL = 0x20,
204   /*! To Mark implicit barriers. */
205   KMP_IDENT_BARRIER_IMPL = 0x0040,
206   KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
207   KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
208   KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
209 
210   KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
211   KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
212 
213   /*! To mark a static loop in OMPT callbacks */
214   KMP_IDENT_WORK_LOOP = 0x200,
215   /*! To mark a sections directive in OMPT callbacks */
216   KMP_IDENT_WORK_SECTIONS = 0x400,
217   /*! To mark a distribute construct in OMPT callbacks */
218   KMP_IDENT_WORK_DISTRIBUTE = 0x800,
219   /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and
220       not currently used. If one day we need more bits, then we can use
221       an invalid combination of hints to mean that another, larger field
222       should be used in a different flag. */
223   KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000,
224   KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
225   KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
226   KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
227   KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
228   KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
229 };
230 
231 /*!
232  * The ident structure that describes a source location.
233  */
234 typedef struct ident {
235   kmp_int32 reserved_1; /**<  might be used in Fortran; see above  */
236   kmp_int32 flags; /**<  also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC
237                       identifies this union member  */
238   kmp_int32 reserved_2; /**<  not really used in Fortran any more; see above */
239 #if USE_ITT_BUILD
240 /*  but currently used for storing region-specific ITT */
241 /*  contextual information. */
242 #endif /* USE_ITT_BUILD */
243   kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++  */
244   char const *psource; /**< String describing the source location.
245                        The string is composed of semi-colon separated fields
246                        which describe the source file, the function and a pair
247                        of line numbers that delimit the construct. */
248   // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
249   kmp_int32 get_openmp_version() {
250     return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
251   }
252 } ident_t;
253 /*!
254 @}
255 */
256 
257 // Some forward declarations.
258 typedef union kmp_team kmp_team_t;
259 typedef struct kmp_taskdata kmp_taskdata_t;
260 typedef union kmp_task_team kmp_task_team_t;
261 typedef union kmp_team kmp_team_p;
262 typedef union kmp_info kmp_info_p;
263 typedef union kmp_root kmp_root_p;
264 
265 template <bool C = false, bool S = true> class kmp_flag_32;
266 template <bool C = false, bool S = true> class kmp_flag_64;
267 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
268 class kmp_flag_oncore;
269 
270 #ifdef __cplusplus
271 extern "C" {
272 #endif
273 
274 /* ------------------------------------------------------------------------ */
275 
276 /* Pack two 32-bit signed integers into a 64-bit signed integer */
277 /* ToDo: Fix word ordering for big-endian machines. */
278 #define KMP_PACK_64(HIGH_32, LOW_32)                                           \
279   ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
280 
281 // Generic string manipulation macros. Assume that _x is of type char *
282 #define SKIP_WS(_x)                                                            \
283   {                                                                            \
284     while (*(_x) == ' ' || *(_x) == '\t')                                      \
285       (_x)++;                                                                  \
286   }
287 #define SKIP_DIGITS(_x)                                                        \
288   {                                                                            \
289     while (*(_x) >= '0' && *(_x) <= '9')                                       \
290       (_x)++;                                                                  \
291   }
292 #define SKIP_TOKEN(_x)                                                         \
293   {                                                                            \
294     while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
295            (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_')                     \
296       (_x)++;                                                                  \
297   }
298 #define SKIP_TO(_x, _c)                                                        \
299   {                                                                            \
300     while (*(_x) != '\0' && *(_x) != (_c))                                     \
301       (_x)++;                                                                  \
302   }
303 
304 /* ------------------------------------------------------------------------ */
305 
306 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
307 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
308 
309 /* ------------------------------------------------------------------------ */
310 /* Enumeration types */
311 
312 enum kmp_state_timer {
313   ts_stop,
314   ts_start,
315   ts_pause,
316 
317   ts_last_state
318 };
319 
320 enum dynamic_mode {
321   dynamic_default,
322 #ifdef USE_LOAD_BALANCE
323   dynamic_load_balance,
324 #endif /* USE_LOAD_BALANCE */
325   dynamic_random,
326   dynamic_thread_limit,
327   dynamic_max
328 };
329 
330 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
331  * not include it here */
332 #ifndef KMP_SCHED_TYPE_DEFINED
333 #define KMP_SCHED_TYPE_DEFINED
334 typedef enum kmp_sched {
335   kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
336   // Note: need to adjust __kmp_sch_map global array in case enum is changed
337   kmp_sched_static = 1, // mapped to kmp_sch_static_chunked           (33)
338   kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked          (35)
339   kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked           (36)
340   kmp_sched_auto = 4, // mapped to kmp_sch_auto                     (38)
341   kmp_sched_upper_std = 5, // upper bound for standard schedules
342   kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
343   kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
344 #if KMP_STATIC_STEAL_ENABLED
345   kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
346 #endif
347   kmp_sched_upper,
348   kmp_sched_default = kmp_sched_static, // default scheduling
349   kmp_sched_monotonic = 0x80000000
350 } kmp_sched_t;
351 #endif
352 
353 /*!
354  @ingroup WORK_SHARING
355  * Describes the loop schedule to be used for a parallel for loop.
356  */
357 enum sched_type : kmp_int32 {
358   kmp_sch_lower = 32, /**< lower bound for unordered values */
359   kmp_sch_static_chunked = 33,
360   kmp_sch_static = 34, /**< static unspecialized */
361   kmp_sch_dynamic_chunked = 35,
362   kmp_sch_guided_chunked = 36, /**< guided unspecialized */
363   kmp_sch_runtime = 37,
364   kmp_sch_auto = 38, /**< auto */
365   kmp_sch_trapezoidal = 39,
366 
367   /* accessible only through KMP_SCHEDULE environment variable */
368   kmp_sch_static_greedy = 40,
369   kmp_sch_static_balanced = 41,
370   /* accessible only through KMP_SCHEDULE environment variable */
371   kmp_sch_guided_iterative_chunked = 42,
372   kmp_sch_guided_analytical_chunked = 43,
373   /* accessible only through KMP_SCHEDULE environment variable */
374   kmp_sch_static_steal = 44,
375 
376   /* static with chunk adjustment (e.g., simd) */
377   kmp_sch_static_balanced_chunked = 45,
378   kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */
379   kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */
380 
381   /* accessible only through KMP_SCHEDULE environment variable */
382   kmp_sch_upper, /**< upper bound for unordered values */
383 
384   kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */
385   kmp_ord_static_chunked = 65,
386   kmp_ord_static = 66, /**< ordered static unspecialized */
387   kmp_ord_dynamic_chunked = 67,
388   kmp_ord_guided_chunked = 68,
389   kmp_ord_runtime = 69,
390   kmp_ord_auto = 70, /**< ordered auto */
391   kmp_ord_trapezoidal = 71,
392   kmp_ord_upper, /**< upper bound for ordered values */
393 
394   /* Schedules for Distribute construct */
395   kmp_distribute_static_chunked = 91, /**< distribute static chunked */
396   kmp_distribute_static = 92, /**< distribute static unspecialized */
397 
398   /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
399      single iteration/chunk, even if the loop is serialized. For the schedule
400      types listed above, the entire iteration vector is returned if the loop is
401      serialized. This doesn't work for gcc/gcomp sections. */
402   kmp_nm_lower = 160, /**< lower bound for nomerge values */
403 
404   kmp_nm_static_chunked =
405       (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
406   kmp_nm_static = 162, /**< static unspecialized */
407   kmp_nm_dynamic_chunked = 163,
408   kmp_nm_guided_chunked = 164, /**< guided unspecialized */
409   kmp_nm_runtime = 165,
410   kmp_nm_auto = 166, /**< auto */
411   kmp_nm_trapezoidal = 167,
412 
413   /* accessible only through KMP_SCHEDULE environment variable */
414   kmp_nm_static_greedy = 168,
415   kmp_nm_static_balanced = 169,
416   /* accessible only through KMP_SCHEDULE environment variable */
417   kmp_nm_guided_iterative_chunked = 170,
418   kmp_nm_guided_analytical_chunked = 171,
419   kmp_nm_static_steal =
420       172, /* accessible only through OMP_SCHEDULE environment variable */
421 
422   kmp_nm_ord_static_chunked = 193,
423   kmp_nm_ord_static = 194, /**< ordered static unspecialized */
424   kmp_nm_ord_dynamic_chunked = 195,
425   kmp_nm_ord_guided_chunked = 196,
426   kmp_nm_ord_runtime = 197,
427   kmp_nm_ord_auto = 198, /**< auto */
428   kmp_nm_ord_trapezoidal = 199,
429   kmp_nm_upper, /**< upper bound for nomerge values */
430 
431   /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
432      we need to distinguish the three possible cases (no modifier, monotonic
433      modifier, nonmonotonic modifier), we need separate bits for each modifier.
434      The absence of monotonic does not imply nonmonotonic, especially since 4.5
435      says that the behaviour of the "no modifier" case is implementation defined
436      in 4.5, but will become "nonmonotonic" in 5.0.
437 
438      Since we're passing a full 32 bit value, we can use a couple of high bits
439      for these flags; out of paranoia we avoid the sign bit.
440 
441      These modifiers can be or-ed into non-static schedules by the compiler to
442      pass the additional information. They will be stripped early in the
443      processing in __kmp_dispatch_init when setting up schedules, so most of the
444      code won't ever see schedules with these bits set.  */
445   kmp_sch_modifier_monotonic =
446       (1 << 29), /**< Set if the monotonic schedule modifier was present */
447   kmp_sch_modifier_nonmonotonic =
448       (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */
449 
450 #define SCHEDULE_WITHOUT_MODIFIERS(s)                                          \
451   (enum sched_type)(                                                           \
452       (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
453 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
454 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
455 #define SCHEDULE_HAS_NO_MODIFIERS(s)                                           \
456   (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
457 #define SCHEDULE_GET_MODIFIERS(s)                                              \
458   ((enum sched_type)(                                                          \
459       (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
460 #define SCHEDULE_SET_MODIFIERS(s, m)                                           \
461   (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
462 #define SCHEDULE_NONMONOTONIC 0
463 #define SCHEDULE_MONOTONIC 1
464 
465   kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */
466 };
467 
468 // Apply modifiers on internal kind to standard kind
469 static inline void
470 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
471                                enum sched_type internal_kind) {
472   if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
473     *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
474   }
475 }
476 
477 // Apply modifiers on standard kind to internal kind
478 static inline void
479 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
480                                enum sched_type *internal_kind) {
481   if ((int)kind & (int)kmp_sched_monotonic) {
482     *internal_kind = (enum sched_type)((int)*internal_kind |
483                                        (int)kmp_sch_modifier_monotonic);
484   }
485 }
486 
487 // Get standard schedule without modifiers
488 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
489   return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
490 }
491 
492 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
493 typedef union kmp_r_sched {
494   struct {
495     enum sched_type r_sched_type;
496     int chunk;
497   };
498   kmp_int64 sched;
499 } kmp_r_sched_t;
500 
501 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
502 // internal schedule types
503 
504 enum library_type {
505   library_none,
506   library_serial,
507   library_turnaround,
508   library_throughput
509 };
510 
511 #if KMP_OS_LINUX
512 enum clock_function_type {
513   clock_function_gettimeofday,
514   clock_function_clock_gettime
515 };
516 #endif /* KMP_OS_LINUX */
517 
518 #if KMP_MIC_SUPPORTED
519 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
520 #endif
521 
522 /* -- fast reduction stuff ------------------------------------------------ */
523 
524 #undef KMP_FAST_REDUCTION_BARRIER
525 #define KMP_FAST_REDUCTION_BARRIER 1
526 
527 #undef KMP_FAST_REDUCTION_CORE_DUO
528 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
529 #define KMP_FAST_REDUCTION_CORE_DUO 1
530 #endif
531 
532 enum _reduction_method {
533   reduction_method_not_defined = 0,
534   critical_reduce_block = (1 << 8),
535   atomic_reduce_block = (2 << 8),
536   tree_reduce_block = (3 << 8),
537   empty_reduce_block = (4 << 8)
538 };
539 
540 // Description of the packed_reduction_method variable:
541 // The packed_reduction_method variable consists of two enum types variables
542 // that are packed together into 0-th byte and 1-st byte:
543 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
544 // barrier that will be used in fast reduction: bs_plain_barrier or
545 // bs_reduction_barrier
546 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
547 // be used in fast reduction;
548 // Reduction method is of 'enum _reduction_method' type and it's defined the way
549 // so that the bits of 0-th byte are empty, so no need to execute a shift
550 // instruction while packing/unpacking
551 
552 #if KMP_FAST_REDUCTION_BARRIER
553 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
554   ((reduction_method) | (barrier_type))
555 
556 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
557   ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
558 
559 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method)                      \
560   ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
561 #else
562 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type)      \
563   (reduction_method)
564 
565 #define UNPACK_REDUCTION_METHOD(packed_reduction_method)                       \
566   (packed_reduction_method)
567 
568 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
569 #endif
570 
571 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block)  \
572   ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) ==                       \
573    (which_reduction_block))
574 
575 #if KMP_FAST_REDUCTION_BARRIER
576 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER                               \
577   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
578 
579 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER                                   \
580   (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
581 #endif
582 
583 typedef int PACKED_REDUCTION_METHOD_T;
584 
585 /* -- end of fast reduction stuff ----------------------------------------- */
586 
587 #if KMP_OS_WINDOWS
588 #define USE_CBLKDATA
589 #if KMP_MSVC_COMPAT
590 #pragma warning(push)
591 #pragma warning(disable : 271 310)
592 #endif
593 #include <windows.h>
594 #if KMP_MSVC_COMPAT
595 #pragma warning(pop)
596 #endif
597 #endif
598 
599 #if KMP_OS_UNIX
600 #include <dlfcn.h>
601 #include <pthread.h>
602 #endif
603 
604 enum kmp_hw_t : int {
605   KMP_HW_UNKNOWN = -1,
606   KMP_HW_SOCKET = 0,
607   KMP_HW_PROC_GROUP,
608   KMP_HW_NUMA,
609   KMP_HW_DIE,
610   KMP_HW_LLC,
611   KMP_HW_L3,
612   KMP_HW_TILE,
613   KMP_HW_MODULE,
614   KMP_HW_L2,
615   KMP_HW_L1,
616   KMP_HW_CORE,
617   KMP_HW_THREAD,
618   KMP_HW_LAST
619 };
620 
621 typedef enum kmp_hw_core_type_t {
622   KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
623 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
624   KMP_HW_CORE_TYPE_ATOM = 0x20,
625   KMP_HW_CORE_TYPE_CORE = 0x40,
626   KMP_HW_MAX_NUM_CORE_TYPES = 3,
627 #else
628   KMP_HW_MAX_NUM_CORE_TYPES = 1,
629 #endif
630 } kmp_hw_core_type_t;
631 
632 #define KMP_HW_MAX_NUM_CORE_EFFS 8
633 
634 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type)                                   \
635   KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
636 #define KMP_ASSERT_VALID_HW_TYPE(type)                                         \
637   KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
638 
639 #define KMP_FOREACH_HW_TYPE(type)                                              \
640   for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST;                        \
641        type = (kmp_hw_t)((int)type + 1))
642 
643 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
644 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
645 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
646 
647 /* Only Linux* OS and Windows* OS support thread affinity. */
648 #if KMP_AFFINITY_SUPPORTED
649 
650 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
651 #if KMP_OS_WINDOWS
652 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
653 typedef struct GROUP_AFFINITY {
654   KAFFINITY Mask;
655   WORD Group;
656   WORD Reserved[3];
657 } GROUP_AFFINITY;
658 #endif /* _MSC_VER < 1600 */
659 #if KMP_GROUP_AFFINITY
660 extern int __kmp_num_proc_groups;
661 #else
662 static const int __kmp_num_proc_groups = 1;
663 #endif /* KMP_GROUP_AFFINITY */
664 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
665 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
666 
667 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
668 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
669 
670 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
671 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
672 
673 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
674                                              GROUP_AFFINITY *);
675 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
676 #endif /* KMP_OS_WINDOWS */
677 
678 #if KMP_USE_HWLOC
679 extern hwloc_topology_t __kmp_hwloc_topology;
680 extern int __kmp_hwloc_error;
681 #endif
682 
683 extern size_t __kmp_affin_mask_size;
684 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
685 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
686 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
687 #define KMP_CPU_SET_ITERATE(i, mask)                                           \
688   for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
689 #define KMP_CPU_SET(i, mask) (mask)->set(i)
690 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
691 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
692 #define KMP_CPU_ZERO(mask) (mask)->zero()
693 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
694 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
695 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
696 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
697 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
698 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
699 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
700 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
701 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
702 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
703 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
704 #define KMP_CPU_ALLOC_ARRAY(arr, n)                                            \
705   (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
706 #define KMP_CPU_FREE_ARRAY(arr, n)                                             \
707   __kmp_affinity_dispatch->deallocate_mask_array(arr)
708 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
709 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
710 #define __kmp_get_system_affinity(mask, abort_bool)                            \
711   (mask)->get_system_affinity(abort_bool)
712 #define __kmp_set_system_affinity(mask, abort_bool)                            \
713   (mask)->set_system_affinity(abort_bool)
714 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
715 
716 class KMPAffinity {
717 public:
718   class Mask {
719   public:
720     void *operator new(size_t n);
721     void operator delete(void *p);
722     void *operator new[](size_t n);
723     void operator delete[](void *p);
724     virtual ~Mask() {}
725     // Set bit i to 1
726     virtual void set(int i) {}
727     // Return bit i
728     virtual bool is_set(int i) const { return false; }
729     // Set bit i to 0
730     virtual void clear(int i) {}
731     // Zero out entire mask
732     virtual void zero() {}
733     // Copy src into this mask
734     virtual void copy(const Mask *src) {}
735     // this &= rhs
736     virtual void bitwise_and(const Mask *rhs) {}
737     // this |= rhs
738     virtual void bitwise_or(const Mask *rhs) {}
739     // this = ~this
740     virtual void bitwise_not() {}
741     // API for iterating over an affinity mask
742     // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
743     virtual int begin() const { return 0; }
744     virtual int end() const { return 0; }
745     virtual int next(int previous) const { return 0; }
746 #if KMP_OS_WINDOWS
747     virtual int set_process_affinity(bool abort_on_error) const { return -1; }
748 #endif
749     // Set the system's affinity to this affinity mask's value
750     virtual int set_system_affinity(bool abort_on_error) const { return -1; }
751     // Set this affinity mask to the current system affinity
752     virtual int get_system_affinity(bool abort_on_error) { return -1; }
753     // Only 1 DWORD in the mask should have any procs set.
