xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp.h (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
1 /*! \file */
2 /*
3  * kmp.h -- KPTS runtime header file.
4  */
5 
6 //===----------------------------------------------------------------------===//
7 //
8 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9 // See https://llvm.org/LICENSE.txt for license information.
10 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef KMP_H
15 #define KMP_H
16 
17 #include "kmp_config.h"
18 
19 /* #define BUILD_PARALLEL_ORDERED 1 */
20 
21 /* This fix replaces gettimeofday with clock_gettime for better scalability on
22    the Altix.  Requires user code to be linked with -lrt. */
23 //#define FIX_SGI_CLOCK
24 
25 /* Defines for OpenMP 3.0 tasking and auto scheduling */
26 
27 #ifndef KMP_STATIC_STEAL_ENABLED
28 #define KMP_STATIC_STEAL_ENABLED 1
29 #endif
30 
31 #define TASK_CURRENT_NOT_QUEUED 0
32 #define TASK_CURRENT_QUEUED 1
33 
34 #ifdef BUILD_TIED_TASK_STACK
35 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37 // Number of entries in each task stack array
38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39 // Mask for determining index into stack block
40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41 #endif // BUILD_TIED_TASK_STACK
42 
43 #define TASK_NOT_PUSHED 1
44 #define TASK_SUCCESSFULLY_PUSHED 0
45 #define TASK_TIED 1
46 #define TASK_UNTIED 0
47 #define TASK_EXPLICIT 1
48 #define TASK_IMPLICIT 0
49 #define TASK_PROXY 1
50 #define TASK_FULL 0
51 #define TASK_DETACHABLE 1
52 #define TASK_UNDETACHABLE 0
53 
54 #define KMP_CANCEL_THREADS
55 #define KMP_THREAD_ATTR
56 
57 // Android does not have pthread_cancel.  Undefine KMP_CANCEL_THREADS if being
58 // built on Android
59 #if defined(__ANDROID__)
60 #undef KMP_CANCEL_THREADS
61 #endif
62 
63 #include <signal.h>
64 #include <stdarg.h>
65 #include <stddef.h>
66 #include <stdio.h>
67 #include <stdlib.h>
68 #include <string.h>
69 #include <limits>
70 #include <type_traits>
71 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72    Microsoft library. Some macros provided below to replace these functions  */
73 #ifndef __ABSOFT_WIN
74 #include <sys/types.h>
75 #endif
76 #include <limits.h>
77 #include <time.h>
78 
79 #include <errno.h>
80 
81 #include "kmp_os.h"
82 
83 #include "kmp_safe_c_api.h"
84 
85 #if KMP_STATS_ENABLED
86 class kmp_stats_list;
87 #endif
88 
89 #if KMP_USE_HIER_SCHED
90 // Only include hierarchical scheduling if affinity is supported
91 #undef KMP_USE_HIER_SCHED
92 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93 #endif
94 
95 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96 #include "hwloc.h"
97 #ifndef HWLOC_OBJ_NUMANODE
98 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99 #endif
100 #ifndef HWLOC_OBJ_PACKAGE
101 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102 #endif
103 #if HWLOC_API_VERSION >= 0x00020000
104 // hwloc 2.0 changed type of depth of object from unsigned to int
105 typedef int kmp_hwloc_depth_t;
106 #else
107 typedef unsigned int kmp_hwloc_depth_t;
108 #endif
109 #endif
110 
111 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
112 #include <xmmintrin.h>
113 #endif
114 
115 #include "kmp_debug.h"
116 #include "kmp_lock.h"
117 #include "kmp_version.h"
118 #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 char *__kmp_cpuinfo_file;
845 
846 #endif /* KMP_AFFINITY_SUPPORTED */
847 
848 // This needs to be kept in sync with the values in omp.h !!!
849 typedef enum kmp_proc_bind_t {
850   proc_bind_false = 0,
851   proc_bind_true,
852   proc_bind_primary,
853   proc_bind_close,
854   proc_bind_spread,
855   proc_bind_intel, // use KMP_AFFINITY interface
856   proc_bind_default
857 } kmp_proc_bind_t;
858 
859 typedef struct kmp_nested_proc_bind_t {
860   kmp_proc_bind_t *bind_types;
861   int size;
862   int used;
863 } kmp_nested_proc_bind_t;
864 
865 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
866 extern kmp_proc_bind_t __kmp_teams_proc_bind;
867 
868 extern int __kmp_display_affinity;
869 extern char *__kmp_affinity_format;
870 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
871 #if OMPT_SUPPORT
872 extern int __kmp_tool;
873 extern char *__kmp_tool_libraries;
874 #endif // OMPT_SUPPORT
875 
876 #if KMP_AFFINITY_SUPPORTED
877 #define KMP_PLACE_ALL (-1)
878 #define KMP_PLACE_UNDEFINED (-2)
879 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
880 #define KMP_AFFINITY_NON_PROC_BIND                                             \
881   ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false ||                 \
882     __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) &&                \
883    (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
884 #endif /* KMP_AFFINITY_SUPPORTED */
885 
886 extern int __kmp_affinity_num_places;
887 
888 typedef enum kmp_cancel_kind_t {
889   cancel_noreq = 0,
890   cancel_parallel = 1,
891   cancel_loop = 2,
892   cancel_sections = 3,
893   cancel_taskgroup = 4
894 } kmp_cancel_kind_t;
895 
896 // KMP_HW_SUBSET support:
897 typedef struct kmp_hws_item {
898   int num;
899   int offset;
900 } kmp_hws_item_t;
901 
902 extern kmp_hws_item_t __kmp_hws_socket;
903 extern kmp_hws_item_t __kmp_hws_die;
904 extern kmp_hws_item_t __kmp_hws_node;
905 extern kmp_hws_item_t __kmp_hws_tile;
906 extern kmp_hws_item_t __kmp_hws_core;
907 extern kmp_hws_item_t __kmp_hws_proc;
908 extern int __kmp_hws_requested;
909 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
910 
911 /* ------------------------------------------------------------------------ */
912 
913 #define KMP_PAD(type, sz)                                                      \
914   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
915 
916 // We need to avoid using -1 as a GTID as +1 is added to the gtid
917 // when storing it in a lock, and the value 0 is reserved.
918 #define KMP_GTID_DNE (-2) /* Does not exist */
919 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
920 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
921 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
922 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
923 
924 /* OpenMP 5.0 Memory Management support */
925 
926 #ifndef __OMP_H
927 // Duplicate type definitions from omp.h
928 typedef uintptr_t omp_uintptr_t;
929 
930 typedef enum {
931   omp_atk_sync_hint = 1,
932   omp_atk_alignment = 2,
933   omp_atk_access = 3,
934   omp_atk_pool_size = 4,
935   omp_atk_fallback = 5,
936   omp_atk_fb_data = 6,
937   omp_atk_pinned = 7,
938   omp_atk_partition = 8
939 } omp_alloctrait_key_t;
940 
941 typedef enum {
942   omp_atv_false = 0,
943   omp_atv_true = 1,
944   omp_atv_contended = 3,
945   omp_atv_uncontended = 4,
946   omp_atv_serialized = 5,
947   omp_atv_sequential = omp_atv_serialized, // (deprecated)
948   omp_atv_private = 6,
949   omp_atv_all = 7,
950   omp_atv_thread = 8,
951   omp_atv_pteam = 9,
952   omp_atv_cgroup = 10,
953   omp_atv_default_mem_fb = 11,
954   omp_atv_null_fb = 12,
955   omp_atv_abort_fb = 13,
956   omp_atv_allocator_fb = 14,
957   omp_atv_environment = 15,
958   omp_atv_nearest = 16,
959   omp_atv_blocked = 17,
960   omp_atv_interleaved = 18
961 } omp_alloctrait_value_t;
962 #define omp_atv_default ((omp_uintptr_t)-1)
963 
964 typedef void *omp_memspace_handle_t;
965 extern omp_memspace_handle_t const omp_default_mem_space;
966 extern omp_memspace_handle_t const omp_large_cap_mem_space;
967 extern omp_memspace_handle_t const omp_const_mem_space;
968 extern omp_memspace_handle_t const omp_high_bw_mem_space;
969 extern omp_memspace_handle_t const omp_low_lat_mem_space;
970 // Preview of target memory support
971 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
972 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
973 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
974 
975 typedef struct {
976   omp_alloctrait_key_t key;
977   omp_uintptr_t value;
978 } omp_alloctrait_t;
979 
980 typedef void *omp_allocator_handle_t;
981 extern omp_allocator_handle_t const omp_null_allocator;
982 extern omp_allocator_handle_t const omp_default_mem_alloc;
983 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
984 extern omp_allocator_handle_t const omp_const_mem_alloc;
985 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
986 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
987 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
988 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
989 extern omp_allocator_handle_t const omp_thread_mem_alloc;
990 // Preview of target memory support
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
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                    : "=r"(flag)
1338                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1339                    :);
1340   return flag;
1341 #else
1342   return _tpause(hint, counter);
1343 #endif
1344 }
1345 KMP_ATTRIBUTE_TARGET_WAITPKG
1346 static inline void __kmp_umonitor(void *cacheline) {
1347 #if !KMP_HAVE_WAITPKG_INTRINSICS
1348   __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1349                    :
1350                    : "a"(cacheline)
1351                    :);
1352 #else
1353   _umonitor(cacheline);
1354 #endif
1355 }
1356 KMP_ATTRIBUTE_TARGET_WAITPKG
1357 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1358 #if !KMP_HAVE_WAITPKG_INTRINSICS
1359   uint32_t timeHi = uint32_t(counter >> 32);
1360   uint32_t timeLo = uint32_t(counter & 0xffffffff);
1361   char flag;
1362   __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1363                    "setb   %0"
1364                    : "=r"(flag)
1365                    : "a"(timeLo), "d"(timeHi), "c"(hint)
1366                    :);
1367   return flag;
1368 #else
1369   return _umwait(hint, counter);
1370 #endif
1371 }
1372 #elif KMP_HAVE_MWAIT
1373 #if KMP_OS_UNIX
1374 #include <pmmintrin.h>
1375 #else
1376 #include <intrin.h>
1377 #endif
1378 #if KMP_OS_UNIX
1379 __attribute__((target("sse3")))
1380 #endif
1381 static inline void
1382 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1383   _mm_monitor(cacheline, extensions, hints);
1384 }
1385 #if KMP_OS_UNIX
1386 __attribute__((target("sse3")))
1387 #endif
1388 static inline void
1389 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1390   _mm_mwait(extensions, hints);
1391 }
1392 #endif // KMP_HAVE_UMWAIT
1393 
1394 #if KMP_ARCH_X86
1395 extern void __kmp_x86_pause(void);
1396 #elif KMP_MIC
1397 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1398 // regression after removal of extra PAUSE from spin loops. Changing
1399 // the delay from 100 to 300 showed even better performance than double PAUSE