754     // Return the appropriate index, or -1 for an invalid mask.
755     virtual int get_proc_group() const { return -1; }
756   };
757   void *operator new(size_t n);
758   void operator delete(void *p);
759   // Need virtual destructor
760   virtual ~KMPAffinity() = default;
761   // Determine if affinity is capable
762   virtual void determine_capable(const char *env_var) {}
763   // Bind the current thread to os proc
764   virtual void bind_thread(int proc) {}
765   // Factory functions to allocate/deallocate a mask
766   virtual Mask *allocate_mask() { return nullptr; }
767   virtual void deallocate_mask(Mask *m) {}
768   virtual Mask *allocate_mask_array(int num) { return nullptr; }
769   virtual void deallocate_mask_array(Mask *m) {}
770   virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
771   static void pick_api();
772   static void destroy_api();
773   enum api_type {
774     NATIVE_OS
775 #if KMP_USE_HWLOC
776     ,
777     HWLOC
778 #endif
779   };
780   virtual api_type get_api_type() const {
781     KMP_ASSERT(0);
782     return NATIVE_OS;
783   }
784 
785 private:
786   static bool picked_api;
787 };
788 
789 typedef KMPAffinity::Mask kmp_affin_mask_t;
790 extern KMPAffinity *__kmp_affinity_dispatch;
791 
792 // Declare local char buffers with this size for printing debug and info
793 // messages, using __kmp_affinity_print_mask().
794 #define KMP_AFFIN_MASK_PRINT_LEN 1024
795 
796 enum affinity_type {
797   affinity_none = 0,
798   affinity_physical,
799   affinity_logical,
800   affinity_compact,
801   affinity_scatter,
802   affinity_explicit,
803   affinity_balanced,
804   affinity_disabled, // not used outsize the env var parser
805   affinity_default
806 };
807 
808 enum affinity_top_method {
809   affinity_top_method_all = 0, // try all (supported) methods, in order
810 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
811   affinity_top_method_apicid,
812   affinity_top_method_x2apicid,
813   affinity_top_method_x2apicid_1f,
814 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
815   affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
816 #if KMP_GROUP_AFFINITY
817   affinity_top_method_group,
818 #endif /* KMP_GROUP_AFFINITY */
819   affinity_top_method_flat,
820 #if KMP_USE_HWLOC
821   affinity_top_method_hwloc,
822 #endif
823   affinity_top_method_default
824 };
825 
826 #define affinity_respect_mask_default (-1)
827 
828 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
829 extern kmp_hw_t __kmp_affinity_gran; /* Affinity granularity */
830 extern int __kmp_affinity_gran_levels; /* corresponding int value */
831 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
832 extern enum affinity_top_method __kmp_affinity_top_method;
833 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
834 extern int __kmp_affinity_offset; /* Affinity offset value  */
835 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
836 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
837 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
838 extern char *__kmp_affinity_proclist; /* proc ID list */
839 extern kmp_affin_mask_t *__kmp_affinity_masks;
840 extern unsigned __kmp_affinity_num_masks;
841 extern void __kmp_affinity_bind_thread(int which);
842 
843 extern kmp_affin_mask_t *__kmp_affin_fullMask;
844 extern kmp_affin_mask_t *__kmp_affin_origMask;
845 extern char *__kmp_cpuinfo_file;
846 extern bool __kmp_affin_reset;
847 
848 #endif /* KMP_AFFINITY_SUPPORTED */
849 
850 // This needs to be kept in sync with the values in omp.h !!!
851 typedef enum kmp_proc_bind_t {
852   proc_bind_false = 0,
853   proc_bind_true,
854   proc_bind_primary,
855   proc_bind_close,
856   proc_bind_spread,
857   proc_bind_intel, // use KMP_AFFINITY interface
858   proc_bind_default
859 } kmp_proc_bind_t;
860 
861 typedef struct kmp_nested_proc_bind_t {
862   kmp_proc_bind_t *bind_types;
863   int size;
864   int used;
865 } kmp_nested_proc_bind_t;
866 
867 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
868 extern kmp_proc_bind_t __kmp_teams_proc_bind;
869 
870 extern int __kmp_display_affinity;
871 extern char *__kmp_affinity_format;
872 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
873 #if OMPT_SUPPORT
874 extern int __kmp_tool;
875 extern char *__kmp_tool_libraries;
876 #endif // OMPT_SUPPORT
877 
878 #if KMP_AFFINITY_SUPPORTED
879 #define KMP_PLACE_ALL (-1)
880 #define KMP_PLACE_UNDEFINED (-2)
881 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
882 #define KMP_AFFINITY_NON_PROC_BIND                                             \
883   ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false ||                 \
884     __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) &&                \
885    (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
886 #endif /* KMP_AFFINITY_SUPPORTED */
887 
888 extern int __kmp_affinity_num_places;
889 
890 typedef enum kmp_cancel_kind_t {
891   cancel_noreq = 0,
892   cancel_parallel = 1,
893   cancel_loop = 2,
894   cancel_sections = 3,
895   cancel_taskgroup = 4
896 } kmp_cancel_kind_t;
897 
898 // KMP_HW_SUBSET support:
899 typedef struct kmp_hws_item {
900   int num;
901   int offset;
902 } kmp_hws_item_t;
903 
904 extern kmp_hws_item_t __kmp_hws_socket;
905 extern kmp_hws_item_t __kmp_hws_die;
906 extern kmp_hws_item_t __kmp_hws_node;
907 extern kmp_hws_item_t __kmp_hws_tile;
908 extern kmp_hws_item_t __kmp_hws_core;
909 extern kmp_hws_item_t __kmp_hws_proc;
910 extern int __kmp_hws_requested;
911 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
912 
913 /* ------------------------------------------------------------------------ */
914 
915 #define KMP_PAD(type, sz)                                                      \
916   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
917 
918 // We need to avoid using -1 as a GTID as +1 is added to the gtid
919 // when storing it in a lock, and the value 0 is reserved.
920 #define KMP_GTID_DNE (-2) /* Does not exist */
921 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
922 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
923 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
924 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
925 
926 /* OpenMP 5.0 Memory Management support */
927 
928 #ifndef __OMP_H
929 // Duplicate type definitions from omp.h
930 typedef uintptr_t omp_uintptr_t;
931 
932 typedef enum {
933   omp_atk_sync_hint = 1,
934   omp_atk_alignment = 2,
935   omp_atk_access = 3,
936   omp_atk_pool_size = 4,
937   omp_atk_fallback = 5,
938   omp_atk_fb_data = 6,
939   omp_atk_pinned = 7,
940   omp_atk_partition = 8
941 } omp_alloctrait_key_t;
942 
943 typedef enum {
944   omp_atv_false = 0,
945   omp_atv_true = 1,
946   omp_atv_contended = 3,
947   omp_atv_uncontended = 4,
948   omp_atv_serialized = 5,
949   omp_atv_sequential = omp_atv_serialized, // (deprecated)
950   omp_atv_private = 6,
951   omp_atv_all = 7,
952   omp_atv_thread = 8,
953   omp_atv_pteam = 9,
954   omp_atv_cgroup = 10,
955   omp_atv_default_mem_fb = 11,
956   omp_atv_null_fb = 12,
957   omp_atv_abort_fb = 13,
958   omp_atv_allocator_fb = 14,
959   omp_atv_environment = 15,
960   omp_atv_nearest = 16,
961   omp_atv_blocked = 17,
962   omp_atv_interleaved = 18
963 } omp_alloctrait_value_t;
964 #define omp_atv_default ((omp_uintptr_t)-1)
965 
966 typedef void *omp_memspace_handle_t;
967 extern omp_memspace_handle_t const omp_default_mem_space;
968 extern omp_memspace_handle_t const omp_large_cap_mem_space;
969 extern omp_memspace_handle_t const omp_const_mem_space;
970 extern omp_memspace_handle_t const omp_high_bw_mem_space;
971 extern omp_memspace_handle_t const omp_low_lat_mem_space;
972 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
973 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
974 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
975 
976 typedef struct {
977   omp_alloctrait_key_t key;
978   omp_uintptr_t value;
979 } omp_alloctrait_t;
980 
981 typedef void *omp_allocator_handle_t;
982 extern omp_allocator_handle_t const omp_null_allocator;
983 extern omp_allocator_handle_t const omp_default_mem_alloc;
984 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
985 extern omp_allocator_handle_t const omp_const_mem_alloc;
986 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
987 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
988 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
989 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
990 extern omp_allocator_handle_t const omp_thread_mem_alloc;
991 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
992 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
993 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
994 extern omp_allocator_handle_t const kmp_max_mem_alloc;
995 extern omp_allocator_handle_t __kmp_def_allocator;
996 
997 // end of duplicate type definitions from omp.h
998 #endif
999 
1000 extern int __kmp_memkind_available;
1001 
1002 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1003 
1004 typedef struct kmp_allocator_t {
1005   omp_memspace_handle_t memspace;
1006   void **memkind; // pointer to memkind
1007   size_t alignment;
1008   omp_alloctrait_value_t fb;
1009   kmp_allocator_t *fb_data;
1010   kmp_uint64 pool_size;
1011   kmp_uint64 pool_used;
1012 } kmp_allocator_t;
1013 
1014 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1015                                                     omp_memspace_handle_t,
1016                                                     int ntraits,
1017                                                     omp_alloctrait_t traits[]);
1018 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1019 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1020 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1021 // external interfaces, may be used by compiler
1022 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1023 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1024                                   omp_allocator_handle_t al);
1025 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1026                            omp_allocator_handle_t al);
1027 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1028                             omp_allocator_handle_t al,
1029                             omp_allocator_handle_t free_al);
1030 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1031 // internal interfaces, contain real implementation
1032 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1033                          omp_allocator_handle_t al);
1034 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1035                           omp_allocator_handle_t al);
1036 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1037                            omp_allocator_handle_t al,
1038                            omp_allocator_handle_t free_al);
1039 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1040 
1041 extern void __kmp_init_memkind();
1042 extern void __kmp_fini_memkind();
1043 extern void __kmp_init_target_mem();
1044 
1045 /* ------------------------------------------------------------------------ */
1046 
1047 #define KMP_UINT64_MAX                                                         \
1048   (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1049 
1050 #define KMP_MIN_NTH 1
1051 
1052 #ifndef KMP_MAX_NTH
1053 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1054 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1055 #else
1056 #define KMP_MAX_NTH INT_MAX
1057 #endif
1058 #endif /* KMP_MAX_NTH */
1059 
1060 #ifdef PTHREAD_STACK_MIN
1061 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1062 #else
1063 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1064 #endif
1065 
1066 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1067 
1068 #if KMP_ARCH_X86
1069 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1070 #elif KMP_ARCH_X86_64
1071 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1072 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1073 #else
1074 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1075 #endif
1076 
1077 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1078 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1079 #define KMP_MAX_MALLOC_POOL_INCR                                               \
1080   (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1081 
1082 #define KMP_MIN_STKOFFSET (0)
1083 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1084 #if KMP_OS_DARWIN
1085 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1086 #else
1087 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1088 #endif
1089 
1090 #define KMP_MIN_STKPADDING (0)
1091 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1092 
1093 #define KMP_BLOCKTIME_MULTIPLIER                                               \
1094   (1000) /* number of blocktime units per second */
1095 #define KMP_MIN_BLOCKTIME (0)
1096 #define KMP_MAX_BLOCKTIME                                                      \
1097   (INT_MAX) /* Must be this for "infinite" setting the work */
1098 
1099 /* __kmp_blocktime is in milliseconds */
1100 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200))
1101 
1102 #if KMP_USE_MONITOR
1103 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1104 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1105 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1106 
1107 /* Calculate new number of monitor wakeups for a specific block time based on
1108    previous monitor_wakeups. Only allow increasing number of wakeups */
1109 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups)                 \
1110   (((blocktime) == KMP_MAX_BLOCKTIME)   ? (monitor_wakeups)                    \
1111    : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS              \
1112    : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime)))            \
1113        ? (monitor_wakeups)                                                     \
1114        : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1115 
1116 /* Calculate number of intervals for a specific block time based on
1117    monitor_wakeups */
1118 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups)               \
1119   (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) /        \
1120    (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1121 #else
1122 #define KMP_BLOCKTIME(team, tid)                                               \
1123   (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1124 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1125 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1126 extern kmp_uint64 __kmp_ticks_per_msec;
1127 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1128 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1129 #else
1130 #define KMP_NOW() __kmp_hardware_timestamp()
1131 #endif
1132 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1133 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1134   (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1135 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1136 #else
1137 // System time is retrieved sporadically while blocking.
1138 extern kmp_uint64 __kmp_now_nsec();
1139 #define KMP_NOW() __kmp_now_nsec()
1140 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1141 #define KMP_BLOCKTIME_INTERVAL(team, tid)                                      \
1142   (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1143 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1144 #endif
1145 #endif // KMP_USE_MONITOR
1146 
1147 #define KMP_MIN_STATSCOLS 40
1148 #define KMP_MAX_STATSCOLS 4096
1149 #define KMP_DEFAULT_STATSCOLS 80
1150 
1151 #define KMP_MIN_INTERVAL 0
1152 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1153 #define KMP_DEFAULT_INTERVAL 0
1154 
1155 #define KMP_MIN_CHUNK 1
1156 #define KMP_MAX_CHUNK (INT_MAX - 1)
1157 #define KMP_DEFAULT_CHUNK 1
1158 
1159 #define KMP_MIN_DISP_NUM_BUFF 1
1160 #define KMP_DFLT_DISP_NUM_BUFF 7
1161 #define KMP_MAX_DISP_NUM_BUFF 4096
1162 
1163 #define KMP_MAX_ORDERED 8
1164 
1165 #define KMP_MAX_FIELDS 32
1166 
1167 #define KMP_MAX_BRANCH_BITS 31
1168 
1169 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1170 
1171 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1172 
1173 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1174 
1175 /* Minimum number of threads before switch to TLS gtid (experimentally
1176    determined) */
1177 /* josh TODO: what about OS X* tuning? */
1178 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1179 #define KMP_TLS_GTID_MIN 5
1180 #else
1181 #define KMP_TLS_GTID_MIN INT_MAX
1182 #endif
1183 
1184 #define KMP_MASTER_TID(tid) (0 == (tid))
1185 #define KMP_WORKER_TID(tid) (0 != (tid))
1186 
1187 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1188 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1189 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1190 
1191 #ifndef TRUE
1192 #define FALSE 0
1193 #define TRUE (!FALSE)
1194 #endif
1195 
1196 /* NOTE: all of the following constants must be even */
1197 
1198 #if KMP_OS_WINDOWS
1199 #define KMP_INIT_WAIT 64U /* initial number of spin-tests   */
1200 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1201 #elif KMP_OS_LINUX
1202 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1203 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1204 #elif KMP_OS_DARWIN
1205 /* TODO: tune for KMP_OS_DARWIN */
1206 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1207 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1208 #elif KMP_OS_DRAGONFLY
1209 /* TODO: tune for KMP_OS_DRAGONFLY */
1210 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1211 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1212 #elif KMP_OS_FREEBSD
1213 /* TODO: tune for KMP_OS_FREEBSD */
1214 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1215 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1216 #elif KMP_OS_NETBSD
1217 /* TODO: tune for KMP_OS_NETBSD */
1218 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1219 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1220 #elif KMP_OS_HURD
1221 /* TODO: tune for KMP_OS_HURD */
1222 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1223 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1224 #elif KMP_OS_OPENBSD
1225 /* TODO: tune for KMP_OS_OPENBSD */
1226 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests   */
1227 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1228 #endif
1229 
1230 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1231 typedef struct kmp_cpuid {
1232   kmp_uint32 eax;
1233   kmp_uint32 ebx;
1234   kmp_uint32 ecx;
1235   kmp_uint32 edx;
1236 } kmp_cpuid_t;
1237 
1238 typedef struct kmp_cpuinfo_flags_t {
1239   unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1240   unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1241   unsigned hybrid : 1;
1242   unsigned reserved : 29; // Ensure size of 32 bits
1243 } kmp_cpuinfo_flags_t;
1244 
1245 typedef struct kmp_cpuinfo {
1246   int initialized; // If 0, other fields are not initialized.
1247   int signature; // CPUID(1).EAX
1248   int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1249   int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1250   // Model << 4 ) + Model)
1251   int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1252   kmp_cpuinfo_flags_t flags;
1253   int apic_id;
1254   int physical_id;
1255   int logical_id;
1256   kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1257   char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1258 } kmp_cpuinfo_t;
1259 
1260 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1261 
1262 #if KMP_OS_UNIX
1263 // subleaf is only needed for cache and topology discovery and can be set to
1264 // zero in most cases
1265 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1266   __asm__ __volatile__("cpuid"
1267                        : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1268                        : "a"(leaf), "c"(subleaf));
1269 }
1270 // Load p into FPU control word
1271 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1272   __asm__ __volatile__("fldcw %0" : : "m"(*p));
1273 }
1274 // Store FPU control word into p
1275 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1276   __asm__ __volatile__("fstcw %0" : "=m"(*p));
1277 }
1278 static inline void __kmp_clear_x87_fpu_status_word() {
1279 #if KMP_MIC
1280   // 32-bit protected mode x87 FPU state
1281   struct x87_fpu_state {
1282     unsigned cw;
1283     unsigned sw;
1284     unsigned tw;
1285     unsigned fip;
1286     unsigned fips;
1287     unsigned fdp;
1288     unsigned fds;
1289   };
1290   struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1291   __asm__ __volatile__("fstenv %0\n\t" // store FP env
1292                        "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1293                        "fldenv %0\n\t" // load FP env back
1294                        : "+m"(fpu_state), "+m"(fpu_state.sw));
1295 #else
1296   __asm__ __volatile__("fnclex");
1297 #endif // KMP_MIC
1298 }
1299 #if __SSE__
1300 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1301 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1302 #else
1303 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1304 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1305 #endif
1306 #else
1307 // Windows still has these as external functions in assembly file
1308 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1309 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1310 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1311 extern void __kmp_clear_x87_fpu_status_word();
1312 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1313 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1314 #endif // KMP_OS_UNIX
1315 
1316 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1317 
1318 // User-level Monitor/Mwait
1319 #if KMP_HAVE_UMWAIT
1320 // We always try for UMWAIT first
1321 #if KMP_HAVE_WAITPKG_INTRINSICS
1322 #if KMP_HAVE_IMMINTRIN_H
1323 #include <immintrin.h>
1324 #elif KMP_HAVE_INTRIN_H
1325 #include <intrin.h>
1326 #endif
1327 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1328 
1329 KMP_ATTRIBUTE_TARGET_WAITPKG
1330 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1331 #if !KMP_HAVE_WAITPKG_INTRINSICS
1332   uint32_t timeHi = uint32_t(counter >> 32);
1333   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1334   char flag;
1335   __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1336                    "setb   %0"
1337                    // The "=q" restraint means any register accessible as rl
1338                    //   in 32-bit mode: a, b, c, and d;
1339                    //   in 64-bit mode: any integer register
1340                    : "=q"(flag)
1341                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1342                    :);
1343   return flag;
1344 #else
1345   return _tpause(hint, counter);
1346 #endif
1347 }
1348 KMP_ATTRIBUTE_TARGET_WAITPKG
1349 static inline void __kmp_umonitor(void *cacheline) {
1350 #if !KMP_HAVE_WAITPKG_INTRINSICS
1351   __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1352                    :
1353                    : "a"(cacheline)
1354                    :);
1355 #else
1356   _umonitor(cacheline);
1357 #endif
1358 }
1359 KMP_ATTRIBUTE_TARGET_WAITPKG
1360 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1361 #if !KMP_HAVE_WAITPKG_INTRINSICS
1362   uint32_t timeHi = uint32_t(counter >> 32);
1363   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1364   char flag;
1365   __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1366                    "setb   %0"
1367                    // The "=q" restraint means any register accessible as rl
1368                    //   in 32-bit mode: a, b, c, and d;
1369                    //   in 64-bit mode: any integer register
1370                    : "=q"(flag)
1371                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1372                    :);
1373   return flag;
1374 #else
1375   return _umwait(hint, counter);
1376 #endif
1377 }
1378 #elif KMP_HAVE_MWAIT
1379 #if KMP_OS_UNIX
1380 #include <pmmintrin.h>
1381 #else
1382 #include <intrin.h>
1383 #endif
1384 #if KMP_OS_UNIX
1385 __attribute__((target("sse3")))
1386 #endif
1387 static inline void
1388 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1389   _mm_monitor(cacheline, extensions, hints);
1390 }
1391 #if KMP_OS_UNIX
1392 __attribute__((target("sse3")))
1393 #endif
1394 static inline void
1395 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1396   _mm_mwait(extensions, hints);
1397 }
1398 #endif // KMP_HAVE_UMWAIT
1399 
1400 #if KMP_ARCH_X86
1401 extern void __kmp_x86_pause(void);
1402 #elif KMP_MIC
1403 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1404 // regression after removal of extra PAUSE from spin loops. Changing
1405 // the delay from 100 to 300 showed even better performance than double PAUSE
1406 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1407 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1408 #else
1409 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1410 #endif
1411 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1412 #elif KMP_ARCH_PPC64
1413 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1414 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1415 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1416 #define KMP_CPU_PAUSE()                                                        \
1417   do {                                                                         \
1418     KMP_PPC64_PRI_LOW();                                                       \
1419     KMP_PPC64_PRI_MED();                                                       \
1420     KMP_PPC64_PRI_LOC_MB();                                                    \
1421   } while (0)
1422 #else
1423 #define KMP_CPU_PAUSE() /* nothing to do */
1424 #endif
1425 
1426 #define KMP_INIT_YIELD(count)                                                  \
1427   { (count) = __kmp_yield_init; }
1428 
1429 #define KMP_INIT_BACKOFF(time)                                                 \
1430   { (time) = __kmp_pause_init; }
1431 
1432 #define KMP_OVERSUBSCRIBED                                                     \
1433   (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1434 
1435 #define KMP_TRY_YIELD                                                          \
1436   ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1437 
1438 #define KMP_TRY_YIELD_OVERSUB                                                  \
1439   ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1440 
1441 #define KMP_YIELD(cond)                                                        \
1442   {                                                                            \
1443     KMP_CPU_PAUSE();                                                           \
1444     if ((cond) && (KMP_TRY_YIELD))                                             \
1445       __kmp_yield();                                                           \
1446   }
1447 
1448 #define KMP_YIELD_OVERSUB()                                                    \
1449   {                                                                            \
1450     KMP_CPU_PAUSE();                                                           \
1451     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1452       __kmp_yield();                                                           \
1453   }
1454 
1455 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1456 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1457 #define KMP_YIELD_SPIN(count)                                                  \
1458   {                                                                            \
1459     KMP_CPU_PAUSE();                                                           \
1460     if (KMP_TRY_YIELD) {                                                       \
1461       (count) -= 2;                                                            \
1462       if (!(count)) {                                                          \
1463         __kmp_yield();                                                         \
1464         (count) = __kmp_yield_next;                                            \
1465       }                                                                        \
1466     }                                                                          \
1467   }
1468 
1469 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1470 // (C0.2) state, which improves performance of other SMT threads on the same
1471 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1472 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1473 // available, fall back to the regular CPU pause and yield combination.
1474 #if KMP_HAVE_UMWAIT
1475 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1476   {                                                                            \
1477     if (__kmp_tpause_enabled) {                                                \
1478       if (KMP_OVERSUBSCRIBED) {                                                \
1479         __kmp_tpause(0, (time));                                               \
1480       } else {                                                                 \
1481         __kmp_tpause(__kmp_tpause_hint, (time));                               \
1482       }                                                                        \
1483       (time) *= 2;                                                             \
1484     } else {                                                                   \
1485       KMP_CPU_PAUSE();                                                         \
1486       if ((KMP_TRY_YIELD_OVERSUB)) {                                           \
1487         __kmp_yield();                                                         \
1488       } else if (__kmp_use_yield == 1) {                                       \
1489         (count) -= 2;                                                          \
1490         if (!(count)) {                                                        \
1491           __kmp_yield();                                                       \
1492           (count) = __kmp_yield_next;                                          \
1493         }                                                                      \
1494       }                                                                        \
1495     }                                                                          \
1496   }
1497 #else
1498 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1499   {                                                                            \
1500     KMP_CPU_PAUSE();                                                           \
1501     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1502       __kmp_yield();                                                           \
1503     else if (__kmp_use_yield == 1) {                                           \
1504       (count) -= 2;                                                            \
1505       if (!(count)) {                                                          \
1506         __kmp_yield();                                                         \
1507         (count) = __kmp_yield_next;                                            \
1508       }                                                                        \
1509     }                                                                          \
1510   }
1511 #endif // KMP_HAVE_UMWAIT
1512 
1513 /* ------------------------------------------------------------------------ */
1514 /* Support datatypes for the orphaned construct nesting checks.             */
1515 /* ------------------------------------------------------------------------ */
1516 
1517 /* When adding to this enum, add its corresponding string in cons_text_c[]
1518  * array in kmp_error.cpp */
1519 enum cons_type {
1520   ct_none,
1521   ct_parallel,
1522   ct_pdo,
1523   ct_pdo_ordered,
1524   ct_psections,
1525   ct_psingle,
1526   ct_critical,
1527   ct_ordered_in_parallel,
1528   ct_ordered_in_pdo,
1529   ct_master,
1530   ct_reduce,
1531   ct_barrier,
1532   ct_masked
1533 };
1534 
1535 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1536 
1537 struct cons_data {
1538   ident_t const *ident;
1539   enum cons_type type;
1540   int prev;
1541   kmp_user_lock_p
1542       name; /* address exclusively for critical section name comparison */
1543 };
1544 
1545 struct cons_header {
1546   int p_top, w_top, s_top;
1547   int stack_size, stack_top;
1548   struct cons_data *stack_data;
1549 };
1550 
1551 struct kmp_region_info {
1552   char *text;
1553   int offset[KMP_MAX_FIELDS];
1554   int length[KMP_MAX_FIELDS];
1555 };
1556 
1557 /* ---------------------------------------------------------------------- */
1558 /* ---------------------------------------------------------------------- */
1559 
1560 #if KMP_OS_WINDOWS
1561 typedef HANDLE kmp_thread_t;
1562 typedef DWORD kmp_key_t;
1563 #endif /* KMP_OS_WINDOWS */
1564 
1565 #if KMP_OS_UNIX
1566 typedef pthread_t kmp_thread_t;
1567 typedef pthread_key_t kmp_key_t;
1568 #endif
1569 
1570 extern kmp_key_t __kmp_gtid_threadprivate_key;
1571 
1572 typedef struct kmp_sys_info {
1573   long maxrss; /* the maximum resident set size utilized (in kilobytes)     */
1574   long minflt; /* the number of page faults serviced without any I/O        */
1575   long majflt; /* the number of page faults serviced that required I/O      */
1576   long nswap; /* the number of times a process was "swapped" out of memory */
1577   long inblock; /* the number of times the file system had to perform input  */
1578   long oublock; /* the number of times the file system had to perform output */
1579   long nvcsw; /* the number of times a context switch was voluntarily      */
1580   long nivcsw; /* the number of times a context switch was forced           */
1581 } kmp_sys_info_t;
1582 
1583 #if USE_ITT_BUILD
1584 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1585 // required type here. Later we will check the type meets requirements.
1586 typedef int kmp_itt_mark_t;
1587 #define KMP_ITT_DEBUG 0
1588 #endif /* USE_ITT_BUILD */
1589 
1590 typedef kmp_int32 kmp_critical_name[8];
1591 
1592 /*!
1593 @ingroup PARALLEL
1594 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1595 The arguments to the outlined function are
1596 @param global_tid the global thread identity of the thread executing the
1597 function.
1598 @param bound_tid  the local identity of the thread executing the function
1599 @param ... pointers to shared variables accessed by the function.
1600 */
1601 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1602 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1603                                  ...);
1604 
1605 /*!
1606 @ingroup THREADPRIVATE
1607 @{
1608 */
1609 /* ---------------------------------------------------------------------------
1610  */
1611 /* Threadprivate initialization/finalization function declarations */
1612 
1613 /*  for non-array objects:  __kmpc_threadprivate_register()  */
1614 
1615 /*!
1616  Pointer to the constructor function.
1617  The first argument is the <tt>this</tt> pointer
1618 */
1619 typedef void *(*kmpc_ctor)(void *);
1620 
1621 /*!
1622  Pointer to the destructor function.
1623  The first argument is the <tt>this</tt> pointer
1624 */
1625 typedef void (*kmpc_dtor)(
1626     void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1627                               compiler */
1628 /*!
1629  Pointer to an alternate constructor.
1630  The first argument is the <tt>this</tt> pointer.
1631 */
1632 typedef void *(*kmpc_cctor)(void *, void *);
1633 
1634 /* for array objects: __kmpc_threadprivate_register_vec() */
1635 /* First arg: "this" pointer */
1636 /* Last arg: number of array elements */
1637 /*!
1638  Array constructor.
1639  First argument is the <tt>this</tt> pointer
1640  Second argument the number of array elements.
1641 */
1642 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1643 /*!
1644  Pointer to the array destructor function.
1645  The first argument is the <tt>this</tt> pointer
1646  Second argument the number of array elements.
1647 */
1648 typedef void (*kmpc_dtor_vec)(void *, size_t);
1649 /*!
1650  Array constructor.
1651  First argument is the <tt>this</tt> pointer
1652  Third argument the number of array elements.
1653 */
1654 typedef void *(*kmpc_cctor_vec)(void *, void *,
1655                                 size_t); /* function unused by compiler */
1656 
1657 /*!
1658 @}
1659 */
1660 
1661 /* keeps tracked of threadprivate cache allocations for cleanup later */
1662 typedef struct kmp_cached_addr {
1663   void **addr; /* address of allocated cache */
1664   void ***compiler_cache; /* pointer to compiler's cache */
1665   void *data; /* pointer to global data */
1666   struct kmp_cached_addr *next; /* pointer to next cached address */
1667 } kmp_cached_addr_t;
1668 
1669 struct private_data {
1670   struct private_data *next; /* The next descriptor in the list      */
1671   void *data; /* The data buffer for this descriptor  */
1672   int more; /* The repeat count for this descriptor */
1673   size_t size; /* The data size for this descriptor    */
1674 };
1675 
1676 struct private_common {
1677   struct private_common *next;
1678   struct private_common *link;
1679   void *gbl_addr;
1680   void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1681   size_t cmn_size;
1682 };
1683 
1684 struct shared_common {
1685   struct shared_common *next;
1686   struct private_data *pod_init;
1687   void *obj_init;
1688   void *gbl_addr;
1689   union {
1690     kmpc_ctor ctor;
1691     kmpc_ctor_vec ctorv;
1692   } ct;
1693   union {
1694     kmpc_cctor cctor;
1695     kmpc_cctor_vec cctorv;
1696   } cct;
1697   union {
1698     kmpc_dtor dtor;
1699     kmpc_dtor_vec dtorv;
1700   } dt;
1701   size_t vec_len;
1702   int is_vec;
1703   size_t cmn_size;
1704 };
1705 
1706 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1707 #define KMP_HASH_TABLE_SIZE                                                    \
1708   (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1709 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1710 #define KMP_HASH(x)                                                            \
1711   ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1712 
1713 struct common_table {
1714   struct private_common *data[KMP_HASH_TABLE_SIZE];
1715 };
1716 
1717 struct shared_table {
1718   struct shared_common *data[KMP_HASH_TABLE_SIZE];
1719 };
1720 
1721 /* ------------------------------------------------------------------------ */
1722 
1723 #if KMP_USE_HIER_SCHED
1724 // Shared barrier data that exists inside a single unit of the scheduling
1725 // hierarchy
1726 typedef struct kmp_hier_private_bdata_t {
1727   kmp_int32 num_active;
1728   kmp_uint64 index;
1729   kmp_uint64 wait_val[2];
1730 } kmp_hier_private_bdata_t;
1731 #endif
1732 
1733 typedef struct kmp_sched_flags {
1734   unsigned ordered : 1;
1735   unsigned nomerge : 1;
1736   unsigned contains_last : 1;
1737 #if KMP_USE_HIER_SCHED
1738   unsigned use_hier : 1;
1739   unsigned unused : 28;
1740 #else
1741   unsigned unused : 29;
1742 #endif
1743 } kmp_sched_flags_t;
1744 
1745 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1746 
1747 #if KMP_STATIC_STEAL_ENABLED
1748 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1749   kmp_int32 count;
1750   kmp_int32 ub;
1751   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1752   kmp_int32 lb;
1753   kmp_int32 st;
1754   kmp_int32 tc;
1755   kmp_lock_t *steal_lock; // lock used for chunk stealing
1756   // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1757   //    a) parm3 is properly aligned and
1758   //    b) all parm1-4 are on the same cache line.
1759   // Because of parm1-4 are used together, performance seems to be better
1760   // if they are on the same cache line (not measured though).
1761 
1762   struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1763     kmp_int32 parm1; //     structures in kmp_dispatch.cpp. This should
1764     kmp_int32 parm2; //     make no real change at least while padding is off.
1765     kmp_int32 parm3;
1766     kmp_int32 parm4;
1767   };
1768 
1769   kmp_uint32 ordered_lower;
1770   kmp_uint32 ordered_upper;
1771 #if KMP_OS_WINDOWS
1772   kmp_int32 last_upper;
1773 #endif /* KMP_OS_WINDOWS */
1774 } dispatch_private_info32_t;
1775 
1776 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1777   kmp_int64 count; // current chunk number for static & static-steal scheduling
1778   kmp_int64 ub; /* upper-bound */
1779   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1780   kmp_int64 lb; /* lower-bound */
1781   kmp_int64 st; /* stride */
1782   kmp_int64 tc; /* trip count (number of iterations) */
1783   kmp_lock_t *steal_lock; // lock used for chunk stealing
1784   /* parm[1-4] are used in different ways by different scheduling algorithms */
1785 
1786   // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1787   //    a) parm3 is properly aligned and
1788   //    b) all parm1-4 are in the same cache line.
1789   // Because of parm1-4 are used together, performance seems to be better
1790   // if they are in the same line (not measured though).