1400 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1401 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1402 #else
1403 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1404 #endif
1405 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1406 #elif KMP_ARCH_PPC64
1407 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1408 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1409 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1410 #define KMP_CPU_PAUSE()                                                        \
1411   do {                                                                         \
1412     KMP_PPC64_PRI_LOW();                                                       \
1413     KMP_PPC64_PRI_MED();                                                       \
1414     KMP_PPC64_PRI_LOC_MB();                                                    \
1415   } while (0)
1416 #else
1417 #define KMP_CPU_PAUSE() /* nothing to do */
1418 #endif
1419 
1420 #define KMP_INIT_YIELD(count)                                                  \
1421   { (count) = __kmp_yield_init; }
1422 
1423 #define KMP_INIT_BACKOFF(time)                                                 \
1424   { (time) = __kmp_pause_init; }
1425 
1426 #define KMP_OVERSUBSCRIBED                                                     \
1427   (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1428 
1429 #define KMP_TRY_YIELD                                                          \
1430   ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1431 
1432 #define KMP_TRY_YIELD_OVERSUB                                                  \
1433   ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1434 
1435 #define KMP_YIELD(cond)                                                        \
1436   {                                                                            \
1437     KMP_CPU_PAUSE();                                                           \
1438     if ((cond) && (KMP_TRY_YIELD))                                             \
1439       __kmp_yield();                                                           \
1440   }
1441 
1442 #define KMP_YIELD_OVERSUB()                                                    \
1443   {                                                                            \
1444     KMP_CPU_PAUSE();                                                           \
1445     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1446       __kmp_yield();                                                           \
1447   }
1448 
1449 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1450 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1451 #define KMP_YIELD_SPIN(count)                                                  \
1452   {                                                                            \
1453     KMP_CPU_PAUSE();                                                           \
1454     if (KMP_TRY_YIELD) {                                                       \
1455       (count) -= 2;                                                            \
1456       if (!(count)) {                                                          \
1457         __kmp_yield();                                                         \
1458         (count) = __kmp_yield_next;                                            \
1459       }                                                                        \
1460     }                                                                          \
1461   }
1462 
1463 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1464 // (C0.2) state, which improves performance of other SMT threads on the same
1465 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1466 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1467 // available, fall back to the regular CPU pause and yield combination.
1468 #if KMP_HAVE_UMWAIT
1469 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1470   {                                                                            \
1471     if (__kmp_tpause_enabled) {                                                \
1472       if (KMP_OVERSUBSCRIBED) {                                                \
1473         __kmp_tpause(0, (time));                                               \
1474       } else {                                                                 \
1475         __kmp_tpause(__kmp_tpause_hint, (time));                               \
1476       }                                                                        \
1477       (time) *= 2;                                                             \
1478     } else {                                                                   \
1479       KMP_CPU_PAUSE();                                                         \
1480       if ((KMP_TRY_YIELD_OVERSUB)) {                                           \
1481         __kmp_yield();                                                         \
1482       } else if (__kmp_use_yield == 1) {                                       \
1483         (count) -= 2;                                                          \
1484         if (!(count)) {                                                        \
1485           __kmp_yield();                                                       \
1486           (count) = __kmp_yield_next;                                          \
1487         }                                                                      \
1488       }                                                                        \
1489     }                                                                          \
1490   }
1491 #else
1492 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time)                               \
1493   {                                                                            \
1494     KMP_CPU_PAUSE();                                                           \
1495     if ((KMP_TRY_YIELD_OVERSUB))                                               \
1496       __kmp_yield();                                                           \
1497     else if (__kmp_use_yield == 1) {                                           \
1498       (count) -= 2;                                                            \
1499       if (!(count)) {                                                          \
1500         __kmp_yield();                                                         \
1501         (count) = __kmp_yield_next;                                            \
1502       }                                                                        \
1503     }                                                                          \
1504   }
1505 #endif // KMP_HAVE_UMWAIT
1506 
1507 /* ------------------------------------------------------------------------ */
1508 /* Support datatypes for the orphaned construct nesting checks.             */
1509 /* ------------------------------------------------------------------------ */
1510 
1511 /* When adding to this enum, add its corresponding string in cons_text_c[]
1512  * array in kmp_error.cpp */
1513 enum cons_type {
1514   ct_none,
1515   ct_parallel,
1516   ct_pdo,
1517   ct_pdo_ordered,
1518   ct_psections,
1519   ct_psingle,
1520   ct_critical,
1521   ct_ordered_in_parallel,
1522   ct_ordered_in_pdo,
1523   ct_master,
1524   ct_reduce,
1525   ct_barrier,
1526   ct_masked
1527 };
1528 
1529 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1530 
1531 struct cons_data {
1532   ident_t const *ident;
1533   enum cons_type type;
1534   int prev;
1535   kmp_user_lock_p
1536       name; /* address exclusively for critical section name comparison */
1537 };
1538 
1539 struct cons_header {
1540   int p_top, w_top, s_top;
1541   int stack_size, stack_top;
1542   struct cons_data *stack_data;
1543 };
1544 
1545 struct kmp_region_info {
1546   char *text;
1547   int offset[KMP_MAX_FIELDS];
1548   int length[KMP_MAX_FIELDS];
1549 };
1550 
1551 /* ---------------------------------------------------------------------- */
1552 /* ---------------------------------------------------------------------- */
1553 
1554 #if KMP_OS_WINDOWS
1555 typedef HANDLE kmp_thread_t;
1556 typedef DWORD kmp_key_t;
1557 #endif /* KMP_OS_WINDOWS */
1558 
1559 #if KMP_OS_UNIX
1560 typedef pthread_t kmp_thread_t;
1561 typedef pthread_key_t kmp_key_t;
1562 #endif
1563 
1564 extern kmp_key_t __kmp_gtid_threadprivate_key;
1565 
1566 typedef struct kmp_sys_info {
1567   long maxrss; /* the maximum resident set size utilized (in kilobytes)     */
1568   long minflt; /* the number of page faults serviced without any I/O        */
1569   long majflt; /* the number of page faults serviced that required I/O      */
1570   long nswap; /* the number of times a process was "swapped" out of memory */
1571   long inblock; /* the number of times the file system had to perform input  */
1572   long oublock; /* the number of times the file system had to perform output */
1573   long nvcsw; /* the number of times a context switch was voluntarily      */
1574   long nivcsw; /* the number of times a context switch was forced           */
1575 } kmp_sys_info_t;
1576 
1577 #if USE_ITT_BUILD
1578 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1579 // required type here. Later we will check the type meets requirements.
1580 typedef int kmp_itt_mark_t;
1581 #define KMP_ITT_DEBUG 0
1582 #endif /* USE_ITT_BUILD */
1583 
1584 typedef kmp_int32 kmp_critical_name[8];
1585 
1586 /*!
1587 @ingroup PARALLEL
1588 The type for a microtask which gets passed to @ref __kmpc_fork_call().
1589 The arguments to the outlined function are
1590 @param global_tid the global thread identity of the thread executing the
1591 function.
1592 @param bound_tid  the local identity of the thread executing the function
1593 @param ... pointers to shared variables accessed by the function.
1594 */
1595 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1596 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1597                                  ...);
1598 
1599 /*!
1600 @ingroup THREADPRIVATE
1601 @{
1602 */
1603 /* ---------------------------------------------------------------------------
1604  */
1605 /* Threadprivate initialization/finalization function declarations */
1606 
1607 /*  for non-array objects:  __kmpc_threadprivate_register()  */
1608 
1609 /*!
1610  Pointer to the constructor function.
1611  The first argument is the <tt>this</tt> pointer
1612 */
1613 typedef void *(*kmpc_ctor)(void *);
1614 
1615 /*!