1791 
1792   struct KMP_ALIGN(32) {
1793     kmp_int64 parm1;
1794     kmp_int64 parm2;
1795     kmp_int64 parm3;
1796     kmp_int64 parm4;
1797   };
1798 
1799   kmp_uint64 ordered_lower;
1800   kmp_uint64 ordered_upper;
1801 #if KMP_OS_WINDOWS
1802   kmp_int64 last_upper;
1803 #endif /* KMP_OS_WINDOWS */
1804 } dispatch_private_info64_t;
1805 #else /* KMP_STATIC_STEAL_ENABLED */
1806 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1807   kmp_int32 lb;
1808   kmp_int32 ub;
1809   kmp_int32 st;
1810   kmp_int32 tc;
1811 
1812   kmp_int32 parm1;
1813   kmp_int32 parm2;
1814   kmp_int32 parm3;
1815   kmp_int32 parm4;
1816 
1817   kmp_int32 count;
1818 
1819   kmp_uint32 ordered_lower;
1820   kmp_uint32 ordered_upper;
1821 #if KMP_OS_WINDOWS
1822   kmp_int32 last_upper;
1823 #endif /* KMP_OS_WINDOWS */
1824 } dispatch_private_info32_t;
1825 
1826 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1827   kmp_int64 lb; /* lower-bound */
1828   kmp_int64 ub; /* upper-bound */
1829   kmp_int64 st; /* stride */
1830   kmp_int64 tc; /* trip count (number of iterations) */
1831 
1832   /* parm[1-4] are used in different ways by different scheduling algorithms */
1833   kmp_int64 parm1;
1834   kmp_int64 parm2;
1835   kmp_int64 parm3;
1836   kmp_int64 parm4;
1837 
1838   kmp_int64 count; /* current chunk number for static scheduling */
1839 
1840   kmp_uint64 ordered_lower;
1841   kmp_uint64 ordered_upper;
1842 #if KMP_OS_WINDOWS
1843   kmp_int64 last_upper;
1844 #endif /* KMP_OS_WINDOWS */
1845 } dispatch_private_info64_t;
1846 #endif /* KMP_STATIC_STEAL_ENABLED */
1847 
1848 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1849   union private_info {
1850     dispatch_private_info32_t p32;
1851     dispatch_private_info64_t p64;
1852   } u;
1853   enum sched_type schedule; /* scheduling algorithm */
1854   kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1855   std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1856   kmp_int32 ordered_bumped;
1857   // Stack of buffers for nest of serial regions
1858   struct dispatch_private_info *next;
1859   kmp_int32 type_size; /* the size of types in private_info */
1860 #if KMP_USE_HIER_SCHED
1861   kmp_int32 hier_id;
1862   void *parent; /* hierarchical scheduling parent pointer */
1863 #endif
1864   enum cons_type pushed_ws;
1865 } dispatch_private_info_t;
1866 
1867 typedef struct dispatch_shared_info32 {
1868   /* chunk index under dynamic, number of idle threads under static-steal;
1869      iteration index otherwise */
1870   volatile kmp_uint32 iteration;
1871   volatile kmp_int32 num_done;
1872   volatile kmp_uint32 ordered_iteration;
1873   // Dummy to retain the structure size after making ordered_iteration scalar
1874   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1875 } dispatch_shared_info32_t;
1876 
1877 typedef struct dispatch_shared_info64 {
1878   /* chunk index under dynamic, number of idle threads under static-steal;
1879      iteration index otherwise */
1880   volatile kmp_uint64 iteration;
1881   volatile kmp_int64 num_done;
1882   volatile kmp_uint64 ordered_iteration;
1883   // Dummy to retain the structure size after making ordered_iteration scalar
1884   kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1885 } dispatch_shared_info64_t;
1886 
1887 typedef struct dispatch_shared_info {
1888   union shared_info {
1889     dispatch_shared_info32_t s32;
1890     dispatch_shared_info64_t s64;
1891   } u;
1892   volatile kmp_uint32 buffer_index;
1893   volatile kmp_int32 doacross_buf_idx; // teamwise index
1894   volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1895   kmp_int32 doacross_num_done; // count finished threads
1896 #if KMP_USE_HIER_SCHED
1897   void *hier;
1898 #endif
1899 #if KMP_USE_HWLOC
1900   // When linking with libhwloc, the ORDERED EPCC test slows down on big
1901   // machines (> 48 cores). Performance analysis showed that a cache thrash
1902   // was occurring and this padding helps alleviate the problem.
1903   char padding[64];
1904 #endif
1905 } dispatch_shared_info_t;
1906 
1907 typedef struct kmp_disp {
1908   /* Vector for ORDERED SECTION */
1909   void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1910   /* Vector for END ORDERED SECTION */
1911   void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1912 
1913   dispatch_shared_info_t *th_dispatch_sh_current;
1914   dispatch_private_info_t *th_dispatch_pr_current;
1915 
1916   dispatch_private_info_t *th_disp_buffer;
1917   kmp_uint32 th_disp_index;
1918   kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1919   volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1920   kmp_int64 *th_doacross_info; // info on loop bounds
1921 #if KMP_USE_INTERNODE_ALIGNMENT
1922   char more_padding[INTERNODE_CACHE_LINE];
1923 #endif
1924 } kmp_disp_t;
1925 
1926 /* ------------------------------------------------------------------------ */
1927 /* Barrier stuff */
1928 
1929 /* constants for barrier state update */
1930 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1931 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1932 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1933 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1934 
1935 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1936 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1937 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1938 
1939 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1940 #error "Barrier sleep bit must be smaller than barrier bump bit"
1941 #endif
1942 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1943 #error "Barrier unused bit must be smaller than barrier bump bit"
1944 #endif
1945 
1946 // Constants for release barrier wait state: currently, hierarchical only
1947 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1948 #define KMP_BARRIER_OWN_FLAG                                                   \
1949   1 // Normal state; worker waiting on own b_go flag in release
1950 #define KMP_BARRIER_PARENT_FLAG                                                \
1951   2 // Special state; worker waiting on parent's b_go flag in release
1952 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG                                         \
1953   3 // Special state; tells worker to shift from parent to own b_go
1954 #define KMP_BARRIER_SWITCHING                                                  \
1955   4 // Special state; worker resets appropriate flag on wake-up
1956 
1957 #define KMP_NOT_SAFE_TO_REAP                                                   \
1958   0 // Thread th_reap_state: not safe to reap (tasking)
1959 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1960 
1961 // The flag_type describes the storage used for the flag.
1962 enum flag_type {
1963   flag32, /**< atomic 32 bit flags */
1964   flag64, /**< 64 bit flags */
1965   atomic_flag64, /**< atomic 64 bit flags */
1966   flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */
1967   flag_unset
1968 };
1969 
1970 enum barrier_type {
1971   bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1972                            barriers if enabled) */
1973   bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1974 #if KMP_FAST_REDUCTION_BARRIER
1975   bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1976 #endif // KMP_FAST_REDUCTION_BARRIER
1977   bs_last_barrier /* Just a placeholder to mark the end */
1978 };
1979 
1980 // to work with reduction barriers just like with plain barriers
1981 #if !KMP_FAST_REDUCTION_BARRIER
1982 #define bs_reduction_barrier bs_plain_barrier
1983 #endif // KMP_FAST_REDUCTION_BARRIER
1984 
1985 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1986                            bp_linear_bar =
1987                                0, /* Single level (degenerate) tree */
1988                            bp_tree_bar =
1989                                1, /* Balanced tree with branching factor 2^n */
1990                            bp_hyper_bar = 2, /* Hypercube-embedded tree with min
1991                                                 branching factor 2^n */
1992                            bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1993                            bp_dist_bar = 4, /* Distributed barrier */
1994                            bp_last_bar /* Placeholder to mark the end */
1995 } kmp_bar_pat_e;
1996 
1997 #define KMP_BARRIER_ICV_PUSH 1
1998 
1999 /* Record for holding the values of the internal controls stack records */
2000 typedef struct kmp_internal_control {
2001   int serial_nesting_level; /* corresponds to the value of the
2002                                th_team_serialized field */
2003   kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2004                        thread) */
2005   kmp_int8
2006       bt_set; /* internal control for whether blocktime is explicitly set */
2007   int blocktime; /* internal control for blocktime */
2008 #if KMP_USE_MONITOR
2009   int bt_intervals; /* internal control for blocktime intervals */
2010 #endif
2011   int nproc; /* internal control for #threads for next parallel region (per
2012                 thread) */
2013   int thread_limit; /* internal control for thread-limit-var */
2014   int max_active_levels; /* internal control for max_active_levels */
2015   kmp_r_sched_t
2016       sched; /* internal control for runtime schedule {sched,chunk} pair */
2017   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
2018   kmp_int32 default_device; /* internal control for default device */
2019   struct kmp_internal_control *next;
2020 } kmp_internal_control_t;
2021 
2022 static inline void copy_icvs(kmp_internal_control_t *dst,
2023                              kmp_internal_control_t *src) {
2024   *dst = *src;
2025 }
2026 
2027 /* Thread barrier needs volatile barrier fields */
2028 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2029   // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2030   // uses of it). It is not explicitly aligned below, because we *don't* want
2031   // it to be padded -- instead, we fit b_go into the same cache line with
2032   // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2033   kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2034   // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2035   // same NGO store
2036   volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2037   KMP_ALIGN_CACHE volatile kmp_uint64
2038       b_arrived; // STATE => task reached synch point.
2039   kmp_uint32 *skip_per_level;
2040   kmp_uint32 my_level;
2041   kmp_int32 parent_tid;
2042   kmp_int32 old_tid;
2043   kmp_uint32 depth;
2044   struct kmp_bstate *parent_bar;
2045   kmp_team_t *team;
2046   kmp_uint64 leaf_state;
2047   kmp_uint32 nproc;
2048   kmp_uint8 base_leaf_kids;
2049   kmp_uint8 leaf_kids;
2050   kmp_uint8 offset;
2051   kmp_uint8 wait_flag;
2052   kmp_uint8 use_oncore_barrier;
2053 #if USE_DEBUGGER
2054   // The following field is intended for the debugger solely. Only the worker
2055   // thread itself accesses this field: the worker increases it by 1 when it
2056   // arrives to a barrier.
2057   KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2058 #endif /* USE_DEBUGGER */
2059 } kmp_bstate_t;
2060 
2061 union KMP_ALIGN_CACHE kmp_barrier_union {
2062   double b_align; /* use worst case alignment */
2063   char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2064   kmp_bstate_t bb;
2065 };
2066 
2067 typedef union kmp_barrier_union kmp_balign_t;
2068 
2069 /* Team barrier needs only non-volatile arrived counter */
2070 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2071   double b_align; /* use worst case alignment */
2072   char b_pad[CACHE_LINE];
2073   struct {
2074     kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2075 #if USE_DEBUGGER
2076     // The following two fields are indended for the debugger solely. Only
2077     // primary thread of the team accesses these fields: the first one is
2078     // increased by 1 when the primary thread arrives to a barrier, the second
2079     // one is increased by one when all the threads arrived.
2080     kmp_uint b_master_arrived;
2081     kmp_uint b_team_arrived;
2082 #endif
2083   };
2084 };
2085 
2086 typedef union kmp_barrier_team_union kmp_balign_team_t;
2087 
2088 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2089    threads when a condition changes.  This is to workaround an NPTL bug where
2090    padding was added to pthread_cond_t which caused the initialization routine
2091    to write outside of the structure if compiled on pre-NPTL threads.  */
2092 #if KMP_OS_WINDOWS
2093 typedef struct kmp_win32_mutex {
2094   /* The Lock */
2095   CRITICAL_SECTION cs;
2096 } kmp_win32_mutex_t;
2097 
2098 typedef struct kmp_win32_cond {
2099   /* Count of the number of waiters. */
2100   int waiters_count_;
2101 
2102   /* Serialize access to <waiters_count_> */
2103   kmp_win32_mutex_t waiters_count_lock_;
2104 
2105   /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2106   int release_count_;
2107 
2108   /* Keeps track of the current "generation" so that we don't allow */
2109   /* one thread to steal all the "releases" from the broadcast. */
2110   int wait_generation_count_;
2111 
2112   /* A manual-reset event that's used to block and release waiting threads. */
2113   HANDLE event_;
2114 } kmp_win32_cond_t;
2115 #endif
2116 
2117 #if KMP_OS_UNIX
2118 
2119 union KMP_ALIGN_CACHE kmp_cond_union {
2120   double c_align;
2121   char c_pad[CACHE_LINE];
2122   pthread_cond_t c_cond;
2123 };
2124 
2125 typedef union kmp_cond_union kmp_cond_align_t;
2126 
2127 union KMP_ALIGN_CACHE kmp_mutex_union {
2128   double m_align;
2129   char m_pad[CACHE_LINE];
2130   pthread_mutex_t m_mutex;
2131 };
2132 
2133 typedef union kmp_mutex_union kmp_mutex_align_t;
2134 
2135 #endif /* KMP_OS_UNIX */
2136 
2137 typedef struct kmp_desc_base {
2138   void *ds_stackbase;
2139   size_t ds_stacksize;
2140   int ds_stackgrow;
2141   kmp_thread_t ds_thread;
2142   volatile int ds_tid;
2143   int ds_gtid;
2144 #if KMP_OS_WINDOWS
2145   volatile int ds_alive;
2146   DWORD ds_thread_id;
2147 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2148    However, debugger support (libomp_db) cannot work with handles, because they
2149    uncomparable. For example, debugger requests info about thread with handle h.
2150    h is valid within debugger process, and meaningless within debugee process.
2151    Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2152    within debugee process, but it is a *new* handle which does *not* equal to
2153    any other handle in debugee... The only way to compare handles is convert
2154    them to system-wide ids. GetThreadId() function is available only in
2155    Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2156    on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2157    thread id by call to GetCurrentThreadId() from within the thread and save it
2158    to let libomp_db identify threads.  */
2159 #endif /* KMP_OS_WINDOWS */
2160 } kmp_desc_base_t;
2161 
2162 typedef union KMP_ALIGN_CACHE kmp_desc {
2163   double ds_align; /* use worst case alignment */
2164   char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2165   kmp_desc_base_t ds;
2166 } kmp_desc_t;
2167 
2168 typedef struct kmp_local {
2169   volatile int this_construct; /* count of single's encountered by thread */
2170   void *reduce_data;
2171 #if KMP_USE_BGET
2172   void *bget_data;
2173   void *bget_list;
2174 #if !USE_CMP_XCHG_FOR_BGET
2175 #ifdef USE_QUEUING_LOCK_FOR_BGET
2176   kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2177 #else
2178   kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2179 // bootstrap lock so we can use it at library
2180 // shutdown.
2181 #endif /* USE_LOCK_FOR_BGET */
2182 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2183 #endif /* KMP_USE_BGET */
2184 
2185   PACKED_REDUCTION_METHOD_T
2186   packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2187                               __kmpc_end_reduce*() */
2188 
2189 } kmp_local_t;
2190 
2191 #define KMP_CHECK_UPDATE(a, b)                                                 \
2192   if ((a) != (b))                                                              \
2193   (a) = (b)
2194 #define KMP_CHECK_UPDATE_SYNC(a, b)                                            \
2195   if ((a) != (b))                                                              \
2196   TCW_SYNC_PTR((a), (b))
2197 
2198 #define get__blocktime(xteam, xtid)                                            \
2199   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2200 #define get__bt_set(xteam, xtid)                                               \
2201   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2202 #if KMP_USE_MONITOR
2203 #define get__bt_intervals(xteam, xtid)                                         \
2204   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2205 #endif
2206 
2207 #define get__dynamic_2(xteam, xtid)                                            \
2208   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2209 #define get__nproc_2(xteam, xtid)                                              \
2210   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2211 #define get__sched_2(xteam, xtid)                                              \
2212   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2213 
2214 #define set__blocktime_team(xteam, xtid, xval)                                 \
2215   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) =     \
2216        (xval))
2217 
2218 #if KMP_USE_MONITOR
2219 #define set__bt_intervals_team(xteam, xtid, xval)                              \
2220   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) =  \
2221        (xval))
2222 #endif
2223 
2224 #define set__bt_set_team(xteam, xtid, xval)                                    \
2225   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2226 
2227 #define set__dynamic(xthread, xval)                                            \
2228   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2229 #define get__dynamic(xthread)                                                  \
2230   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2231 
2232 #define set__nproc(xthread, xval)                                              \
2233   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2234 
2235 #define set__thread_limit(xthread, xval)                                       \
2236   (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2237 
2238 #define set__max_active_levels(xthread, xval)                                  \
2239   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2240 
2241 #define get__max_active_levels(xthread)                                        \
2242   ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2243 
2244 #define set__sched(xthread, xval)                                              \
2245   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2246 
2247 #define set__proc_bind(xthread, xval)                                          \
2248   (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2249 #define get__proc_bind(xthread)                                                \
2250   ((xthread)->th.th_current_task->td_icvs.proc_bind)
2251 
2252 // OpenMP tasking data structures
2253 
2254 typedef enum kmp_tasking_mode {
2255   tskm_immediate_exec = 0,
2256   tskm_extra_barrier = 1,
2257   tskm_task_teams = 2,
2258   tskm_max = 2
2259 } kmp_tasking_mode_t;
2260 
2261 extern kmp_tasking_mode_t
2262     __kmp_tasking_mode; /* determines how/when to execute tasks */
2263 extern int __kmp_task_stealing_constraint;
2264 extern int __kmp_enable_task_throttling;
2265 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2266 // specified, defaults to 0 otherwise
2267 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2268 extern kmp_int32 __kmp_max_task_priority;
2269 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2270 extern kmp_uint64 __kmp_taskloop_min_tasks;
2271 
2272 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2273    taskdata first */
2274 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2275 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2276 
2277 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2278 // were spawned and queued since the previous barrier release.
2279 #define KMP_TASKING_ENABLED(task_team)                                         \
2280   (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2281 /*!
2282 @ingroup BASIC_TYPES
2283 @{
2284 */
2285 
2286 /*!
2287  */
2288 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2289 
2290 typedef union kmp_cmplrdata {
2291   kmp_int32 priority; /**< priority specified by user for the task */
2292   kmp_routine_entry_t
2293       destructors; /* pointer to function to invoke deconstructors of
2294                       firstprivate C++ objects */
2295   /* future data */
2296 } kmp_cmplrdata_t;
2297 
2298 /*  sizeof_kmp_task_t passed as arg to kmpc_omp_task call  */
2299 /*!
2300  */
2301 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2302   void *shareds; /**< pointer to block of pointers to shared vars   */
2303   kmp_routine_entry_t
2304       routine; /**< pointer to routine to call for executing task */
2305   kmp_int32 part_id; /**< part id for the task                          */
2306   kmp_cmplrdata_t
2307       data1; /* Two known optional additions: destructors and priority */
2308   kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2309   /* future data */
2310   /*  private vars  */
2311 } kmp_task_t;
2312 
2313 /*!