1616  Pointer to the destructor function.
1617  The first argument is the <tt>this</tt> pointer
1618 */
1619 typedef void (*kmpc_dtor)(
1620     void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1621                               compiler */
1622 /*!
1623  Pointer to an alternate constructor.
1624  The first argument is the <tt>this</tt> pointer.
1625 */
1626 typedef void *(*kmpc_cctor)(void *, void *);
1627 
1628 /* for array objects: __kmpc_threadprivate_register_vec() */
1629 /* First arg: "this" pointer */
1630 /* Last arg: number of array elements */
1631 /*!
1632  Array constructor.
1633  First argument is the <tt>this</tt> pointer
1634  Second argument the number of array elements.
1635 */
1636 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1637 /*!
1638  Pointer to the array destructor function.
1639  The first argument is the <tt>this</tt> pointer
1640  Second argument the number of array elements.
1641 */
1642 typedef void (*kmpc_dtor_vec)(void *, size_t);
1643 /*!
1644  Array constructor.
1645  First argument is the <tt>this</tt> pointer
1646  Third argument the number of array elements.
1647 */
1648 typedef void *(*kmpc_cctor_vec)(void *, void *,
1649                                 size_t); /* function unused by compiler */
1650 
1651 /*!
1652 @}
1653 */
1654 
1655 /* keeps tracked of threadprivate cache allocations for cleanup later */
1656 typedef struct kmp_cached_addr {
1657   void **addr; /* address of allocated cache */
1658   void ***compiler_cache; /* pointer to compiler's cache */
1659   void *data; /* pointer to global data */
1660   struct kmp_cached_addr *next; /* pointer to next cached address */
1661 } kmp_cached_addr_t;
1662 
1663 struct private_data {
1664   struct private_data *next; /* The next descriptor in the list      */
1665   void *data; /* The data buffer for this descriptor  */
1666   int more; /* The repeat count for this descriptor */
1667   size_t size; /* The data size for this descriptor    */
1668 };
1669 
1670 struct private_common {
1671   struct private_common *next;
1672   struct private_common *link;
1673   void *gbl_addr;
1674   void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1675   size_t cmn_size;
1676 };
1677 
1678 struct shared_common {
1679   struct shared_common *next;
1680   struct private_data *pod_init;
1681   void *obj_init;
1682   void *gbl_addr;
1683   union {
1684     kmpc_ctor ctor;
1685     kmpc_ctor_vec ctorv;
1686   } ct;
1687   union {
1688     kmpc_cctor cctor;
1689     kmpc_cctor_vec cctorv;
1690   } cct;
1691   union {
1692     kmpc_dtor dtor;
1693     kmpc_dtor_vec dtorv;
1694   } dt;
1695   size_t vec_len;
1696   int is_vec;
1697   size_t cmn_size;
1698 };
1699 
1700 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1701 #define KMP_HASH_TABLE_SIZE                                                    \
1702   (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1703 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1704 #define KMP_HASH(x)                                                            \
1705   ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1706 
1707 struct common_table {
1708   struct private_common *data[KMP_HASH_TABLE_SIZE];
1709 };
1710 
1711 struct shared_table {
1712   struct shared_common *data[KMP_HASH_TABLE_SIZE];
1713 };
1714 
1715 /* ------------------------------------------------------------------------ */
1716 
1717 #if KMP_USE_HIER_SCHED
1718 // Shared barrier data that exists inside a single unit of the scheduling
1719 // hierarchy
1720 typedef struct kmp_hier_private_bdata_t {
1721   kmp_int32 num_active;
1722   kmp_uint64 index;
1723   kmp_uint64 wait_val[2];
1724 } kmp_hier_private_bdata_t;
1725 #endif
1726 
1727 typedef struct kmp_sched_flags {
1728   unsigned ordered : 1;
1729   unsigned nomerge : 1;
1730   unsigned contains_last : 1;
1731 #if KMP_USE_HIER_SCHED
1732   unsigned use_hier : 1;
1733   unsigned unused : 28;
1734 #else
1735   unsigned unused : 29;
1736 #endif
1737 } kmp_sched_flags_t;
1738 
1739 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1740 
1741 #if KMP_STATIC_STEAL_ENABLED
1742 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1743   kmp_int32 count;
1744   kmp_int32 ub;
1745   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1746   kmp_int32 lb;
1747   kmp_int32 st;
1748   kmp_int32 tc;
1749   kmp_lock_t *steal_lock; // lock used for chunk stealing
1750   // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1751   //    a) parm3 is properly aligned and
1752   //    b) all parm1-4 are on the same cache line.
1753   // Because of parm1-4 are used together, performance seems to be better
1754   // if they are on the same cache line (not measured though).
1755 
1756   struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1757     kmp_int32 parm1; //     structures in kmp_dispatch.cpp. This should
1758     kmp_int32 parm2; //     make no real change at least while padding is off.
1759     kmp_int32 parm3;
1760     kmp_int32 parm4;
1761   };
1762 
1763   kmp_uint32 ordered_lower;
1764   kmp_uint32 ordered_upper;
1765 #if KMP_OS_WINDOWS
1766   kmp_int32 last_upper;
1767 #endif /* KMP_OS_WINDOWS */
1768 } dispatch_private_info32_t;
1769 
1770 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1771   kmp_int64 count; // current chunk number for static & static-steal scheduling
1772   kmp_int64 ub; /* upper-bound */
1773   /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1774   kmp_int64 lb; /* lower-bound */
1775   kmp_int64 st; /* stride */
1776   kmp_int64 tc; /* trip count (number of iterations) */
1777   kmp_lock_t *steal_lock; // lock used for chunk stealing
1778   /* parm[1-4] are used in different ways by different scheduling algorithms */
1779 
1780   // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1781   //    a) parm3 is properly aligned and
1782   //    b) all parm1-4 are in the same cache line.
1783   // Because of parm1-4 are used together, performance seems to be better
1784   // if they are in the same line (not measured though).
1785 
1786   struct KMP_ALIGN(32) {
1787     kmp_int64 parm1;
1788     kmp_int64 parm2;
1789     kmp_int64 parm3;
1790     kmp_int64 parm4;
1791   };
1792 
1793   kmp_uint64 ordered_lower;
1794   kmp_uint64 ordered_upper;
1795 #if KMP_OS_WINDOWS
1796   kmp_int64 last_upper;
1797 #endif /* KMP_OS_WINDOWS */
1798 } dispatch_private_info64_t;
1799 #else /* KMP_STATIC_STEAL_ENABLED */
1800 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1801   kmp_int32 lb;
1802   kmp_int32 ub;
1803   kmp_int32 st;
1804   kmp_int32 tc;
1805 
1806   kmp_int32 parm1;
1807   kmp_int32 parm2;
1808   kmp_int32 parm3;
1809   kmp_int32 parm4;
1810 
1811   kmp_int32 count;
1812 
1813   kmp_uint32 ordered_lower;
1814   kmp_uint32 ordered_upper;
1815 #if KMP_OS_WINDOWS
1816   kmp_int32 last_upper;
1817 #endif /* KMP_OS_WINDOWS */
1818 } dispatch_private_info32_t;
1819 
1820 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1821   kmp_int64 lb; /* lower-bound */
1822   kmp_int64 ub; /* upper-bound */
1823   kmp_int64 st; /* stride */
1824   kmp_int64 tc; /* trip count (number of iterations) */
1825 
1826   /* parm[1-4] are used in different ways by different scheduling algorithms */
1827   kmp_int64 parm1;
1828   kmp_int64 parm2;
1829   kmp_int64 parm3;
1830   kmp_int64 parm4;
1831 
1832   kmp_int64 count; /* current chunk number for static scheduling */
1833 
1834   kmp_uint64 ordered_lower;
1835   kmp_uint64 ordered_upper;
1836 #if KMP_OS_WINDOWS
1837   kmp_int64 last_upper;
1838 #endif /* KMP_OS_WINDOWS */
1839 } dispatch_private_info64_t;
1840 #endif /* KMP_STATIC_STEAL_ENABLED */
1841 
1842 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1843   union private_info {
1844     dispatch_private_info32_t p32;
1845     dispatch_private_info64_t p64;
1846   } u;
1847   enum sched_type schedule; /* scheduling algorithm */
1848   kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1849   std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1850   kmp_int32 ordered_bumped;
1851   // Stack of buffers for nest of serial regions
1852   struct dispatch_private_info *next;
1853   kmp_int32 type_size; /* the size of types in private_info */
1854 #if KMP_USE_HIER_SCHED
1855   kmp_int32 hier_id;
1856   void *parent; /* hierarchical scheduling parent pointer */
1857 #endif
1858   enum cons_type pushed_ws;
1859 } dispatch_private_info_t;
1860 
1861 typedef struct dispatch_shared_info32 {
1862   /* chunk index under dynamic, number of idle threads under static-steal;
1863      iteration index otherwise */
1864   volatile kmp_uint32 iteration;
1865   volatile kmp_int32 num_done;
1866   volatile kmp_uint32 ordered_iteration;
1867   // Dummy to retain the structure size after making ordered_iteration scalar
1868   kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1869 } dispatch_shared_info32_t;
1870 
1871 typedef struct dispatch_shared_info64 {
1872   /* chunk index under dynamic, number of idle threads under static-steal;
1873      iteration index otherwise */
1874   volatile kmp_uint64 iteration;
1875   volatile kmp_int64 num_done;
1876   volatile kmp_uint64 ordered_iteration;
1877   // Dummy to retain the structure size after making ordered_iteration scalar
1878   kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1879 } dispatch_shared_info64_t;
1880 
1881 typedef struct dispatch_shared_info {
1882   union shared_info {
1883     dispatch_shared_info32_t s32;
1884     dispatch_shared_info64_t s64;
1885   } u;
1886   volatile kmp_uint32 buffer_index;
1887   volatile kmp_int32 doacross_buf_idx; // teamwise index
1888   volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1889   kmp_int32 doacross_num_done; // count finished threads
1890 #if KMP_USE_HIER_SCHED
1891   void *hier;
1892 #endif
1893 #if KMP_USE_HWLOC
1894   // When linking with libhwloc, the ORDERED EPCC test slows down on big
1895   // machines (> 48 cores). Performance analysis showed that a cache thrash
1896   // was occurring and this padding helps alleviate the problem.
1897   char padding[64];
1898 #endif
1899 } dispatch_shared_info_t;
1900 
1901 typedef struct kmp_disp {
1902   /* Vector for ORDERED SECTION */
1903   void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1904   /* Vector for END ORDERED SECTION */
1905   void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1906 
1907   dispatch_shared_info_t *th_dispatch_sh_current;
1908   dispatch_private_info_t *th_dispatch_pr_current;
1909 
1910   dispatch_private_info_t *th_disp_buffer;
1911   kmp_uint32 th_disp_index;
1912   kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1913   volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1914   kmp_int64 *th_doacross_info; // info on loop bounds
1915 #if KMP_USE_INTERNODE_ALIGNMENT
1916   char more_padding[INTERNODE_CACHE_LINE];
1917 #endif
1918 } kmp_disp_t;
1919 
1920 /* ------------------------------------------------------------------------ */
1921 /* Barrier stuff */
1922 
1923 /* constants for barrier state update */
1924 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1925 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1926 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1927 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1928 
1929 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1930 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1931 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1932 
1933 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1934 #error "Barrier sleep bit must be smaller than barrier bump bit"
1935 #endif
1936 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1937 #error "Barrier unused bit must be smaller than barrier bump bit"
1938 #endif
1939 
1940 // Constants for release barrier wait state: currently, hierarchical only
1941 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1942 #define KMP_BARRIER_OWN_FLAG                                                   \
1943   1 // Normal state; worker waiting on own b_go flag in release
1944 #define KMP_BARRIER_PARENT_FLAG                                                \
1945   2 // Special state; worker waiting on parent's b_go flag in release
1946 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG                                         \
1947   3 // Special state; tells worker to shift from parent to own b_go
1948 #define KMP_BARRIER_SWITCHING                                                  \
1949   4 // Special state; worker resets appropriate flag on wake-up
1950 
1951 #define KMP_NOT_SAFE_TO_REAP                                                   \
1952   0 // Thread th_reap_state: not safe to reap (tasking)
1953 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1954 
1955 // The flag_type describes the storage used for the flag.
1956 enum flag_type {
1957   flag32, /**< atomic 32 bit flags */
1958   flag64, /**< 64 bit flags */
1959   atomic_flag64, /**< atomic 64 bit flags */
1960   flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */
1961   flag_unset
1962 };
1963 
1964 enum barrier_type {
1965   bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1966                            barriers if enabled) */
1967   bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1968 #if KMP_FAST_REDUCTION_BARRIER
1969   bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1970 #endif // KMP_FAST_REDUCTION_BARRIER
1971   bs_last_barrier /* Just a placeholder to mark the end */
1972 };
1973 
1974 // to work with reduction barriers just like with plain barriers
1975 #if !KMP_FAST_REDUCTION_BARRIER
1976 #define bs_reduction_barrier bs_plain_barrier
1977 #endif // KMP_FAST_REDUCTION_BARRIER
1978 
1979 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1980                            bp_linear_bar =
1981                                0, /* Single level (degenerate) tree */
1982                            bp_tree_bar =
1983                                1, /* Balanced tree with branching factor 2^n */
1984                            bp_hyper_bar = 2, /* Hypercube-embedded tree with min
1985                                                 branching factor 2^n */
1986                            bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1987                            bp_dist_bar = 4, /* Distributed barrier */
1988                            bp_last_bar /* Placeholder to mark the end */
1989 } kmp_bar_pat_e;
1990 
1991 #define KMP_BARRIER_ICV_PUSH 1
1992 
1993 /* Record for holding the values of the internal controls stack records */
1994 typedef struct kmp_internal_control {
1995   int serial_nesting_level; /* corresponds to the value of the
1996                                th_team_serialized field */
1997   kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1998                        thread) */
1999   kmp_int8
2000       bt_set; /* internal control for whether blocktime is explicitly set */
2001   int blocktime; /* internal control for blocktime */
2002 #if KMP_USE_MONITOR
2003   int bt_intervals; /* internal control for blocktime intervals */
2004 #endif
2005   int nproc; /* internal control for #threads for next parallel region (per
2006                 thread) */
2007   int thread_limit; /* internal control for thread-limit-var */
2008   int max_active_levels; /* internal control for max_active_levels */
2009   kmp_r_sched_t
2010       sched; /* internal control for runtime schedule {sched,chunk} pair */
2011   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
2012   kmp_int32 default_device; /* internal control for default device */
2013   struct kmp_internal_control *next;
2014 } kmp_internal_control_t;
2015 
2016 static inline void copy_icvs(kmp_internal_control_t *dst,
2017                              kmp_internal_control_t *src) {
2018   *dst = *src;
2019 }
2020 
2021 /* Thread barrier needs volatile barrier fields */
2022 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2023   // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2024   // uses of it). It is not explicitly aligned below, because we *don't* want
2025   // it to be padded -- instead, we fit b_go into the same cache line with
2026   // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2027   kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2028   // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2029   // same NGO store
2030   volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2031   KMP_ALIGN_CACHE volatile kmp_uint64
2032       b_arrived; // STATE => task reached synch point.
2033   kmp_uint32 *skip_per_level;
2034   kmp_uint32 my_level;
2035   kmp_int32 parent_tid;
2036   kmp_int32 old_tid;
2037   kmp_uint32 depth;
2038   struct kmp_bstate *parent_bar;
2039   kmp_team_t *team;
2040   kmp_uint64 leaf_state;
2041   kmp_uint32 nproc;
2042   kmp_uint8 base_leaf_kids;
2043   kmp_uint8 leaf_kids;
2044   kmp_uint8 offset;
2045   kmp_uint8 wait_flag;
2046   kmp_uint8 use_oncore_barrier;
2047 #if USE_DEBUGGER
2048   // The following field is intended for the debugger solely. Only the worker
2049   // thread itself accesses this field: the worker increases it by 1 when it
2050   // arrives to a barrier.
2051   KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2052 #endif /* USE_DEBUGGER */
2053 } kmp_bstate_t;
2054 
2055 union KMP_ALIGN_CACHE kmp_barrier_union {
2056   double b_align; /* use worst case alignment */
2057   char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2058   kmp_bstate_t bb;
2059 };
2060 
2061 typedef union kmp_barrier_union kmp_balign_t;
2062 
2063 /* Team barrier needs only non-volatile arrived counter */
2064 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2065   double b_align; /* use worst case alignment */
2066   char b_pad[CACHE_LINE];
2067   struct {
2068     kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2069 #if USE_DEBUGGER
2070     // The following two fields are indended for the debugger solely. Only
2071     // primary thread of the team accesses these fields: the first one is
2072     // increased by 1 when the primary thread arrives to a barrier, the second
2073     // one is increased by one when all the threads arrived.
2074     kmp_uint b_master_arrived;
2075     kmp_uint b_team_arrived;
2076 #endif
2077   };
2078 };
2079 
2080 typedef union kmp_barrier_team_union kmp_balign_team_t;
2081 
2082 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2083    threads when a condition changes.  This is to workaround an NPTL bug where
2084    padding was added to pthread_cond_t which caused the initialization routine
2085    to write outside of the structure if compiled on pre-NPTL threads.  */
2086 #if KMP_OS_WINDOWS
2087 typedef struct kmp_win32_mutex {
2088   /* The Lock */
2089   CRITICAL_SECTION cs;
2090 } kmp_win32_mutex_t;
2091 
2092 typedef struct kmp_win32_cond {
2093   /* Count of the number of waiters. */
2094   int waiters_count_;
2095 
2096   /* Serialize access to <waiters_count_> */
2097   kmp_win32_mutex_t waiters_count_lock_;
2098 
2099   /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2100   int release_count_;
2101 
2102   /* Keeps track of the current "generation" so that we don't allow */
2103   /* one thread to steal all the "releases" from the broadcast. */
2104   int wait_generation_count_;
2105 
2106   /* A manual-reset event that's used to block and release waiting threads. */
2107   HANDLE event_;
2108 } kmp_win32_cond_t;
2109 #endif
2110 
2111 #if KMP_OS_UNIX
2112 
2113 union KMP_ALIGN_CACHE kmp_cond_union {
2114   double c_align;
2115   char c_pad[CACHE_LINE];
2116   pthread_cond_t c_cond;
2117 };
2118 
2119 typedef union kmp_cond_union kmp_cond_align_t;
2120 
2121 union KMP_ALIGN_CACHE kmp_mutex_union {
2122   double m_align;
2123   char m_pad[CACHE_LINE];
2124   pthread_mutex_t m_mutex;
2125 };
2126 
2127 typedef union kmp_mutex_union kmp_mutex_align_t;
2128 
2129 #endif /* KMP_OS_UNIX */
2130 
2131 typedef struct kmp_desc_base {
2132   void *ds_stackbase;
2133   size_t ds_stacksize;
2134   int ds_stackgrow;
2135   kmp_thread_t ds_thread;
2136   volatile int ds_tid;
2137   int ds_gtid;
2138 #if KMP_OS_WINDOWS
2139   volatile int ds_alive;
2140   DWORD ds_thread_id;
2141 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2142    However, debugger support (libomp_db) cannot work with handles, because they
2143    uncomparable. For example, debugger requests info about thread with handle h.