2314 @}
2315 */
2316 
2317 typedef struct kmp_taskgroup {
2318   std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2319   std::atomic<kmp_int32>
2320       cancel_request; // request for cancellation of this taskgroup
2321   struct kmp_taskgroup *parent; // parent taskgroup
2322   // Block of data to perform task reduction
2323   void *reduce_data; // reduction related info
2324   kmp_int32 reduce_num_data; // number of data items to reduce
2325   uintptr_t *gomp_data; // gomp reduction data
2326 } kmp_taskgroup_t;
2327 
2328 // forward declarations
2329 typedef union kmp_depnode kmp_depnode_t;
2330 typedef struct kmp_depnode_list kmp_depnode_list_t;
2331 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2332 
2333 // macros for checking dep flag as an integer
2334 #define KMP_DEP_IN 0x1
2335 #define KMP_DEP_OUT 0x2
2336 #define KMP_DEP_INOUT 0x3
2337 #define KMP_DEP_MTX 0x4
2338 #define KMP_DEP_SET 0x8
2339 #define KMP_DEP_ALL 0x80
2340 // Compiler sends us this info:
2341 typedef struct kmp_depend_info {
2342   kmp_intptr_t base_addr;
2343   size_t len;
2344   union {
2345     kmp_uint8 flag; // flag as an unsigned char
2346     struct { // flag as a set of 8 bits
2347       unsigned in : 1;
2348       unsigned out : 1;
2349       unsigned mtx : 1;
2350       unsigned set : 1;
2351       unsigned unused : 3;
2352       unsigned all : 1;
2353     } flags;
2354   };
2355 } kmp_depend_info_t;
2356 
2357 // Internal structures to work with task dependencies:
2358 struct kmp_depnode_list {
2359   kmp_depnode_t *node;
2360   kmp_depnode_list_t *next;
2361 };
2362 
2363 // Max number of mutexinoutset dependencies per node
2364 #define MAX_MTX_DEPS 4
2365 
2366 typedef struct kmp_base_depnode {
2367   kmp_depnode_list_t *successors; /* used under lock */
2368   kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2369   kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2370   kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2371   kmp_lock_t lock; /* guards shared fields: task, successors */
2372 #if KMP_SUPPORT_GRAPH_OUTPUT
2373   kmp_uint32 id;
2374 #endif
2375   std::atomic<kmp_int32> npredecessors;
2376   std::atomic<kmp_int32> nrefs;
2377 } kmp_base_depnode_t;
2378 
2379 union KMP_ALIGN_CACHE kmp_depnode {
2380   double dn_align; /* use worst case alignment */
2381   char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2382   kmp_base_depnode_t dn;
2383 };
2384 
2385 struct kmp_dephash_entry {
2386   kmp_intptr_t addr;
2387   kmp_depnode_t *last_out;
2388   kmp_depnode_list_t *last_set;
2389   kmp_depnode_list_t *prev_set;
2390   kmp_uint8 last_flag;
2391   kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2392   kmp_dephash_entry_t *next_in_bucket;
2393 };
2394 
2395 typedef struct kmp_dephash {
2396   kmp_dephash_entry_t **buckets;
2397   size_t size;
2398   kmp_depnode_t *last_all;
2399   size_t generation;
2400   kmp_uint32 nelements;
2401   kmp_uint32 nconflicts;
2402 } kmp_dephash_t;
2403 
2404 typedef struct kmp_task_affinity_info {
2405   kmp_intptr_t base_addr;
2406   size_t len;
2407   struct {
2408     bool flag1 : 1;
2409     bool flag2 : 1;
2410     kmp_int32 reserved : 30;
2411   } flags;
2412 } kmp_task_affinity_info_t;
2413 
2414 typedef enum kmp_event_type_t {
2415   KMP_EVENT_UNINITIALIZED = 0,
2416   KMP_EVENT_ALLOW_COMPLETION = 1
2417 } kmp_event_type_t;
2418 
2419 typedef struct {
2420   kmp_event_type_t type;
2421   kmp_tas_lock_t lock;
2422   union {
2423     kmp_task_t *task;
2424   } ed;
2425 } kmp_event_t;
2426 
2427 #ifdef BUILD_TIED_TASK_STACK
2428 
2429 /* Tied Task stack definitions */
2430 typedef struct kmp_stack_block {
2431   kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2432   struct kmp_stack_block *sb_next;
2433   struct kmp_stack_block *sb_prev;
2434 } kmp_stack_block_t;
2435 
2436 typedef struct kmp_task_stack {
2437   kmp_stack_block_t ts_first_block; // first block of stack entries
2438   kmp_taskdata_t **ts_top; // pointer to the top of stack
2439   kmp_int32 ts_entries; // number of entries on the stack
2440 } kmp_task_stack_t;
2441 
2442 #endif // BUILD_TIED_TASK_STACK
2443 
2444 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2445   /* Compiler flags */ /* Total compiler flags must be 16 bits */
2446   unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2447   unsigned final : 1; /* task is final(1) so execute immediately */
2448   unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2449                               code path */
2450   unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2451                                      invoke destructors from the runtime */
2452   unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2453                          context of the RTL) */
2454   unsigned priority_specified : 1; /* set if the compiler provides priority
2455                                       setting for the task */
2456   unsigned detachable : 1; /* 1 == can detach */
2457   unsigned hidden_helper : 1; /* 1 == hidden helper task */
2458   unsigned reserved : 8; /* reserved for compiler use */
2459 
2460   /* Library flags */ /* Total library flags must be 16 bits */
2461   unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2462   unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2463   unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2464   // (1) or may be deferred (0)
2465   unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2466   // (0) [>= 2 threads]
2467   /* If either team_serial or tasking_ser is set, task team may be NULL */
2468   /* Task State Flags: */
2469   unsigned started : 1; /* 1==started, 0==not started     */
2470   unsigned executing : 1; /* 1==executing, 0==not executing */
2471   unsigned complete : 1; /* 1==complete, 0==not complete   */
2472   unsigned freed : 1; /* 1==freed, 0==allocated        */
2473   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2474   unsigned reserved31 : 7; /* reserved for library use */
2475 
2476 } kmp_tasking_flags_t;
2477 
2478 struct kmp_taskdata { /* aligned during dynamic allocation       */
2479   kmp_int32 td_task_id; /* id, assigned by debugger                */
2480   kmp_tasking_flags_t td_flags; /* task flags                              */
2481   kmp_team_t *td_team; /* team for this task                      */
2482   kmp_info_p *td_alloc_thread; /* thread that allocated data structures   */
2483   /* Currently not used except for perhaps IDB */
2484   kmp_taskdata_t *td_parent; /* parent task                             */
2485   kmp_int32 td_level; /* task nesting level                      */
2486   std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2487   ident_t *td_ident; /* task identifier                         */
2488   // Taskwait data.
2489   ident_t *td_taskwait_ident;
2490   kmp_uint32 td_taskwait_counter;
2491   kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2492   KMP_ALIGN_CACHE kmp_internal_control_t
2493       td_icvs; /* Internal control variables for the task */
2494   KMP_ALIGN_CACHE std::atomic<kmp_int32>
2495       td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2496                                    deallocated */
2497   std::atomic<kmp_int32>
2498       td_incomplete_child_tasks; /* Child tasks not yet complete */
2499   kmp_taskgroup_t
2500       *td_taskgroup; // Each task keeps pointer to its current taskgroup
2501   kmp_dephash_t
2502       *td_dephash; // Dependencies for children tasks are tracked from here
2503   kmp_depnode_t
2504       *td_depnode; // Pointer to graph node if this task has dependencies
2505   kmp_task_team_t *td_task_team;
2506   size_t td_size_alloc; // Size of task structure, including shareds etc.
2507 #if defined(KMP_GOMP_COMPAT)
2508   // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2509   kmp_int32 td_size_loop_bounds;
2510 #endif
2511   kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2512 #if defined(KMP_GOMP_COMPAT)
2513   // GOMP sends in a copy function for copy constructors
2514   void (*td_copy_func)(void *, void *);
2515 #endif
2516   kmp_event_t td_allow_completion_event;
2517 #if OMPT_SUPPORT
2518   ompt_task_info_t ompt_task_info;
2519 #endif
2520 }; // struct kmp_taskdata
2521 
2522 // Make sure padding above worked
2523 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2524 
2525 // Data for task team but per thread
2526 typedef struct kmp_base_thread_data {
2527   kmp_info_p *td_thr; // Pointer back to thread info
2528   // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2529   // queued?
2530   kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2531   kmp_taskdata_t *
2532       *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2533   kmp_int32 td_deque_size; // Size of deck
2534   kmp_uint32 td_deque_head; // Head of deque (will wrap)
2535   kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2536   kmp_int32 td_deque_ntasks; // Number of tasks in deque
2537   // GEH: shouldn't this be volatile since used in while-spin?
2538   kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2539 #ifdef BUILD_TIED_TASK_STACK
2540   kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2541 // scheduling constraint
2542 #endif // BUILD_TIED_TASK_STACK
2543 } kmp_base_thread_data_t;
2544 
2545 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2546 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2547 
2548 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2549 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2550 
2551 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2552   kmp_base_thread_data_t td;
2553   double td_align; /* use worst case alignment */
2554   char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2555 } kmp_thread_data_t;
2556 
2557 typedef struct kmp_task_pri {
2558   kmp_thread_data_t td;
2559   kmp_int32 priority;
2560   kmp_task_pri *next;
2561 } kmp_task_pri_t;
2562 
2563 // Data for task teams which are used when tasking is enabled for the team
2564 typedef struct kmp_base_task_team {
2565   kmp_bootstrap_lock_t
2566       tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2567   /* must be bootstrap lock since used at library shutdown*/
2568 
2569   // TODO: check performance vs kmp_tas_lock_t
2570   kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2571   kmp_task_pri_t *tt_task_pri_list;
2572 
2573   kmp_task_team_t *tt_next; /* For linking the task team free list */
2574   kmp_thread_data_t
2575       *tt_threads_data; /* Array of per-thread structures for task team */
2576   /* Data survives task team deallocation */
2577   kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2578                                executing this team? */
2579   /* TRUE means tt_threads_data is set up and initialized */
2580   kmp_int32 tt_nproc; /* #threads in team           */
2581   kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2582   kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2583   kmp_int32 tt_untied_task_encountered;
2584   std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2585   // There is hidden helper thread encountered in this task team so that we must
2586   // wait when waiting on task team
2587   kmp_int32 tt_hidden_helper_task_encountered;
2588 
2589   KMP_ALIGN_CACHE
2590   std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2591 
2592   KMP_ALIGN_CACHE
2593   volatile kmp_uint32
2594       tt_active; /* is the team still actively executing tasks */
2595 } kmp_base_task_team_t;
2596 
2597 union KMP_ALIGN_CACHE kmp_task_team {
2598   kmp_base_task_team_t tt;
2599   double tt_align; /* use worst case alignment */
2600   char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2601 };
2602 
2603 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2604 // Free lists keep same-size free memory slots for fast memory allocation
2605 // routines
2606 typedef struct kmp_free_list {
2607   void *th_free_list_self; // Self-allocated tasks free list
2608   void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2609   // threads
2610   void *th_free_list_other; // Non-self free list (to be returned to owner's
2611   // sync list)
2612 } kmp_free_list_t;
2613 #endif
2614 #if KMP_NESTED_HOT_TEAMS
2615 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2616 // are not put in teams pool, and they don't put threads in threads pool.
2617 typedef struct kmp_hot_team_ptr {
2618   kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2619   kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2620 } kmp_hot_team_ptr_t;
2621 #endif
2622 typedef struct kmp_teams_size {
2623   kmp_int32 nteams; // number of teams in a league
2624   kmp_int32 nth; // number of threads in each team of the league
2625 } kmp_teams_size_t;
2626 
2627 // This struct stores a thread that acts as a "root" for a contention
2628 // group. Contention groups are rooted at kmp_root threads, but also at
2629 // each primary thread of each team created in the teams construct.
2630 // This struct therefore also stores a thread_limit associated with
2631 // that contention group, and a counter to track the number of threads
2632 // active in that contention group. Each thread has a list of these: CG
2633 // root threads have an entry in their list in which cg_root refers to
2634 // the thread itself, whereas other workers in the CG will have a
2635 // single entry where cg_root is same as the entry containing their CG
2636 // root. When a thread encounters a teams construct, it will add a new
2637 // entry to the front of its list, because it now roots a new CG.
2638 typedef struct kmp_cg_root {
2639   kmp_info_p *cg_root; // "root" thread for a contention group
2640   // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2641   // thread_limit clause for teams primary threads
2642   kmp_int32 cg_thread_limit;
2643   kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2644   struct kmp_cg_root *up; // pointer to higher level CG root in list
2645 } kmp_cg_root_t;
2646 
2647 // OpenMP thread data structures
2648 
2649 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2650   /* Start with the readonly data which is cache aligned and padded. This is
2651      written before the thread starts working by the primary thread. Uber
2652      masters may update themselves later. Usage does not consider serialized
2653      regions.  */
2654   kmp_desc_t th_info;
2655   kmp_team_p *th_team; /* team we belong to */
2656   kmp_root_p *th_root; /* pointer to root of task hierarchy */
2657   kmp_info_p *th_next_pool; /* next available thread in the pool */
2658   kmp_disp_t *th_dispatch; /* thread's dispatch data */
2659   int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2660 
2661   /* The following are cached from the team info structure */
2662   /* TODO use these in more places as determined to be needed via profiling */
2663   int th_team_nproc; /* number of threads in a team */
2664   kmp_info_p *th_team_master; /* the team's primary thread */
2665   int th_team_serialized; /* team is serialized */
2666   microtask_t th_teams_microtask; /* save entry address for teams construct */
2667   int th_teams_level; /* save initial level of teams construct */
2668 /* it is 0 on device but may be any on host */
2669 
2670 /* The blocktime info is copied from the team struct to the thread struct */
2671 /* at the start of a barrier, and the values stored in the team are used  */
2672 /* at points in the code where the team struct is no longer guaranteed    */
2673 /* to exist (from the POV of worker threads).                             */
2674 #if KMP_USE_MONITOR
2675   int th_team_bt_intervals;
2676   int th_team_bt_set;
2677 #else
2678   kmp_uint64 th_team_bt_intervals;
2679 #endif
2680 
2681 #if KMP_AFFINITY_SUPPORTED
2682   kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2683 #endif
2684   omp_allocator_handle_t th_def_allocator; /* default allocator */
2685   /* The data set by the primary thread at reinit, then R/W by the worker */
2686   KMP_ALIGN_CACHE int
2687       th_set_nproc; /* if > 0, then only use this request for the next fork */
2688 #if KMP_NESTED_HOT_TEAMS
2689   kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2690 #endif
2691   kmp_proc_bind_t
2692       th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2693   kmp_teams_size_t
2694       th_teams_size; /* number of teams/threads in teams construct */
2695 #if KMP_AFFINITY_SUPPORTED
2696   int th_current_place; /* place currently bound to */
2697   int th_new_place; /* place to bind to in par reg */
2698   int th_first_place; /* first place in partition */
2699   int th_last_place; /* last place in partition */
2700 #endif
2701   int th_prev_level; /* previous level for affinity format */
2702   int th_prev_num_threads; /* previous num_threads for affinity format */
2703 #if USE_ITT_BUILD
2704   kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2705   kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2706   kmp_uint64 th_frame_time; /* frame timestamp */
2707 #endif /* USE_ITT_BUILD */
2708   kmp_local_t th_local;
2709   struct private_common *th_pri_head;
2710 
2711   /* Now the data only used by the worker (after initial allocation) */
2712   /* TODO the first serial team should actually be stored in the info_t
2713      structure.  this will help reduce initial allocation overhead */
2714   KMP_ALIGN_CACHE kmp_team_p
2715       *th_serial_team; /*serialized team held in reserve*/
2716 
2717 #if OMPT_SUPPORT
2718   ompt_thread_info_t ompt_thread_info;
2719 #endif
2720 
2721   /* The following are also read by the primary thread during reinit */
2722   struct common_table *th_pri_common;
2723 
2724   volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2725   /* while awaiting queuing lock acquire */
2726 
2727   volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2728   flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2729 
2730   ident_t *th_ident;
2731   unsigned th_x; // Random number generator data
2732   unsigned th_a; // Random number generator data
2733 
2734   /* Tasking-related data for the thread */
2735   kmp_task_team_t *th_task_team; // Task team struct
2736   kmp_taskdata_t *th_current_task; // Innermost Task being executed
2737   kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2738   kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2739   // at nested levels
2740   kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2741   kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2742   kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2743   // tasking, thus safe to reap
2744 
2745   /* More stuff for keeping track of active/sleeping threads (this part is
2746      written by the worker thread) */
2747   kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2748   int th_active; // ! sleeping; 32 bits for TCR/TCW
2749   std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2750   // 0 = not used in team; 1 = used in team;
2751   // 2 = transitioning to not used in team; 3 = transitioning to used in team
2752   struct cons_header *th_cons; // used for consistency check
2753 #if KMP_USE_HIER_SCHED
2754   // used for hierarchical scheduling
2755   kmp_hier_private_bdata_t *th_hier_bar_data;
2756 #endif
2757 
2758   /* Add the syncronizing data which is cache aligned and padded. */
2759   KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2760 
2761   KMP_ALIGN_CACHE volatile kmp_int32
2762       th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2763 
2764 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2765 #define NUM_LISTS 4
2766   kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2767 // allocation routines
2768 #endif
2769 
2770 #if KMP_OS_WINDOWS
2771   kmp_win32_cond_t th_suspend_cv;
2772   kmp_win32_mutex_t th_suspend_mx;
2773   std::atomic<int> th_suspend_init;
2774 #endif
2775 #if KMP_OS_UNIX
2776   kmp_cond_align_t th_suspend_cv;
2777   kmp_mutex_align_t th_suspend_mx;
2778   std::atomic<int> th_suspend_init_count;
2779 #endif
2780 
2781 #if USE_ITT_BUILD
2782   kmp_itt_mark_t th_itt_mark_single;
2783 // alignment ???