2144    h is valid within debugger process, and meaningless within debugee process.
2145    Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2146    within debugee process, but it is a *new* handle which does *not* equal to
2147    any other handle in debugee... The only way to compare handles is convert
2148    them to system-wide ids. GetThreadId() function is available only in
2149    Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2150    on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2151    thread id by call to GetCurrentThreadId() from within the thread and save it
2152    to let libomp_db identify threads.  */
2153 #endif /* KMP_OS_WINDOWS */
2154 } kmp_desc_base_t;
2155 
2156 typedef union KMP_ALIGN_CACHE kmp_desc {
2157   double ds_align; /* use worst case alignment */
2158   char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2159   kmp_desc_base_t ds;
2160 } kmp_desc_t;
2161 
2162 typedef struct kmp_local {
2163   volatile int this_construct; /* count of single's encountered by thread */
2164   void *reduce_data;
2165 #if KMP_USE_BGET
2166   void *bget_data;
2167   void *bget_list;
2168 #if !USE_CMP_XCHG_FOR_BGET
2169 #ifdef USE_QUEUING_LOCK_FOR_BGET
2170   kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2171 #else
2172   kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2173 // bootstrap lock so we can use it at library
2174 // shutdown.
2175 #endif /* USE_LOCK_FOR_BGET */
2176 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2177 #endif /* KMP_USE_BGET */
2178 
2179   PACKED_REDUCTION_METHOD_T
2180   packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2181                               __kmpc_end_reduce*() */
2182 
2183 } kmp_local_t;
2184 
2185 #define KMP_CHECK_UPDATE(a, b)                                                 \
2186   if ((a) != (b))                                                              \
2187   (a) = (b)
2188 #define KMP_CHECK_UPDATE_SYNC(a, b)                                            \
2189   if ((a) != (b))                                                              \
2190   TCW_SYNC_PTR((a), (b))
2191 
2192 #define get__blocktime(xteam, xtid)                                            \
2193   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2194 #define get__bt_set(xteam, xtid)                                               \
2195   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2196 #if KMP_USE_MONITOR
2197 #define get__bt_intervals(xteam, xtid)                                         \
2198   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2199 #endif
2200 
2201 #define get__dynamic_2(xteam, xtid)                                            \
2202   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2203 #define get__nproc_2(xteam, xtid)                                              \
2204   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2205 #define get__sched_2(xteam, xtid)                                              \
2206   ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2207 
2208 #define set__blocktime_team(xteam, xtid, xval)                                 \
2209   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) =     \
2210        (xval))
2211 
2212 #if KMP_USE_MONITOR
2213 #define set__bt_intervals_team(xteam, xtid, xval)                              \
2214   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) =  \
2215        (xval))
2216 #endif
2217 
2218 #define set__bt_set_team(xteam, xtid, xval)                                    \
2219   (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2220 
2221 #define set__dynamic(xthread, xval)                                            \
2222   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2223 #define get__dynamic(xthread)                                                  \
2224   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2225 
2226 #define set__nproc(xthread, xval)                                              \
2227   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2228 
2229 #define set__thread_limit(xthread, xval)                                       \
2230   (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2231 
2232 #define set__max_active_levels(xthread, xval)                                  \
2233   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2234 
2235 #define get__max_active_levels(xthread)                                        \
2236   ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2237 
2238 #define set__sched(xthread, xval)                                              \
2239   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2240 
2241 #define set__proc_bind(xthread, xval)                                          \
2242   (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2243 #define get__proc_bind(xthread)                                                \
2244   ((xthread)->th.th_current_task->td_icvs.proc_bind)
2245 
2246 // OpenMP tasking data structures
2247 
2248 typedef enum kmp_tasking_mode {
2249   tskm_immediate_exec = 0,
2250   tskm_extra_barrier = 1,
2251   tskm_task_teams = 2,
2252   tskm_max = 2
2253 } kmp_tasking_mode_t;
2254 
2255 extern kmp_tasking_mode_t
2256     __kmp_tasking_mode; /* determines how/when to execute tasks */
2257 extern int __kmp_task_stealing_constraint;
2258 extern int __kmp_enable_task_throttling;
2259 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2260 // specified, defaults to 0 otherwise
2261 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2262 extern kmp_int32 __kmp_max_task_priority;
2263 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2264 extern kmp_uint64 __kmp_taskloop_min_tasks;
2265 
2266 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2267    taskdata first */
2268 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2269 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2270 
2271 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2272 // were spawned and queued since the previous barrier release.
2273 #define KMP_TASKING_ENABLED(task_team)                                         \
2274   (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2275 /*!
2276 @ingroup BASIC_TYPES
2277 @{
2278 */
2279 
2280 /*!
2281  */
2282 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2283 
2284 typedef union kmp_cmplrdata {
2285   kmp_int32 priority; /**< priority specified by user for the task */
2286   kmp_routine_entry_t
2287       destructors; /* pointer to function to invoke deconstructors of
2288                       firstprivate C++ objects */
2289   /* future data */
2290 } kmp_cmplrdata_t;
2291 
2292 /*  sizeof_kmp_task_t passed as arg to kmpc_omp_task call  */
2293 /*!
2294  */
2295 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2296   void *shareds; /**< pointer to block of pointers to shared vars   */
2297   kmp_routine_entry_t
2298       routine; /**< pointer to routine to call for executing task */
2299   kmp_int32 part_id; /**< part id for the task                          */
2300   kmp_cmplrdata_t
2301       data1; /* Two known optional additions: destructors and priority */
2302   kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2303   /* future data */
2304   /*  private vars  */
2305 } kmp_task_t;
2306 
2307 /*!
2308 @}
2309 */
2310 
2311 typedef struct kmp_taskgroup {
2312   std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2313   std::atomic<kmp_int32>
2314       cancel_request; // request for cancellation of this taskgroup
2315   struct kmp_taskgroup *parent; // parent taskgroup
2316   // Block of data to perform task reduction
2317   void *reduce_data; // reduction related info
2318   kmp_int32 reduce_num_data; // number of data items to reduce
2319   uintptr_t *gomp_data; // gomp reduction data
2320 } kmp_taskgroup_t;
2321 
2322 // forward declarations
2323 typedef union kmp_depnode kmp_depnode_t;
2324 typedef struct kmp_depnode_list kmp_depnode_list_t;
2325 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2326 
2327 // macros for checking dep flag as an integer
2328 #define KMP_DEP_IN 0x1
2329 #define KMP_DEP_OUT 0x2
2330 #define KMP_DEP_INOUT 0x3
2331 #define KMP_DEP_MTX 0x4
2332 #define KMP_DEP_SET 0x8
2333 #define KMP_DEP_ALL 0x80
2334 // Compiler sends us this info:
2335 typedef struct kmp_depend_info {
2336   kmp_intptr_t base_addr;
2337   size_t len;
2338   union {
2339     kmp_uint8 flag; // flag as an unsigned char
2340     struct { // flag as a set of 8 bits
2341       unsigned in : 1;
2342       unsigned out : 1;
2343       unsigned mtx : 1;
2344       unsigned set : 1;
2345       unsigned unused : 3;
2346       unsigned all : 1;
2347     } flags;
2348   };
2349 } kmp_depend_info_t;
2350 
2351 // Internal structures to work with task dependencies:
2352 struct kmp_depnode_list {
2353   kmp_depnode_t *node;
2354   kmp_depnode_list_t *next;
2355 };
2356 
2357 // Max number of mutexinoutset dependencies per node
2358 #define MAX_MTX_DEPS 4
2359 
2360 typedef struct kmp_base_depnode {
2361   kmp_depnode_list_t *successors; /* used under lock */
2362   kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2363   kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2364   kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2365   kmp_lock_t lock; /* guards shared fields: task, successors */
2366 #if KMP_SUPPORT_GRAPH_OUTPUT
2367   kmp_uint32 id;
2368 #endif
2369   std::atomic<kmp_int32> npredecessors;
2370   std::atomic<kmp_int32> nrefs;
2371 } kmp_base_depnode_t;
2372 
2373 union KMP_ALIGN_CACHE kmp_depnode {
2374   double dn_align; /* use worst case alignment */
2375   char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2376   kmp_base_depnode_t dn;
2377 };
2378 
2379 struct kmp_dephash_entry {
2380   kmp_intptr_t addr;
2381   kmp_depnode_t *last_out;
2382   kmp_depnode_list_t *last_set;
2383   kmp_depnode_list_t *prev_set;
2384   kmp_uint8 last_flag;
2385   kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2386   kmp_dephash_entry_t *next_in_bucket;
2387 };
2388 
2389 typedef struct kmp_dephash {
2390   kmp_dephash_entry_t **buckets;
2391   size_t size;
2392   kmp_depnode_t *last_all;
2393   size_t generation;
2394   kmp_uint32 nelements;
2395   kmp_uint32 nconflicts;