2784 #endif /* USE_ITT_BUILD */
2785 #if KMP_STATS_ENABLED
2786   kmp_stats_list *th_stats;
2787 #endif
2788 #if KMP_OS_UNIX
2789   std::atomic<bool> th_blocking;
2790 #endif
2791   kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2792 } kmp_base_info_t;
2793 
2794 typedef union KMP_ALIGN_CACHE kmp_info {
2795   double th_align; /* use worst case alignment */
2796   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2797   kmp_base_info_t th;
2798 } kmp_info_t;
2799 
2800 // OpenMP thread team data structures
2801 
2802 typedef struct kmp_base_data {
2803   volatile kmp_uint32 t_value;
2804 } kmp_base_data_t;
2805 
2806 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2807   double dt_align; /* use worst case alignment */
2808   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2809   kmp_base_data_t dt;
2810 } kmp_sleep_team_t;
2811 
2812 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2813   double dt_align; /* use worst case alignment */
2814   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2815   kmp_base_data_t dt;
2816 } kmp_ordered_team_t;
2817 
2818 typedef int (*launch_t)(int gtid);
2819 
2820 /* Minimum number of ARGV entries to malloc if necessary */
2821 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2822 
2823 // Set up how many argv pointers will fit in cache lines containing
2824 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2825 // larger value for more space between the primary write/worker read section and
2826 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2827 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2828 #define KMP_INLINE_ARGV_BYTES                                                  \
2829   (4 * CACHE_LINE -                                                            \
2830    ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) +               \
2831      sizeof(kmp_int16) + sizeof(kmp_uint32)) %                                 \
2832     CACHE_LINE))
2833 #else
2834 #define KMP_INLINE_ARGV_BYTES                                                  \
2835   (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2836 #endif
2837 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2838 
2839 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2840   // Synchronization Data
2841   // ---------------------------------------------------------------------------
2842   KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2843   kmp_balign_team_t t_bar[bs_last_barrier];
2844   std::atomic<int> t_construct; // count of single directive encountered by team
2845   char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2846 
2847   // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2848   std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2849   std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2850 
2851   // Primary thread only
2852   // ---------------------------------------------------------------------------
2853   KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2854   int t_master_this_cons; // "this_construct" single counter of primary thread
2855   // in parent team
2856   ident_t *t_ident; // if volatile, have to change too much other crud to
2857   // volatile too
2858   kmp_team_p *t_parent; // parent team
2859   kmp_team_p *t_next_pool; // next free team in the team pool
2860   kmp_disp_t *t_dispatch; // thread's dispatch data
2861   kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2862   kmp_proc_bind_t t_proc_bind; // bind type for par region
2863 #if USE_ITT_BUILD
2864   kmp_uint64 t_region_time; // region begin timestamp
2865 #endif /* USE_ITT_BUILD */
2866 
2867   // Primary thread write, workers read
2868   // --------------------------------------------------------------------------
2869   KMP_ALIGN_CACHE void **t_argv;
2870   int t_argc;
2871   int t_nproc; // number of threads in team
2872   microtask_t t_pkfn;
2873   launch_t t_invoke; // procedure to launch the microtask
2874 
2875 #if OMPT_SUPPORT
2876   ompt_team_info_t ompt_team_info;
2877   ompt_lw_taskteam_t *ompt_serialized_team_info;
2878 #endif
2879 
2880 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2881   kmp_int8 t_fp_control_saved;
2882   kmp_int8 t_pad2b;
2883   kmp_int16 t_x87_fpu_control_word; // FP control regs
2884   kmp_uint32 t_mxcsr;
2885 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2886 
2887   void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2888 
2889   KMP_ALIGN_CACHE kmp_info_t **t_threads;
2890   kmp_taskdata_t
2891       *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2892   int t_level; // nested parallel level
2893 
2894   KMP_ALIGN_CACHE int t_max_argc;
2895   int t_max_nproc; // max threads this team can handle (dynamically expandable)
2896   int t_serialized; // levels deep of serialized teams
2897   dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2898   int t_id; // team's id, assigned by debugger.
2899   int t_active_level; // nested active parallel level
2900   kmp_r_sched_t t_sched; // run-time schedule for the team
2901 #if KMP_AFFINITY_SUPPORTED
2902   int t_first_place; // first & last place in parent thread's partition.
2903   int t_last_place; // Restore these values to primary thread after par region.
2904 #endif // KMP_AFFINITY_SUPPORTED
2905   int t_display_affinity;
2906   int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2907   // omp_set_num_threads() call
2908   omp_allocator_handle_t t_def_allocator; /* default allocator */
2909 
2910 // Read/write by workers as well
2911 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2912   // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2913   // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2914   // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2915   // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2916   char dummy_padding[1024];
2917 #endif
2918   // Internal control stack for additional nested teams.
2919   KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2920   // for SERIALIZED teams nested 2 or more levels deep
2921   // typed flag to store request state of cancellation
2922   std::atomic<kmp_int32> t_cancel_request;
2923   int t_master_active; // save on fork, restore on join
2924   void *t_copypriv_data; // team specific pointer to copyprivate data array
2925 #if KMP_OS_WINDOWS
2926   std::atomic<kmp_uint32> t_copyin_counter;
2927 #endif
2928 #if USE_ITT_BUILD
2929   void *t_stack_id; // team specific stack stitching id (for ittnotify)
2930 #endif /* USE_ITT_BUILD */
2931   distributedBarrier *b; // Distributed barrier data associated with team
2932 } kmp_base_team_t;
2933 
2934 union KMP_ALIGN_CACHE kmp_team {
2935   kmp_base_team_t t;
2936   double t_align; /* use worst case alignment */
2937   char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2938 };
2939 
2940 typedef union KMP_ALIGN_CACHE kmp_time_global {
2941   double dt_align; /* use worst case alignment */
2942   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2943   kmp_base_data_t dt;
2944 } kmp_time_global_t;
2945 
2946 typedef struct kmp_base_global {
2947   /* cache-aligned */
2948   kmp_time_global_t g_time;
2949 
2950   /* non cache-aligned */
2951   volatile int g_abort;
2952   volatile int g_done;
2953 
2954   int g_dynamic;
2955   enum dynamic_mode g_dynamic_mode;
2956 } kmp_base_global_t;
2957 
2958 typedef union KMP_ALIGN_CACHE kmp_global {
2959   kmp_base_global_t g;
2960   double g_align; /* use worst case alignment */
2961   char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2962 } kmp_global_t;
2963 
2964 typedef struct kmp_base_root {
2965   // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2966   // (r_in_parallel>= 0)
2967   // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2968   // the synch overhead or keeping r_active
2969   volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2970   // keeps a count of active parallel regions per root
2971   std::atomic<int> r_in_parallel;
2972   // GEH: This is misnamed, should be r_active_levels
2973   kmp_team_t *r_root_team;
2974   kmp_team_t *r_hot_team;
2975   kmp_info_t *r_uber_thread;
2976   kmp_lock_t r_begin_lock;
2977   volatile int r_begin;
2978   int r_blocktime; /* blocktime for this root and descendants */
2979 #if KMP_AFFINITY_SUPPORTED
2980   int r_affinity_assigned;
2981 #endif // KMP_AFFINITY_SUPPORTED
2982 } kmp_base_root_t;
2983 
2984 typedef union KMP_ALIGN_CACHE kmp_root {
2985   kmp_base_root_t r;
2986   double r_align; /* use worst case alignment */
2987   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2988 } kmp_root_t;
2989 
2990 struct fortran_inx_info {
2991   kmp_int32 data;
2992 };
2993 
2994 // This list type exists to hold old __kmp_threads arrays so that
2995 // old references to them may complete while reallocation takes place when
2996 // expanding the array. The items in this list are kept alive until library
2997 // shutdown.
2998 typedef struct kmp_old_threads_list_t {
2999   kmp_info_t **threads;
3000   struct kmp_old_threads_list_t *next;
3001 } kmp_old_threads_list_t;
3002 
3003 /* ------------------------------------------------------------------------ */
3004 
3005 extern int __kmp_settings;
3006 extern int __kmp_duplicate_library_ok;
3007 #if USE_ITT_BUILD
3008 extern int __kmp_forkjoin_frames;
3009 extern int __kmp_forkjoin_frames_mode;
3010 #endif
3011 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3012 extern int __kmp_determ_red;
3013 
3014 #ifdef KMP_DEBUG
3015 extern int kmp_a_debug;
3016 extern int kmp_b_debug;
3017 extern int kmp_c_debug;
3018 extern int kmp_d_debug;
3019 extern int kmp_e_debug;
3020 extern int kmp_f_debug;
3021 #endif /* KMP_DEBUG */
3022 
3023 /* For debug information logging using rotating buffer */
3024 #define KMP_DEBUG_BUF_LINES_INIT 512
3025 #define KMP_DEBUG_BUF_LINES_MIN 1
3026 
3027 #define KMP_DEBUG_BUF_CHARS_INIT 128
3028 #define KMP_DEBUG_BUF_CHARS_MIN 2
3029 
3030 extern int
3031     __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3032 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3033 extern int
3034     __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3035 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3036                                       entry pointer */
3037 
3038 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3039 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3040                                               printed in buffer so far */
3041 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3042                                           recommended in warnings */
3043 /* end rotating debug buffer */
3044 
3045 #ifdef KMP_DEBUG
3046 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3047 
3048 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3049 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3050 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3051 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3052 extern int __kmp_par_range_lb;
3053 extern int __kmp_par_range_ub;
3054 #endif
3055 
3056 /* For printing out dynamic storage map for threads and teams */
3057 extern int
3058     __kmp_storage_map; /* True means print storage map for threads and teams */
3059 extern int __kmp_storage_map_verbose; /* True means storage map includes
3060                                          placement info */
3061 extern int __kmp_storage_map_verbose_specified;
3062 
3063 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3064 extern kmp_cpuinfo_t __kmp_cpuinfo;
3065 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3066 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3067 static inline bool __kmp_is_hybrid_cpu() { return true; }
3068 #else
3069 static inline bool __kmp_is_hybrid_cpu() { return false; }
3070 #endif
3071 
3072 extern volatile int __kmp_init_serial;
3073 extern volatile int __kmp_init_gtid;
3074 extern volatile int __kmp_init_common;
3075 extern volatile int __kmp_need_register_serial;
3076 extern volatile int __kmp_init_middle;
3077 extern volatile int __kmp_init_parallel;
3078 #if KMP_USE_MONITOR
3079 extern volatile int __kmp_init_monitor;
3080 #endif
3081 extern volatile int __kmp_init_user_locks;
3082 extern volatile int __kmp_init_hidden_helper_threads;
3083 extern int __kmp_init_counter;
3084 extern int __kmp_root_counter;
3085 extern int __kmp_version;
3086 
3087 /* list of address of allocated caches for commons */
3088 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3089 
3090 /* Barrier algorithm types and options */
3091 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3092 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3093 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3094 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3095 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3096 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3097 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3098 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3099 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3100 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3101 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3102 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3103 
3104 /* Global Locks */
3105 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3106 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3107 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3108 extern kmp_bootstrap_lock_t
3109     __kmp_exit_lock; /* exit() is not always thread-safe */
3110 #if KMP_USE_MONITOR
3111 extern kmp_bootstrap_lock_t
3112     __kmp_monitor_lock; /* control monitor thread creation */
3113 #endif
3114 extern kmp_bootstrap_lock_t
3115     __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3116                              __kmp_threads expansion to co-exist */
3117 
3118 extern kmp_lock_t __kmp_global_lock; /* control OS/global access  */
3119 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access  */
3120 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3121 
3122 extern enum library_type __kmp_library;
3123 
3124 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3125 extern enum sched_type __kmp_static; /* default static scheduling method */
3126 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3127 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3128 extern int __kmp_chunk; /* default runtime chunk size */
3129 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3130 
3131 extern size_t __kmp_stksize; /* stack size per thread         */
3132 #if KMP_USE_MONITOR
3133 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3134 #endif
3135 extern size_t __kmp_stkoffset; /* stack offset per thread       */
3136 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3137 
3138 extern size_t
3139     __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3140 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3141 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3142 extern int __kmp_env_checks; /* was KMP_CHECKS specified?    */
3143 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3144 extern int __kmp_generate_warnings; /* should we issue warnings? */
3145 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3146 
3147 #ifdef DEBUG_SUSPEND
3148 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3149 #endif
3150 
3151 extern kmp_int32 __kmp_use_yield;
3152 extern kmp_int32 __kmp_use_yield_exp_set;
3153 extern kmp_uint32 __kmp_yield_init;
3154 extern kmp_uint32 __kmp_yield_next;
3155 extern kmp_uint64 __kmp_pause_init;
3156 
3157 /* ------------------------------------------------------------------------- */
3158 extern int __kmp_allThreadsSpecified;
3159 
3160 extern size_t __kmp_align_alloc;
3161 /* following data protected by initialization routines */
3162 extern int __kmp_xproc; /* number of processors in the system */
3163 extern int __kmp_avail_proc; /* number of processors available to the process */
3164 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3165 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3166 // maximum total number of concurrently-existing threads on device
3167 extern int __kmp_max_nth;
3168 // maximum total number of concurrently-existing threads in a contention group
3169 extern int __kmp_cg_max_nth;
3170 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3171 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3172                                       __kmp_root */
3173 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3174                                    region a la OMP_NUM_THREADS */
3175 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3176                                       initialization */
3177 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3178                                  used (fixed) */
3179 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3180                                (__kmpc_threadprivate_cached()) */
3181 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3182                                     blocking (env setting) */
3183 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3184 #if KMP_USE_MONITOR
3185 extern int
3186     __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3187 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3188                                   blocking */
3189 #endif
3190 #ifdef KMP_ADJUST_BLOCKTIME
3191 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3192 #endif /* KMP_ADJUST_BLOCKTIME */
3193 #ifdef KMP_DFLT_NTH_CORES
3194 extern int __kmp_ncores; /* Total number of cores for threads placement */
3195 #endif
3196 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3197 extern int __kmp_abort_delay;
3198 
3199 extern int __kmp_need_register_atfork_specified;
3200 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3201                                           to install fork handler */
3202 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3203                                0 - not set, will be set at runtime
3204                                1 - using stack search
3205                                2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3206                                    X*) or TlsGetValue(Windows* OS))
3207                                3 - static TLS (__declspec(thread) __kmp_gtid),
3208                                    Linux* OS .so only.  */
3209 extern int
3210     __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3211 #ifdef KMP_TDATA_GTID
3212 extern KMP_THREAD_LOCAL int __kmp_gtid;
3213 #endif
3214 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3215 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3216 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3217 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3218 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3219 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3220 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3221 
3222 // max_active_levels for nested parallelism enabled by default via
3223 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3224 extern int __kmp_dflt_max_active_levels;
3225 // Indicates whether value of __kmp_dflt_max_active_levels was already
3226 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3227 extern bool __kmp_dflt_max_active_levels_set;
3228 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3229                                           concurrent execution per team */
3230 #if KMP_NESTED_HOT_TEAMS
3231 extern int __kmp_hot_teams_mode;
3232 extern int __kmp_hot_teams_max_level;
3233 #endif
3234 
3235 #if KMP_OS_LINUX
3236 extern enum clock_function_type __kmp_clock_function;
3237 extern int __kmp_clock_function_param;
3238 #endif /* KMP_OS_LINUX */
3239 
3240 #if KMP_MIC_SUPPORTED
3241 extern enum mic_type __kmp_mic_type;
3242 #endif
3243 
3244 #ifdef USE_LOAD_BALANCE
3245 extern double __kmp_load_balance_interval; // load balance algorithm interval
3246 #endif /* USE_LOAD_BALANCE */
3247 
3248 // OpenMP 3.1 - Nested num threads array
3249 typedef struct kmp_nested_nthreads_t {
3250   int *nth;
3251   int size;
3252   int used;
3253 } kmp_nested_nthreads_t;
3254 
3255 extern kmp_nested_nthreads_t __kmp_nested_nth;
3256 
3257 #if KMP_USE_ADAPTIVE_LOCKS
3258 
3259 // Parameters for the speculative lock backoff system.
3260 struct kmp_adaptive_backoff_params_t {
3261   // Number of soft retries before it counts as a hard retry.
3262   kmp_uint32 max_soft_retries;
3263   // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3264   // the right
3265   kmp_uint32 max_badness;
3266 };
3267 
3268 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3269 
3270 #if KMP_DEBUG_ADAPTIVE_LOCKS
3271 extern const char *__kmp_speculative_statsfile;
3272 #endif
3273 
3274 #endif // KMP_USE_ADAPTIVE_LOCKS
3275 
3276 extern int __kmp_display_env; /* TRUE or FALSE */
3277 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3278 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3279 extern int __kmp_nteams;
3280 extern int __kmp_teams_thread_limit;
3281 
3282 /* ------------------------------------------------------------------------- */
3283 
3284 /* the following are protected by the fork/join lock */
3285 /* write: lock  read: anytime */
3286 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3287 /* Holds old arrays of __kmp_threads until library shutdown */
3288 extern kmp_old_threads_list_t *__kmp_old_threads_list;
3289 /* read/write: lock */
3290 extern volatile kmp_team_t *__kmp_team_pool;
3291 extern volatile kmp_info_t *__kmp_thread_pool;
3292 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3293 
3294 // total num threads reachable from some root thread including all root threads
3295 extern volatile int __kmp_nth;
3296 /* total number of threads reachable from some root thread including all root
3297    threads, and those in the thread pool */
3298 extern volatile int __kmp_all_nth;
3299 extern std::atomic<int> __kmp_thread_pool_active_nth;
3300 
3301 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3302 /* end data protected by fork/join lock */
3303 /* ------------------------------------------------------------------------- */
3304 
3305 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3306 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3307 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3308 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3309 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3310 
3311 // AT: Which way is correct?
3312 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3313 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3314 #define __kmp_get_team_num_threads(gtid)                                       \
3315   (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3316 
3317 static inline bool KMP_UBER_GTID(int gtid) {
3318   KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3319   KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3320   return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3321           __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3322 }
3323 
3324 static inline int __kmp_tid_from_gtid(int gtid) {
3325   KMP_DEBUG_ASSERT(gtid >= 0);
3326   return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3327 }
3328 
3329 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3330   KMP_DEBUG_ASSERT(tid >= 0 && team);
3331   return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3332 }
3333 
3334 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3335   KMP_DEBUG_ASSERT(thr);
3336   return thr->th.th_info.ds.ds_gtid;
3337 }
3338 
3339 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3340   KMP_DEBUG_ASSERT(gtid >= 0);
3341   return __kmp_threads[gtid];
3342 }
3343 
3344 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3345   KMP_DEBUG_ASSERT(gtid >= 0);
3346   return __kmp_threads[gtid]->th.th_team;
3347 }
3348 
3349 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3350   if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3351     KMP_FATAL(ThreadIdentInvalid);
3352 }
3353 
3354 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3355 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3356 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3357 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3358 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3359 #endif
3360 
3361 #if KMP_HAVE_UMWAIT
3362 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3363 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3364 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3365 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3366 #endif
3367 
3368 /* ------------------------------------------------------------------------- */
3369 
3370 extern kmp_global_t __kmp_global; /* global status */
3371 
3372 extern kmp_info_t __kmp_monitor;
3373 // For Debugging Support Library
3374 extern std::atomic<kmp_int32> __kmp_team_counter;
3375 // For Debugging Support Library
3376 extern std::atomic<kmp_int32> __kmp_task_counter;
3377 
3378 #if USE_DEBUGGER
3379 #define _KMP_GEN_ID(counter)                                                   \
3380   (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3381 #else
3382 #define _KMP_GEN_ID(counter) (~0)
3383 #endif /* USE_DEBUGGER */
3384 
3385 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3386 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3387 
3388 /* ------------------------------------------------------------------------ */
3389 
3390 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3391                                          size_t size, char const *format, ...);
3392 
3393 extern void __kmp_serial_initialize(void);
3394 extern void __kmp_middle_initialize(void);
3395 extern void __kmp_parallel_initialize(void);
3396 
3397 extern void __kmp_internal_begin(void);
3398 extern void __kmp_internal_end_library(int gtid);
3399 extern void __kmp_internal_end_thread(int gtid);
3400 extern void __kmp_internal_end_atexit(void);
3401 extern void __kmp_internal_end_dtor(void);
3402 extern void __kmp_internal_end_dest(void *);
3403 
3404 extern int __kmp_register_root(int initial_thread);
3405 extern void __kmp_unregister_root(int gtid);
3406 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3407 
3408 extern int __kmp_ignore_mppbeg(void);
3409 extern int __kmp_ignore_mppend(void);
3410 
3411 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3412 extern void __kmp_exit_single(int gtid);
3413 
3414 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3415 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3416 
3417 #ifdef USE_LOAD_BALANCE
3418 extern int __kmp_get_load_balance(int);
3419 #endif
3420 
3421 extern int __kmp_get_global_thread_id(void);
3422 extern int __kmp_get_global_thread_id_reg(void);
3423 extern void __kmp_exit_thread(int exit_status);
3424 extern void __kmp_abort(char const *format, ...);
3425 extern void __kmp_abort_thread(void);
3426 KMP_NORETURN extern void __kmp_abort_process(void);
3427 extern void __kmp_warn(char const *format, ...);
3428 
3429 extern void __kmp_set_num_threads(int new_nth, int gtid);
3430 
3431 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3432 // registered.
3433 static inline kmp_info_t *__kmp_entry_thread() {
3434   int gtid = __kmp_entry_gtid();
3435 
3436   return __kmp_threads[gtid];
3437 }
3438 
3439 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3440 extern int __kmp_get_max_active_levels(int gtid);
3441 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3442 extern int __kmp_get_team_size(int gtid, int level);
3443 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3444 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3445 
3446 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3447 extern void __kmp_init_random(kmp_info_t *thread);
3448 
3449 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3450 extern void __kmp_adjust_num_threads(int new_nproc);
3451 extern void __kmp_check_stksize(size_t *val);
3452 
3453 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3454 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3455 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3456 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3457 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3458 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3459 
3460 #if USE_FAST_MEMORY
3461 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3462                                   size_t size KMP_SRC_LOC_DECL);
3463 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3464 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3465 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3466 #define __kmp_fast_allocate(this_thr, size)                                    \
3467   ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3468 #define __kmp_fast_free(this_thr, ptr)                                         \
3469   ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3470 #endif
3471 
3472 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3473 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3474                                   size_t elsize KMP_SRC_LOC_DECL);
3475 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3476                                    size_t size KMP_SRC_LOC_DECL);
3477 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3478 #define __kmp_thread_malloc(th, size)                                          \
3479   ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3480 #define __kmp_thread_calloc(th, nelem, elsize)                                 \
3481   ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3482 #define __kmp_thread_realloc(th, ptr, size)                                    \
3483   ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3484 #define __kmp_thread_free(th, ptr)                                             \
3485   ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3486 
3487 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3488 
3489 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3490                                  kmp_proc_bind_t proc_bind);
3491 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3492                                  int num_threads);
3493 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3494                                     int num_teams_ub, int num_threads);
3495 
3496 extern void __kmp_yield();
3497 
3498 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3499                                    enum sched_type schedule, kmp_int32 lb,
3500                                    kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3501 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3502                                     enum sched_type schedule, kmp_uint32 lb,
3503                                     kmp_uint32 ub, kmp_int32 st,
3504                                     kmp_int32 chunk);
3505 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3506                                    enum sched_type schedule, kmp_int64 lb,
3507                                    kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3508 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3509                                     enum sched_type schedule, kmp_uint64 lb,
3510                                     kmp_uint64 ub, kmp_int64 st,
3511                                     kmp_int64 chunk);
3512 
3513 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3514                                   kmp_int32 *p_last, kmp_int32 *p_lb,
3515                                   kmp_int32 *p_ub, kmp_int32 *p_st);
3516 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3517                                    kmp_int32 *p_last, kmp_uint32 *p_lb,
3518                                    kmp_uint32 *p_ub, kmp_int32 *p_st);
3519 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3520                                   kmp_int32 *p_last, kmp_int64 *p_lb,
3521                                   kmp_int64 *p_ub, kmp_int64 *p_st);
3522 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3523                                    kmp_int32 *p_last, kmp_uint64 *p_lb,
3524                                    kmp_uint64 *p_ub, kmp_int64 *p_st);
3525 
3526 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3527 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3528 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3529 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3530 
3531 #ifdef KMP_GOMP_COMPAT
3532 
3533 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3534                                       enum sched_type schedule, kmp_int32 lb,
3535                                       kmp_int32 ub, kmp_int32 st,
3536                                       kmp_int32 chunk, int push_ws);
3537 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3538                                        enum sched_type schedule, kmp_uint32 lb,
3539                                        kmp_uint32 ub, kmp_int32 st,
3540                                        kmp_int32 chunk, int push_ws);
3541 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3542                                       enum sched_type schedule, kmp_int64 lb,
3543                                       kmp_int64 ub, kmp_int64 st,
3544                                       kmp_int64 chunk, int push_ws);
3545 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3546                                        enum sched_type schedule, kmp_uint64 lb,
3547                                        kmp_uint64 ub, kmp_int64 st,
3548                                        kmp_int64 chunk, int push_ws);
3549 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3550 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3551 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3552 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3553 
3554 #endif /* KMP_GOMP_COMPAT */
3555 
3556 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3557 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3558 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3559 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3560 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3561 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3562                                kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3563                                void *obj);
3564 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3565                              kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3566 
3567 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3568                           int final_spin
3569 #if USE_ITT_BUILD
3570                           ,
3571                           void *itt_sync_obj
3572 #endif
3573 );
3574 extern void __kmp_release_64(kmp_flag_64<> *flag);
3575 
3576 extern void __kmp_infinite_loop(void);
3577 
3578 extern void __kmp_cleanup(void);
3579 
3580 #if KMP_HANDLE_SIGNALS
3581 extern int __kmp_handle_signals;
3582 extern void __kmp_install_signals(int parallel_init);
3583 extern void __kmp_remove_signals(void);
3584 #endif
3585 
3586 extern void __kmp_clear_system_time(void);
3587 extern void __kmp_read_system_time(double *delta);
3588 
3589 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3590 
3591 extern void __kmp_expand_host_name(char *buffer, size_t size);
3592 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3593 
3594 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64)
3595 extern void
3596 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3597 #endif
3598 
3599 extern void
3600 __kmp_runtime_initialize(void); /* machine specific initialization */
3601 extern void __kmp_runtime_destroy(void);
3602 
3603 #if KMP_AFFINITY_SUPPORTED
3604 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3605                                        kmp_affin_mask_t *mask);
3606 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3607                                                   kmp_affin_mask_t *mask);
3608 extern void __kmp_affinity_initialize(void);
3609 extern void __kmp_affinity_uninitialize(void);
3610 extern void __kmp_affinity_set_init_mask(
3611     int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3612 extern void __kmp_affinity_set_place(int gtid);
3613 extern void __kmp_affinity_determine_capable(const char *env_var);
3614 extern int __kmp_aux_set_affinity(void **mask);
3615 extern int __kmp_aux_get_affinity(void **mask);
3616 extern int __kmp_aux_get_affinity_max_proc();
3617 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3618 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3619 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3620 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3621 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3622 extern int kmp_set_thread_affinity_mask_initial(void);
3623 #endif
3624 static inline void __kmp_assign_root_init_mask() {
3625   int gtid = __kmp_entry_gtid();
3626   kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3627   if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3628     __kmp_affinity_set_init_mask(gtid, TRUE);
3629     r->r.r_affinity_assigned = TRUE;
3630   }
3631 }
3632 static inline void __kmp_reset_root_init_mask(int gtid) {
3633   kmp_info_t *th = __kmp_threads[gtid];
3634   kmp_root_t *r = th->th.th_root;
3635   if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3636     __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3637     KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3638     r->r.r_affinity_assigned = FALSE;
3639   }
3640 }
3641 #else /* KMP_AFFINITY_SUPPORTED */
3642 #define __kmp_assign_root_init_mask() /* Nothing */
3643 static inline void __kmp_reset_root_init_mask(int gtid) {}
3644 #endif /* KMP_AFFINITY_SUPPORTED */
3645 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3646 // format string is for affinity, so platforms that do not support
3647 // affinity can still use the other fields, e.g., %n for num_threads
3648 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3649                                          kmp_str_buf_t *buffer);
3650 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3651 
3652 extern void __kmp_cleanup_hierarchy();
3653 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3654 
3655 #if KMP_USE_FUTEX
3656 
3657 extern int __kmp_futex_determine_capable(void);
3658 
3659 #endif // KMP_USE_FUTEX
3660 
3661 extern void __kmp_gtid_set_specific(int gtid);
3662 extern int __kmp_gtid_get_specific(void);
3663 
3664 extern double __kmp_read_cpu_time(void);
3665 
3666 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3667 
3668 #if KMP_USE_MONITOR
3669 extern void __kmp_create_monitor(kmp_info_t *th);
3670 #endif
3671 
3672 extern void *__kmp_launch_thread(kmp_info_t *thr);
3673 
3674 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3675 
3676 #if KMP_OS_WINDOWS
3677 extern int __kmp_still_running(kmp_info_t *th);
3678 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3679 extern void __kmp_free_handle(kmp_thread_t tHandle);
3680 #endif
3681 
3682 #if KMP_USE_MONITOR
3683 extern void __kmp_reap_monitor(kmp_info_t *th);
3684 #endif
3685 extern void __kmp_reap_worker(kmp_info_t *th);
3686 extern void __kmp_terminate_thread(int gtid);
3687 
3688 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3689 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3690 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3691 
3692 extern void __kmp_elapsed(double *);
3693 extern void __kmp_elapsed_tick(double *);
3694 
3695 extern void __kmp_enable(int old_state);
3696 extern void __kmp_disable(int *old_state);
3697 
3698 extern void __kmp_thread_sleep(int millis);
3699 
3700 extern void __kmp_common_initialize(void);
3701 extern void __kmp_common_destroy(void);
3702 extern void __kmp_common_destroy_gtid(int gtid);
3703 
3704 #if KMP_OS_UNIX
3705 extern void __kmp_register_atfork(void);
3706 #endif
3707 extern void __kmp_suspend_initialize(void);
3708 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3709 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3710 
3711 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3712                                          int tid);
3713 extern kmp_team_t *
3714 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3715 #if OMPT_SUPPORT
3716                     ompt_data_t ompt_parallel_data,
3717 #endif
3718                     kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3719                     int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3720 extern void __kmp_free_thread(kmp_info_t *);
3721 extern void __kmp_free_team(kmp_root_t *,
3722                             kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3723 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3724 
3725 /* ------------------------------------------------------------------------ */
3726 
3727 extern void __kmp_initialize_bget(kmp_info_t *th);
3728 extern void __kmp_finalize_bget(kmp_info_t *th);
3729 
3730 KMP_EXPORT void *kmpc_malloc(size_t size);
3731 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3732 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3733 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3734 KMP_EXPORT void kmpc_free(void *ptr);
3735 
3736 /* declarations for internal use */
3737 
3738 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3739                          size_t reduce_size, void *reduce_data,
3740                          void (*reduce)(void *, void *));
3741 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3742 extern int __kmp_barrier_gomp_cancel(int gtid);
3743 
3744 /*!
3745  * Tell the fork call which compiler generated the fork call, and therefore how
3746  * to deal with the call.
3747  */
3748 enum fork_context_e {
3749   fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3750                        microtask internally. */
3751   fork_context_intel, /**< Called from Intel generated code.  */
3752   fork_context_last
3753 };
3754 extern int __kmp_fork_call(ident_t *loc, int gtid,
3755                            enum fork_context_e fork_context, kmp_int32 argc,
3756                            microtask_t microtask, launch_t invoker,
3757                            kmp_va_list ap);
3758 
3759 extern void __kmp_join_call(ident_t *loc, int gtid
3760 #if OMPT_SUPPORT
3761                             ,
3762                             enum fork_context_e fork_context
3763 #endif
3764                             ,
3765                             int exit_teams = 0);
3766 
3767 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3768 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3769 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3770 extern int __kmp_invoke_task_func(int gtid);
3771 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3772                                           kmp_info_t *this_thr,
3773                                           kmp_team_t *team);
3774 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3775                                          kmp_info_t *this_thr,
3776                                          kmp_team_t *team);
3777 
3778 // should never have been exported
3779 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3780 extern int __kmp_invoke_teams_master(int gtid);
3781 extern void __kmp_teams_master(int gtid);
3782 extern int __kmp_aux_get_team_num();
3783 extern int __kmp_aux_get_num_teams();
3784 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3785 extern void __kmp_user_set_library(enum library_type arg);
3786 extern void __kmp_aux_set_library(enum library_type arg);
3787 extern void __kmp_aux_set_stacksize(size_t arg);
3788 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3789 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3790 
3791 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3792 void kmpc_set_blocktime(int arg);
3793 void ompc_set_nested(int flag);
3794 void ompc_set_dynamic(int flag);
3795 void ompc_set_num_threads(int arg);
3796 
3797 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3798                                               kmp_team_t *team, int tid);
3799 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3800 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3801                                     kmp_tasking_flags_t *flags,
3802                                     size_t sizeof_kmp_task_t,
3803                                     size_t sizeof_shareds,
3804                                     kmp_routine_entry_t task_entry);
3805 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3806                                      kmp_team_t *team, int tid,
3807                                      int set_curr_task);
3808 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3809 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3810 
3811 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3812                                                        int gtid,
3813                                                        kmp_task_t *task);
3814 extern void __kmp_fulfill_event(kmp_event_t *event);
3815 
3816 extern void __kmp_free_task_team(kmp_info_t *thread,
3817                                  kmp_task_team_t *task_team);
3818 extern void __kmp_reap_task_teams(void);
3819 extern void __kmp_wait_to_unref_task_teams(void);
3820 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3821                                   int always);
3822 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3823 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3824 #if USE_ITT_BUILD
3825                                  ,
3826                                  void *itt_sync_obj
3827 #endif /* USE_ITT_BUILD */
3828                                  ,
3829                                  int wait = 1);
3830 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3831                                   int gtid);
3832 
3833 extern int __kmp_is_address_mapped(void *addr);
3834 extern kmp_uint64 __kmp_hardware_timestamp(void);
3835 
3836 #if KMP_OS_UNIX
3837 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3838 #endif
3839 
3840 /* ------------------------------------------------------------------------ */
3841 //
3842 // Assembly routines that have no compiler intrinsic replacement
3843 //
3844 
3845 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3846                                   void *argv[]
3847 #if OMPT_SUPPORT
3848                                   ,
3849                                   void **exit_frame_ptr
3850 #endif
3851 );
3852 
3853 /* ------------------------------------------------------------------------ */
3854 
3855 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3856 KMP_EXPORT void __kmpc_end(ident_t *);
3857 
3858 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3859                                                   kmpc_ctor_vec ctor,
3860                                                   kmpc_cctor_vec cctor,
3861                                                   kmpc_dtor_vec dtor,
3862                                                   size_t vector_length);
3863 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3864                                               kmpc_ctor ctor, kmpc_cctor cctor,
3865                                               kmpc_dtor dtor);
3866 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3867                                       void *data, size_t size);
3868 
3869 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3870 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3871 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3872 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3873 
3874 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3875 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3876                                  kmpc_micro microtask, ...);
3877 
3878 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3879 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3880 
3881 KMP_EXPORT void __kmpc_flush(ident_t *);
3882 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3883 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3884 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3885 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3886                                    kmp_int32 filter);
3887 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3888 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3889 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3890 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3891                                 kmp_critical_name *);
3892 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3893                                     kmp_critical_name *);
3894 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3895                                           kmp_critical_name *, uint32_t hint);
3896 
3897 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3898 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3899 
3900 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3901                                                   kmp_int32 global_tid);
3902 
3903 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3904 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3905 
3906 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
3907 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
3908                                          kmp_int32 numberOfSections);
3909 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
3910 
3911 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3912                                      kmp_int32 schedtype, kmp_int32 *plastiter,
3913                                      kmp_int *plower, kmp_int *pupper,
3914                                      kmp_int *pstride, kmp_int incr,
3915                                      kmp_int chunk);
3916 
3917 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3918 
3919 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3920                                    size_t cpy_size, void *cpy_data,
3921                                    void (*cpy_func)(void *, void *),
3922                                    kmp_int32 didit);
3923 
3924 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
3925                                           void *cpy_data);
3926 
3927 extern void KMPC_SET_NUM_THREADS(int arg);
3928 extern void KMPC_SET_DYNAMIC(int flag);
3929 extern void KMPC_SET_NESTED(int flag);
3930 
3931 /* OMP 3.0 tasking interface routines */
3932 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3933                                      kmp_task_t *new_task);
3934 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3935                                              kmp_int32 flags,
3936                                              size_t sizeof_kmp_task_t,
3937                                              size_t sizeof_shareds,
3938                                              kmp_routine_entry_t task_entry);
3939 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
3940     ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
3941     size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
3942 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3943                                           kmp_task_t *task);
3944 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3945                                              kmp_task_t *task);
3946 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3947                                            kmp_task_t *new_task);
3948 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3949 
3950 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3951                                           int end_part);
3952 
3953 #if TASK_UNUSED
3954 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3955 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3956                               kmp_task_t *task);
3957 #endif // TASK_UNUSED
3958 
3959 /* ------------------------------------------------------------------------ */
3960 
3961 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3962 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3963 
3964 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3965     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3966     kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3967     kmp_depend_info_t *noalias_dep_list);
3968 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3969                                      kmp_int32 ndeps,
3970                                      kmp_depend_info_t *dep_list,
3971                                      kmp_int32 ndeps_noalias,
3972                                      kmp_depend_info_t *noalias_dep_list);
3973 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3974                                 bool serialize_immediate);
3975 
3976 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3977                                    kmp_int32 cncl_kind);
3978 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3979                                               kmp_int32 cncl_kind);
3980 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3981 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3982 
3983 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3984 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3985 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3986                                 kmp_int32 if_val, kmp_uint64 *lb,
3987                                 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3988                                 kmp_int32 sched, kmp_uint64 grainsize,
3989                                 void *task_dup);
3990 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3991                                   kmp_task_t *task, kmp_int32 if_val,
3992                                   kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3993                                   kmp_int32 nogroup, kmp_int32 sched,
3994                                   kmp_uint64 grainsize, kmp_int32 modifier,
3995                                   void *task_dup);
3996 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3997 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3998 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3999 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4000                                                      int is_ws, int num,
4001                                                      void *data);
4002 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4003                                               int num, void *data);
4004 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4005                                                     int is_ws);
4006 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4007     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4008     kmp_task_affinity_info_t *affin_list);
4009 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4010 KMP_EXPORT int __kmp_get_max_teams(void);
4011 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4012 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4013 
4014 /* Lock interface routines (fast versions with gtid passed in) */
4015 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4016                                  void **user_lock);
4017 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4018                                       void **user_lock);
4019 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4020                                     void **user_lock);
4021 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4022                                          void **user_lock);
4023 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4024 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4025                                      void **user_lock);
4026 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4027                                   void **user_lock);
4028 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4029                                        void **user_lock);
4030 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4031 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4032                                      void **user_lock);
4033 
4034 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4035                                            void **user_lock, uintptr_t hint);
4036 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4037                                                 void **user_lock,
4038                                                 uintptr_t hint);
4039 
4040 /* Interface to fast scalable reduce methods routines */
4041 
4042 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4043     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4044     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4045     kmp_critical_name *lck);
4046 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4047                                          kmp_critical_name *lck);
4048 KMP_EXPORT kmp_int32 __kmpc_reduce(
4049     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4050     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4051     kmp_critical_name *lck);
4052 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4053                                   kmp_critical_name *lck);
4054 
4055 /* Internal fast reduction routines */
4056 
4057 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4058     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4059     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4060     kmp_critical_name *lck);
4061 
4062 // this function is for testing set/get/determine reduce method
4063 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4064 
4065 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4066 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4067 
4068 // C++ port
4069 // missing 'extern "C"' declarations
4070 
4071 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4072 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4073 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4074                                         kmp_int32 num_threads);
4075 
4076 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4077                                       int proc_bind);
4078 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4079                                       kmp_int32 num_teams,
4080                                       kmp_int32 num_threads);
4081 /* Function for OpenMP 5.1 num_teams clause */
4082 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4083                                          kmp_int32 num_teams_lb,
4084                                          kmp_int32 num_teams_ub,
4085                                          kmp_int32 num_threads);
4086 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4087                                   kmpc_micro microtask, ...);
4088 struct kmp_dim { // loop bounds info casted to kmp_int64
4089   kmp_int64 lo; // lower
4090   kmp_int64 up; // upper
4091   kmp_int64 st; // stride
4092 };
4093 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4094                                      kmp_int32 num_dims,
4095                                      const struct kmp_dim *dims);
4096 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4097                                      const kmp_int64 *vec);
4098 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4099                                      const kmp_int64 *vec);
4100 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4101 
4102 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4103                                              void *data, size_t size,
4104                                              void ***cache);
4105 
4106 // Symbols for MS mutual detection.