2396 } kmp_dephash_t;
2397 
2398 typedef struct kmp_task_affinity_info {
2399   kmp_intptr_t base_addr;
2400   size_t len;
2401   struct {
2402     bool flag1 : 1;
2403     bool flag2 : 1;
2404     kmp_int32 reserved : 30;
2405   } flags;
2406 } kmp_task_affinity_info_t;
2407 
2408 typedef enum kmp_event_type_t {
2409   KMP_EVENT_UNINITIALIZED = 0,
2410   KMP_EVENT_ALLOW_COMPLETION = 1
2411 } kmp_event_type_t;
2412 
2413 typedef struct {
2414   kmp_event_type_t type;
2415   kmp_tas_lock_t lock;
2416   union {
2417     kmp_task_t *task;
2418   } ed;
2419 } kmp_event_t;
2420 
2421 #ifdef BUILD_TIED_TASK_STACK
2422 
2423 /* Tied Task stack definitions */
2424 typedef struct kmp_stack_block {
2425   kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2426   struct kmp_stack_block *sb_next;
2427   struct kmp_stack_block *sb_prev;
2428 } kmp_stack_block_t;
2429 
2430 typedef struct kmp_task_stack {
2431   kmp_stack_block_t ts_first_block; // first block of stack entries
2432   kmp_taskdata_t **ts_top; // pointer to the top of stack
2433   kmp_int32 ts_entries; // number of entries on the stack
2434 } kmp_task_stack_t;
2435 
2436 #endif // BUILD_TIED_TASK_STACK
2437 
2438 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2439   /* Compiler flags */ /* Total compiler flags must be 16 bits */
2440   unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2441   unsigned final : 1; /* task is final(1) so execute immediately */
2442   unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2443                               code path */
2444   unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2445                                      invoke destructors from the runtime */
2446   unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2447                          context of the RTL) */
2448   unsigned priority_specified : 1; /* set if the compiler provides priority
2449                                       setting for the task */
2450   unsigned detachable : 1; /* 1 == can detach */
2451   unsigned hidden_helper : 1; /* 1 == hidden helper task */
2452   unsigned reserved : 8; /* reserved for compiler use */
2453 
2454   /* Library flags */ /* Total library flags must be 16 bits */
2455   unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2456   unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2457   unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2458   // (1) or may be deferred (0)
2459   unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2460   // (0) [>= 2 threads]
2461   /* If either team_serial or tasking_ser is set, task team may be NULL */
2462   /* Task State Flags: */
2463   unsigned started : 1; /* 1==started, 0==not started     */
2464   unsigned executing : 1; /* 1==executing, 0==not executing */
2465   unsigned complete : 1; /* 1==complete, 0==not complete   */
2466   unsigned freed : 1; /* 1==freed, 0==allocated        */
2467   unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2468   unsigned reserved31 : 7; /* reserved for library use */
2469 
2470 } kmp_tasking_flags_t;
2471 
2472 struct kmp_taskdata { /* aligned during dynamic allocation       */
2473   kmp_int32 td_task_id; /* id, assigned by debugger                */
2474   kmp_tasking_flags_t td_flags; /* task flags                              */
2475   kmp_team_t *td_team; /* team for this task                      */
2476   kmp_info_p *td_alloc_thread; /* thread that allocated data structures   */
2477   /* Currently not used except for perhaps IDB */
2478   kmp_taskdata_t *td_parent; /* parent task                             */
2479   kmp_int32 td_level; /* task nesting level                      */
2480   std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2481   ident_t *td_ident; /* task identifier                         */
2482   // Taskwait data.
2483   ident_t *td_taskwait_ident;
2484   kmp_uint32 td_taskwait_counter;
2485   kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2486   KMP_ALIGN_CACHE kmp_internal_control_t
2487       td_icvs; /* Internal control variables for the task */
2488   KMP_ALIGN_CACHE std::atomic<kmp_int32>
2489       td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2490                                    deallocated */
2491   std::atomic<kmp_int32>
2492       td_incomplete_child_tasks; /* Child tasks not yet complete */
2493   kmp_taskgroup_t
2494       *td_taskgroup; // Each task keeps pointer to its current taskgroup
2495   kmp_dephash_t
2496       *td_dephash; // Dependencies for children tasks are tracked from here
2497   kmp_depnode_t
2498       *td_depnode; // Pointer to graph node if this task has dependencies
2499   kmp_task_team_t *td_task_team;
2500   size_t td_size_alloc; // Size of task structure, including shareds etc.
2501 #if defined(KMP_GOMP_COMPAT)
2502   // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2503   kmp_int32 td_size_loop_bounds;
2504 #endif
2505   kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2506 #if defined(KMP_GOMP_COMPAT)
2507   // GOMP sends in a copy function for copy constructors
2508   void (*td_copy_func)(void *, void *);
2509 #endif
2510   kmp_event_t td_allow_completion_event;
2511 #if OMPT_SUPPORT
2512   ompt_task_info_t ompt_task_info;
2513 #endif
2514 }; // struct kmp_taskdata
2515 
2516 // Make sure padding above worked
2517 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2518 
2519 // Data for task team but per thread
2520 typedef struct kmp_base_thread_data {
2521   kmp_info_p *td_thr; // Pointer back to thread info
2522   // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2523   // queued?
2524   kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2525   kmp_taskdata_t *
2526       *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2527   kmp_int32 td_deque_size; // Size of deck
2528   kmp_uint32 td_deque_head; // Head of deque (will wrap)
2529   kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2530   kmp_int32 td_deque_ntasks; // Number of tasks in deque
2531   // GEH: shouldn't this be volatile since used in while-spin?
2532   kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2533 #ifdef BUILD_TIED_TASK_STACK
2534   kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2535 // scheduling constraint
2536 #endif // BUILD_TIED_TASK_STACK
2537 } kmp_base_thread_data_t;
2538 
2539 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2540 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2541 
2542 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2543 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2544 
2545 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2546   kmp_base_thread_data_t td;
2547   double td_align; /* use worst case alignment */
2548   char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2549 } kmp_thread_data_t;
2550 
2551 // Data for task teams which are used when tasking is enabled for the team
2552 typedef struct kmp_base_task_team {
2553   kmp_bootstrap_lock_t
2554       tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2555   /* must be bootstrap lock since used at library shutdown*/
2556   kmp_task_team_t *tt_next; /* For linking the task team free list */
2557   kmp_thread_data_t
2558       *tt_threads_data; /* Array of per-thread structures for task team */
2559   /* Data survives task team deallocation */
2560   kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2561                                executing this team? */
2562   /* TRUE means tt_threads_data is set up and initialized */
2563   kmp_int32 tt_nproc; /* #threads in team           */
2564   kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2565   kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2566   kmp_int32 tt_untied_task_encountered;
2567   // There is hidden helper thread encountered in this task team so that we must
2568   // wait when waiting on task team
2569   kmp_int32 tt_hidden_helper_task_encountered;
2570 
2571   KMP_ALIGN_CACHE
2572   std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2573 
2574   KMP_ALIGN_CACHE
2575   volatile kmp_uint32
2576       tt_active; /* is the team still actively executing tasks */
2577 } kmp_base_task_team_t;
2578 
2579 union KMP_ALIGN_CACHE kmp_task_team {
2580   kmp_base_task_team_t tt;
2581   double tt_align; /* use worst case alignment */
2582   char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2583 };
2584 
2585 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2586 // Free lists keep same-size free memory slots for fast memory allocation
2587 // routines
2588 typedef struct kmp_free_list {
2589   void *th_free_list_self; // Self-allocated tasks free list
2590   void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2591   // threads
2592   void *th_free_list_other; // Non-self free list (to be returned to owner's
2593   // sync list)
2594 } kmp_free_list_t;
2595 #endif
2596 #if KMP_NESTED_HOT_TEAMS
2597 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2598 // are not put in teams pool, and they don't put threads in threads pool.
2599 typedef struct kmp_hot_team_ptr {
2600   kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2601   kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2602 } kmp_hot_team_ptr_t;
2603 #endif
2604 typedef struct kmp_teams_size {
2605   kmp_int32 nteams; // number of teams in a league
2606   kmp_int32 nth; // number of threads in each team of the league
2607 } kmp_teams_size_t;
2608 
2609 // This struct stores a thread that acts as a "root" for a contention
2610 // group. Contention groups are rooted at kmp_root threads, but also at
2611 // each primary thread of each team created in the teams construct.
2612 // This struct therefore also stores a thread_limit associated with
2613 // that contention group, and a counter to track the number of threads
2614 // active in that contention group. Each thread has a list of these: CG
2615 // root threads have an entry in their list in which cg_root refers to
2616 // the thread itself, whereas other workers in the CG will have a
2617 // single entry where cg_root is same as the entry containing their CG
2618 // root. When a thread encounters a teams construct, it will add a new
2619 // entry to the front of its list, because it now roots a new CG.