4107 extern int _You_must_link_with_exactly_one_OpenMP_library;
4108 extern int _You_must_link_with_Intel_OpenMP_library;
4109 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
4110 extern int _You_must_link_with_Microsoft_OpenMP_library;
4111 #endif
4112 
4113 // The routines below are not exported.
4114 // Consider making them 'static' in corresponding source files.
4115 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4116                                            void *data_addr, size_t pc_size);
4117 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4118                                                 void *data_addr,
4119                                                 size_t pc_size);
4120 void __kmp_threadprivate_resize_cache(int newCapacity);
4121 void __kmp_cleanup_threadprivate_caches();
4122 
4123 // ompc_, kmpc_ entries moved from omp.h.
4124 #if KMP_OS_WINDOWS
4125 #define KMPC_CONVENTION __cdecl
4126 #else
4127 #define KMPC_CONVENTION
4128 #endif
4129 
4130 #ifndef __OMP_H
4131 typedef enum omp_sched_t {
4132   omp_sched_static = 1,
4133   omp_sched_dynamic = 2,
4134   omp_sched_guided = 3,
4135   omp_sched_auto = 4
4136 } omp_sched_t;
4137 typedef void *kmp_affinity_mask_t;
4138 #endif
4139 
4140 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4141 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4142 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4143 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4144 KMP_EXPORT int KMPC_CONVENTION
4145 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4146 KMP_EXPORT int KMPC_CONVENTION
4147 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4148 KMP_EXPORT int KMPC_CONVENTION
4149 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4150 
4151 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4152 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4153 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4154 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4155 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4156 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4157 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4158 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4159 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4160                                               char const *format);
4161 
4162 enum kmp_target_offload_kind {
4163   tgt_disabled = 0,
4164   tgt_default = 1,
4165   tgt_mandatory = 2
4166 };
4167 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4168 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4169 extern kmp_target_offload_kind_t __kmp_target_offload;
4170 extern int __kmpc_get_target_offload();
4171 
4172 // Constants used in libomptarget
4173 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4174 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4175 
4176 // OMP Pause Resource
4177 
4178 // The following enum is used both to set the status in __kmp_pause_status, and
4179 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4180 typedef enum kmp_pause_status_t {
4181   kmp_not_paused = 0, // status is not paused, or, requesting resume
4182   kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4183   kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4184 } kmp_pause_status_t;
4185 
4186 // This stores the pause state of the runtime
4187 extern kmp_pause_status_t __kmp_pause_status;
4188 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4189 extern int __kmp_pause_resource(kmp_pause_status_t level);
4190 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4191 extern void __kmp_resume_if_soft_paused();
4192 // Hard resume simply resets the status to not paused. Library will appear to
4193 // be uninitialized after hard pause. Let OMP constructs trigger required
4194 // initializations.
4195 static inline void __kmp_resume_if_hard_paused() {
4196   if (__kmp_pause_status == kmp_hard_paused) {
4197     __kmp_pause_status = kmp_not_paused;
4198   }
4199 }
4200 
4201 extern void __kmp_omp_display_env(int verbose);
4202 
4203 // 1: it is initializing hidden helper team
4204 extern volatile int __kmp_init_hidden_helper;
4205 // 1: the hidden helper team is done
4206 extern volatile int __kmp_hidden_helper_team_done;
4207 // 1: enable hidden helper task
4208 extern kmp_int32 __kmp_enable_hidden_helper;
4209 // Main thread of hidden helper team
4210 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4211 // Descriptors for the hidden helper threads
4212 extern kmp_info_t **__kmp_hidden_helper_threads;
4213 // Number of hidden helper threads
4214 extern kmp_int32 __kmp_hidden_helper_threads_num;
4215 // Number of hidden helper tasks that have not been executed yet
4216 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4217 
4218 extern void __kmp_hidden_helper_initialize();
4219 extern void __kmp_hidden_helper_threads_initz_routine();
4220 extern void __kmp_do_initialize_hidden_helper_threads();
4221 extern void __kmp_hidden_helper_threads_initz_wait();
4222 extern void __kmp_hidden_helper_initz_release();
4223 extern void __kmp_hidden_helper_threads_deinitz_wait();
4224 extern void __kmp_hidden_helper_threads_deinitz_release();
4225 extern void __kmp_hidden_helper_main_thread_wait();
4226 extern void __kmp_hidden_helper_worker_thread_wait();
4227 extern void __kmp_hidden_helper_worker_thread_signal();
4228 extern void __kmp_hidden_helper_main_thread_release();
4229 
4230 // Check whether a given thread is a hidden helper thread
4231 #define KMP_HIDDEN_HELPER_THREAD(gtid)                                         \
4232   ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4233 
4234 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)                                  \
4235   ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4236 
4237 #define KMP_HIDDEN_HELPER_TEAM(team)                                           \
4238   (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4239 
4240 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4241 // main thread, is skipped.
4242 #define KMP_GTID_TO_SHADOW_GTID(gtid)                                          \
4243   ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4244 
4245 // Return the adjusted gtid value by subtracting from gtid the number
4246 // of hidden helper threads. This adjusted value is the gtid the thread would
4247 // have received if there were no hidden helper threads.
4248 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4249   int adjusted_gtid = gtid;
4250   if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4251       gtid - __kmp_hidden_helper_threads_num >= 0) {
4252     adjusted_gtid -= __kmp_hidden_helper_threads_num;
4253   }
4254   return adjusted_gtid;
4255 }
4256 
4257 // Support for error directive
4258 typedef enum kmp_severity_t {
4259   severity_warning = 1,
4260   severity_fatal = 2
4261 } kmp_severity_t;
4262 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4263 
4264 // Support for scope directive
4265 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4266 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4267 
4268 #ifdef __cplusplus
4269 }
4270 #endif
4271 
4272 template <bool C, bool S>
4273 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4274 template <bool C, bool S>
4275 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4276 template <bool C, bool S>
4277 extern void __kmp_atomic_suspend_64(int th_gtid,
4278                                     kmp_atomic_flag_64<C, S> *flag);
4279 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4280 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4281 template <bool C, bool S>
4282 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4283 template <bool C, bool S>
4284 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4285 template <bool C, bool S>
4286 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4287 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4288 #endif
4289 template <bool C, bool S>
4290 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4291 template <bool C, bool S>
4292 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4293 template <bool C, bool S>
4294 extern void __kmp_atomic_resume_64(int target_gtid,
4295                                    kmp_atomic_flag_64<C, S> *flag);
4296 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4297 
4298 template <bool C, bool S>
4299 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4300                            kmp_flag_32<C, S> *flag, int final_spin,
4301                            int *thread_finished,
4302 #if USE_ITT_BUILD
4303                            void *itt_sync_obj,
4304 #endif /* USE_ITT_BUILD */
4305                            kmp_int32 is_constrained);
4306 template <bool C, bool S>
4307 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4308                            kmp_flag_64<C, S> *flag, int final_spin,
4309                            int *thread_finished,
4310 #if USE_ITT_BUILD
4311                            void *itt_sync_obj,
4312 #endif /* USE_ITT_BUILD */
4313                            kmp_int32 is_constrained);
4314 template <bool C, bool S>
4315 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4316                                   kmp_atomic_flag_64<C, S> *flag,
4317                                   int final_spin, int *thread_finished,
4318 #if USE_ITT_BUILD
4319                                   void *itt_sync_obj,
4320 #endif /* USE_ITT_BUILD */
4321                                   kmp_int32 is_constrained);
4322 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4323                                kmp_flag_oncore *flag, int final_spin,
4324                                int *thread_finished,
4325 #if USE_ITT_BUILD
4326                                void *itt_sync_obj,
4327 #endif /* USE_ITT_BUILD */
4328                                kmp_int32 is_constrained);
4329 
4330 extern int __kmp_nesting_mode;
4331 extern int __kmp_nesting_mode_nlevels;
4332 extern int *__kmp_nesting_nth_level;
4333 extern void __kmp_init_nesting_mode();
4334 extern void __kmp_set_nesting_mode_threads();
4335 
4336 /// This class safely opens and closes a C-style FILE* object using RAII
4337 /// semantics. There are also methods which allow using stdout or stderr as
4338 /// the underlying FILE* object. With the implicit conversion operator to
4339 /// FILE*, an object with this type can be used in any function which takes
4340 /// a FILE* object e.g., fprintf().
4341 /// No close method is needed at use sites.
4342 class kmp_safe_raii_file_t {
4343   FILE *f;
4344 
4345   void close() {
4346     if (f && f != stdout && f != stderr) {
4347       fclose(f);
4348       f = nullptr;
4349     }
4350   }
4351 
4352 public:
4353   kmp_safe_raii_file_t() : f(nullptr) {}
4354   kmp_safe_raii_file_t(const char *filename, const char *mode,
4355                        const char *env_var = nullptr)
4356       : f(nullptr) {
4357     open(filename, mode, env_var);
4358   }
4359   ~kmp_safe_raii_file_t() { close(); }
4360 
4361   /// Open filename using mode. This is automatically closed in the destructor.
4362   /// The env_var parameter indicates the environment variable the filename
4363   /// came from if != nullptr.
4364   void open(const char *filename, const char *mode,
4365             const char *env_var = nullptr) {
4366     KMP_ASSERT(!f);
4367     f = fopen(filename, mode);
4368     if (!f) {
4369       int code = errno;
4370       if (env_var) {
4371         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4372                     KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4373       } else {
4374         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4375                     __kmp_msg_null);
4376       }
4377     }
4378   }
4379   /// Instead of erroring out, return non-zero when
4380   /// unsuccessful fopen() for any reason
4381   int try_open(const char *filename, const char *mode) {
4382     KMP_ASSERT(!f);
4383     f = fopen(filename, mode);
4384     if (!f)
4385       return errno;
4386     return 0;
4387   }
4388   /// Set the FILE* object to stdout and output there
4389   /// No open call should happen before this call.
4390   void set_stdout() {
4391     KMP_ASSERT(!f);
4392     f = stdout;
4393   }
4394   /// Set the FILE* object to stderr and output there
4395   /// No open call should happen before this call.
4396   void set_stderr() {
4397     KMP_ASSERT(!f);
4398     f = stderr;
4399   }
4400   operator bool() { return bool(f); }
4401   operator FILE *() { return f; }
4402 };
4403 
4404 template <typename SourceType, typename TargetType,
4405           bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4406           bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4407           bool isSourceSigned = std::is_signed<SourceType>::value,
4408           bool isTargetSigned = std::is_signed<TargetType>::value>
4409 struct kmp_convert {};
4410 
4411 // Both types are signed; Source smaller
4412 template <typename SourceType, typename TargetType>
4413 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4414   static TargetType to(SourceType src) { return (TargetType)src; }
4415 };
4416 // Source equal
4417 template <typename SourceType, typename TargetType>
4418 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4419   static TargetType to(SourceType src) { return src; }
4420 };
4421 // Source bigger
4422 template <typename SourceType, typename TargetType>
4423 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4424   static TargetType to(SourceType src) {
4425     KMP_ASSERT(src <= static_cast<SourceType>(
4426                           (std::numeric_limits<TargetType>::max)()));
4427     KMP_ASSERT(src >= static_cast<SourceType>(
4428                           (std::numeric_limits<TargetType>::min)()));
4429     return (TargetType)src;
4430   }
4431 };
4432 
4433 // Source signed, Target unsigned
4434 // Source smaller
4435 template <typename SourceType, typename TargetType>
4436 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4437   static TargetType to(SourceType src) {
4438     KMP_ASSERT(src >= 0);
4439     return (TargetType)src;
4440   }
4441 };
4442 // Source equal
4443 template <typename SourceType, typename TargetType>
4444 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4445   static TargetType to(SourceType src) {
4446     KMP_ASSERT(src >= 0);
4447     return (TargetType)src;
4448   }
4449 };
4450 // Source bigger
4451 template <typename SourceType, typename TargetType>
4452 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4453   static TargetType to(SourceType src) {
4454     KMP_ASSERT(src >= 0);
4455     KMP_ASSERT(src <= static_cast<SourceType>(
4456                           (std::numeric_limits<TargetType>::max)()));
4457     return (TargetType)src;
4458   }
4459 };
4460 
4461 // Source unsigned, Target signed
4462 // Source smaller
4463 template <typename SourceType, typename TargetType>
4464 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4465   static TargetType to(SourceType src) { return (TargetType)src; }
4466 };
4467 // Source equal
4468 template <typename SourceType, typename TargetType>
4469 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4470   static TargetType to(SourceType src) {
4471     KMP_ASSERT(src <= static_cast<SourceType>(
4472                           (std::numeric_limits<TargetType>::max)()));
4473     return (TargetType)src;
4474   }
4475 };
4476 // Source bigger
4477 template <typename SourceType, typename TargetType>
4478 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4479   static TargetType to(SourceType src) {
4480     KMP_ASSERT(src <= static_cast<SourceType>(
4481                           (std::numeric_limits<TargetType>::max)()));
4482     return (TargetType)src;
4483   }
4484 };
4485 
4486 // Source unsigned, Target unsigned
4487 // Source smaller
4488 template <typename SourceType, typename TargetType>
4489 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4490   static TargetType to(SourceType src) { return (TargetType)src; }
4491 };
4492 // Source equal
4493 template <typename SourceType, typename TargetType>
4494 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4495   static TargetType to(SourceType src) { return src; }
4496 };
4497 // Source bigger
4498 template <typename SourceType, typename TargetType>
4499 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4500   static TargetType to(SourceType src) {
4501     KMP_ASSERT(src <= static_cast<SourceType>(
4502                           (std::numeric_limits<TargetType>::max)()));
4503     return (TargetType)src;
4504   }
4505 };
4506 
4507 template <typename T1, typename T2>
4508 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4509   *dest = kmp_convert<T1, T2>::to(src);
4510 }
4511 
4512 #endif /* KMP_H */
4513