2620 typedef struct kmp_cg_root {
2621   kmp_info_p *cg_root; // "root" thread for a contention group
2622   // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2623   // thread_limit clause for teams primary threads
2624   kmp_int32 cg_thread_limit;
2625   kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2626   struct kmp_cg_root *up; // pointer to higher level CG root in list
2627 } kmp_cg_root_t;
2628 
2629 // OpenMP thread data structures
2630 
2631 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2632   /* Start with the readonly data which is cache aligned and padded. This is
2633      written before the thread starts working by the primary thread. Uber
2634      masters may update themselves later. Usage does not consider serialized
2635      regions.  */
2636   kmp_desc_t th_info;
2637   kmp_team_p *th_team; /* team we belong to */
2638   kmp_root_p *th_root; /* pointer to root of task hierarchy */
2639   kmp_info_p *th_next_pool; /* next available thread in the pool */
2640   kmp_disp_t *th_dispatch; /* thread's dispatch data */
2641   int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2642 
2643   /* The following are cached from the team info structure */
2644   /* TODO use these in more places as determined to be needed via profiling */
2645   int th_team_nproc; /* number of threads in a team */
2646   kmp_info_p *th_team_master; /* the team's primary thread */
2647   int th_team_serialized; /* team is serialized */
2648   microtask_t th_teams_microtask; /* save entry address for teams construct */
2649   int th_teams_level; /* save initial level of teams construct */
2650 /* it is 0 on device but may be any on host */
2651 
2652 /* The blocktime info is copied from the team struct to the thread struct */
2653 /* at the start of a barrier, and the values stored in the team are used  */
2654 /* at points in the code where the team struct is no longer guaranteed    */
2655 /* to exist (from the POV of worker threads).                             */
2656 #if KMP_USE_MONITOR
2657   int th_team_bt_intervals;
2658   int th_team_bt_set;
2659 #else
2660   kmp_uint64 th_team_bt_intervals;
2661 #endif
2662 
2663 #if KMP_AFFINITY_SUPPORTED
2664   kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2665 #endif
2666   omp_allocator_handle_t th_def_allocator; /* default allocator */
2667   /* The data set by the primary thread at reinit, then R/W by the worker */
2668   KMP_ALIGN_CACHE int
2669       th_set_nproc; /* if > 0, then only use this request for the next fork */
2670 #if KMP_NESTED_HOT_TEAMS
2671   kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2672 #endif
2673   kmp_proc_bind_t
2674       th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2675   kmp_teams_size_t
2676       th_teams_size; /* number of teams/threads in teams construct */
2677 #if KMP_AFFINITY_SUPPORTED
2678   int th_current_place; /* place currently bound to */
2679   int th_new_place; /* place to bind to in par reg */
2680   int th_first_place; /* first place in partition */
2681   int th_last_place; /* last place in partition */
2682 #endif
2683   int th_prev_level; /* previous level for affinity format */
2684   int th_prev_num_threads; /* previous num_threads for affinity format */
2685 #if USE_ITT_BUILD
2686   kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2687   kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2688   kmp_uint64 th_frame_time; /* frame timestamp */
2689 #endif /* USE_ITT_BUILD */
2690   kmp_local_t th_local;
2691   struct private_common *th_pri_head;
2692 
2693   /* Now the data only used by the worker (after initial allocation) */
2694   /* TODO the first serial team should actually be stored in the info_t
2695      structure.  this will help reduce initial allocation overhead */
2696   KMP_ALIGN_CACHE kmp_team_p
2697       *th_serial_team; /*serialized team held in reserve*/
2698 
2699 #if OMPT_SUPPORT
2700   ompt_thread_info_t ompt_thread_info;
2701 #endif
2702 
2703   /* The following are also read by the primary thread during reinit */
2704   struct common_table *th_pri_common;
2705 
2706   volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2707   /* while awaiting queuing lock acquire */
2708 
2709   volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2710   flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2711 
2712   ident_t *th_ident;
2713   unsigned th_x; // Random number generator data
2714   unsigned th_a; // Random number generator data
2715 
2716   /* Tasking-related data for the thread */
2717   kmp_task_team_t *th_task_team; // Task team struct
2718   kmp_taskdata_t *th_current_task; // Innermost Task being executed
2719   kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2720   kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2721   // at nested levels
2722   kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2723   kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2724   kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2725   // tasking, thus safe to reap
2726 
2727   /* More stuff for keeping track of active/sleeping threads (this part is
2728      written by the worker thread) */
2729   kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2730   int th_active; // ! sleeping; 32 bits for TCR/TCW
2731   std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2732   // 0 = not used in team; 1 = used in team;
2733   // 2 = transitioning to not used in team; 3 = transitioning to used in team
2734   struct cons_header *th_cons; // used for consistency check
2735 #if KMP_USE_HIER_SCHED
2736   // used for hierarchical scheduling
2737   kmp_hier_private_bdata_t *th_hier_bar_data;
2738 #endif
2739 
2740   /* Add the syncronizing data which is cache aligned and padded. */
2741   KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2742 
2743   KMP_ALIGN_CACHE volatile kmp_int32
2744       th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2745 
2746 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2747 #define NUM_LISTS 4
2748   kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2749 // allocation routines
2750 #endif
2751 
2752 #if KMP_OS_WINDOWS
2753   kmp_win32_cond_t th_suspend_cv;
2754   kmp_win32_mutex_t th_suspend_mx;
2755   std::atomic<int> th_suspend_init;
2756 #endif
2757 #if KMP_OS_UNIX
2758   kmp_cond_align_t th_suspend_cv;
2759   kmp_mutex_align_t th_suspend_mx;
2760   std::atomic<int> th_suspend_init_count;
2761 #endif
2762 
2763 #if USE_ITT_BUILD
2764   kmp_itt_mark_t th_itt_mark_single;
2765 // alignment ???
2766 #endif /* USE_ITT_BUILD */
2767 #if KMP_STATS_ENABLED
2768   kmp_stats_list *th_stats;
2769 #endif
2770 #if KMP_OS_UNIX
2771   std::atomic<bool> th_blocking;
2772 #endif
2773   kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2774 } kmp_base_info_t;
2775 
2776 typedef union KMP_ALIGN_CACHE kmp_info {
2777   double th_align; /* use worst case alignment */
2778   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2779   kmp_base_info_t th;
2780 } kmp_info_t;
2781 
2782 // OpenMP thread team data structures
2783 
2784 typedef struct kmp_base_data {
2785   volatile kmp_uint32 t_value;
2786 } kmp_base_data_t;
2787 
2788 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2789   double dt_align; /* use worst case alignment */
2790   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2791   kmp_base_data_t dt;
2792 } kmp_sleep_team_t;
2793 
2794 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2795   double dt_align; /* use worst case alignment */
2796   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2797   kmp_base_data_t dt;
2798 } kmp_ordered_team_t;
2799 
2800 typedef int (*launch_t)(int gtid);
2801 
2802 /* Minimum number of ARGV entries to malloc if necessary */
2803 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2804 
2805 // Set up how many argv pointers will fit in cache lines containing
2806 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2807 // larger value for more space between the primary write/worker read section and
2808 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2809 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2810 #define KMP_INLINE_ARGV_BYTES                                                  \
2811   (4 * CACHE_LINE -                                                            \
2812    ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) +               \
2813      sizeof(kmp_int16) + sizeof(kmp_uint32)) %                                 \
2814     CACHE_LINE))
2815 #else
2816 #define KMP_INLINE_ARGV_BYTES                                                  \
2817   (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2818 #endif
2819 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2820 
2821 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2822   // Synchronization Data
2823   // ---------------------------------------------------------------------------
2824   KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2825   kmp_balign_team_t t_bar[bs_last_barrier];
2826   std::atomic<int> t_construct; // count of single directive encountered by team
2827   char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2828 
2829   // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2830   std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2831   std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2832 
2833   // Primary thread only
2834   // ---------------------------------------------------------------------------
2835   KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2836   int t_master_this_cons; // "this_construct" single counter of primary thread
2837   // in parent team
2838   ident_t *t_ident; // if volatile, have to change too much other crud to
2839   // volatile too
2840   kmp_team_p *t_parent; // parent team
2841   kmp_team_p *t_next_pool; // next free team in the team pool
2842   kmp_disp_t *t_dispatch; // thread's dispatch data
2843   kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2844   kmp_proc_bind_t t_proc_bind; // bind type for par region
2845 #if USE_ITT_BUILD
2846   kmp_uint64 t_region_time; // region begin timestamp
2847 #endif /* USE_ITT_BUILD */
2848 
2849   // Primary thread write, workers read
2850   // --------------------------------------------------------------------------
2851   KMP_ALIGN_CACHE void **t_argv;
2852   int t_argc;
2853   int t_nproc; // number of threads in team
2854   microtask_t t_pkfn;
2855   launch_t t_invoke; // procedure to launch the microtask
2856 
2857 #if OMPT_SUPPORT
2858   ompt_team_info_t ompt_team_info;
2859   ompt_lw_taskteam_t *ompt_serialized_team_info;
2860 #endif
2861 
2862 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2863   kmp_int8 t_fp_control_saved;
2864   kmp_int8 t_pad2b;
2865   kmp_int16 t_x87_fpu_control_word; // FP control regs
2866   kmp_uint32 t_mxcsr;
2867 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2868 
2869   void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2870 
2871   KMP_ALIGN_CACHE kmp_info_t **t_threads;
2872   kmp_taskdata_t
2873       *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2874   int t_level; // nested parallel level
2875 
2876   KMP_ALIGN_CACHE int t_max_argc;
2877   int t_max_nproc; // max threads this team can handle (dynamically expandable)
2878   int t_serialized; // levels deep of serialized teams
2879   dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2880   int t_id; // team's id, assigned by debugger.
2881   int t_active_level; // nested active parallel level
2882   kmp_r_sched_t t_sched; // run-time schedule for the team
2883 #if KMP_AFFINITY_SUPPORTED
2884   int t_first_place; // first & last place in parent thread's partition.
2885   int t_last_place; // Restore these values to primary thread after par region.
2886 #endif // KMP_AFFINITY_SUPPORTED
2887   int t_display_affinity;
2888   int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2889   // omp_set_num_threads() call
2890   omp_allocator_handle_t t_def_allocator; /* default allocator */
2891 
2892 // Read/write by workers as well
2893 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2894   // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2895   // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2896   // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2897   // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2898   char dummy_padding[1024];
2899 #endif
2900   // Internal control stack for additional nested teams.
2901   KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2902   // for SERIALIZED teams nested 2 or more levels deep
2903   // typed flag to store request state of cancellation
2904   std::atomic<kmp_int32> t_cancel_request;
2905   int t_master_active; // save on fork, restore on join
2906   void *t_copypriv_data; // team specific pointer to copyprivate data array
2907 #if KMP_OS_WINDOWS
2908   std::atomic<kmp_uint32> t_copyin_counter;
2909 #endif
2910 #if USE_ITT_BUILD
2911   void *t_stack_id; // team specific stack stitching id (for ittnotify)
2912 #endif /* USE_ITT_BUILD */
2913   distributedBarrier *b; // Distributed barrier data associated with team
2914 } kmp_base_team_t;
2915 
2916 union KMP_ALIGN_CACHE kmp_team {
2917   kmp_base_team_t t;
2918   double t_align; /* use worst case alignment */
2919   char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2920 };
2921 
2922 typedef union KMP_ALIGN_CACHE kmp_time_global {
2923   double dt_align; /* use worst case alignment */
2924   char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2925   kmp_base_data_t dt;
2926 } kmp_time_global_t;
2927 
2928 typedef struct kmp_base_global {
2929   /* cache-aligned */
2930   kmp_time_global_t g_time;
2931 
2932   /* non cache-aligned */
2933   volatile int g_abort;
2934   volatile int g_done;
2935 
2936   int g_dynamic;
2937   enum dynamic_mode g_dynamic_mode;
2938 } kmp_base_global_t;
2939 
2940 typedef union KMP_ALIGN_CACHE kmp_global {
2941   kmp_base_global_t g;
2942   double g_align; /* use worst case alignment */
2943   char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2944 } kmp_global_t;
2945 
2946 typedef struct kmp_base_root {
2947   // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2948   // (r_in_parallel>= 0)
2949   // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2950   // the synch overhead or keeping r_active
2951   volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2952   // keeps a count of active parallel regions per root
2953   std::atomic<int> r_in_parallel;
2954   // GEH: This is misnamed, should be r_active_levels
2955   kmp_team_t *r_root_team;
2956   kmp_team_t *r_hot_team;
2957   kmp_info_t *r_uber_thread;
2958   kmp_lock_t r_begin_lock;
2959   volatile int r_begin;
2960   int r_blocktime; /* blocktime for this root and descendants */
2961 #if KMP_AFFINITY_SUPPORTED
2962   int r_affinity_assigned;
2963 #endif // KMP_AFFINITY_SUPPORTED
2964 } kmp_base_root_t;
2965 
2966 typedef union KMP_ALIGN_CACHE kmp_root {
2967   kmp_base_root_t r;
2968   double r_align; /* use worst case alignment */
2969   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2970 } kmp_root_t;
2971 
2972 struct fortran_inx_info {
2973   kmp_int32 data;
2974 };
2975 
2976 /* ------------------------------------------------------------------------ */
2977 
2978 extern int __kmp_settings;
2979 extern int __kmp_duplicate_library_ok;
2980 #if USE_ITT_BUILD
2981 extern int __kmp_forkjoin_frames;
2982 extern int __kmp_forkjoin_frames_mode;
2983 #endif
2984 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2985 extern int __kmp_determ_red;
2986 
2987 #ifdef KMP_DEBUG
2988 extern int kmp_a_debug;
2989 extern int kmp_b_debug;
2990 extern int kmp_c_debug;
2991 extern int kmp_d_debug;
2992 extern int kmp_e_debug;
2993 extern int kmp_f_debug;
2994 #endif /* KMP_DEBUG */
2995 
2996 /* For debug information logging using rotating buffer */
2997 #define KMP_DEBUG_BUF_LINES_INIT 512
2998 #define KMP_DEBUG_BUF_LINES_MIN 1
2999 
3000 #define KMP_DEBUG_BUF_CHARS_INIT 128
3001 #define KMP_DEBUG_BUF_CHARS_MIN 2
3002 
3003 extern int
3004     __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3005 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3006 extern int
3007     __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3008 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3009                                       entry pointer */
3010 
3011 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3012 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3013                                               printed in buffer so far */
3014 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3015                                           recommended in warnings */
3016 /* end rotating debug buffer */
3017 
3018 #ifdef KMP_DEBUG
3019 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3020 
3021 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3022 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3023 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3024 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3025 extern int __kmp_par_range_lb;
3026 extern int __kmp_par_range_ub;
3027 #endif
3028 
3029 /* For printing out dynamic storage map for threads and teams */
3030 extern int
3031     __kmp_storage_map; /* True means print storage map for threads and teams */
3032 extern int __kmp_storage_map_verbose; /* True means storage map includes
3033                                          placement info */
3034 extern int __kmp_storage_map_verbose_specified;
3035 
3036 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3037 extern kmp_cpuinfo_t __kmp_cpuinfo;
3038 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3039 #else
3040 static inline bool __kmp_is_hybrid_cpu() { return false; }
3041 #endif
3042 
3043 extern volatile int __kmp_init_serial;
3044 extern volatile int __kmp_init_gtid;
3045 extern volatile int __kmp_init_common;
3046 extern volatile int __kmp_init_middle;
3047 extern volatile int __kmp_init_parallel;
3048 #if KMP_USE_MONITOR
3049 extern volatile int __kmp_init_monitor;
3050 #endif
3051 extern volatile int __kmp_init_user_locks;
3052 extern volatile int __kmp_init_hidden_helper_threads;
3053 extern int __kmp_init_counter;
3054 extern int __kmp_root_counter;
3055 extern int __kmp_version;
3056 
3057 /* list of address of allocated caches for commons */
3058 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3059 
3060 /* Barrier algorithm types and options */
3061 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3062 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3063 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3064 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3065 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3066 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3067 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3068 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3069 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3070 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3071 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3072 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3073 
3074 /* Global Locks */
3075 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3076 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3077 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3078 extern kmp_bootstrap_lock_t
3079     __kmp_exit_lock; /* exit() is not always thread-safe */
3080 #if KMP_USE_MONITOR
3081 extern kmp_bootstrap_lock_t
3082     __kmp_monitor_lock; /* control monitor thread creation */
3083 #endif
3084 extern kmp_bootstrap_lock_t
3085     __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3086                              __kmp_threads expansion to co-exist */
3087 
3088 extern kmp_lock_t __kmp_global_lock; /* control OS/global access  */
3089 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access  */
3090 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3091 
3092 extern enum library_type __kmp_library;
3093 
3094 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3095 extern enum sched_type __kmp_static; /* default static scheduling method */
3096 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3097 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3098 extern int __kmp_chunk; /* default runtime chunk size */
3099 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3100 
3101 extern size_t __kmp_stksize; /* stack size per thread         */
3102 #if KMP_USE_MONITOR
3103 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3104 #endif
3105 extern size_t __kmp_stkoffset; /* stack offset per thread       */
3106 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3107 
3108 extern size_t
3109     __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3110 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3111 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3112 extern int __kmp_env_checks; /* was KMP_CHECKS specified?    */
3113 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3114 extern int __kmp_generate_warnings; /* should we issue warnings? */
3115 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3116 
3117 #ifdef DEBUG_SUSPEND
3118 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3119 #endif
3120 
3121 extern kmp_int32 __kmp_use_yield;
3122 extern kmp_int32 __kmp_use_yield_exp_set;
3123 extern kmp_uint32 __kmp_yield_init;
3124 extern kmp_uint32 __kmp_yield_next;
3125 extern kmp_uint64 __kmp_pause_init;
3126 
3127 /* ------------------------------------------------------------------------- */
3128 extern int __kmp_allThreadsSpecified;
3129 
3130 extern size_t __kmp_align_alloc;
3131 /* following data protected by initialization routines */
3132 extern int __kmp_xproc; /* number of processors in the system */
3133 extern int __kmp_avail_proc; /* number of processors available to the process */
3134 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3135 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3136 // maximum total number of concurrently-existing threads on device
3137 extern int __kmp_max_nth;
3138 // maximum total number of concurrently-existing threads in a contention group
3139 extern int __kmp_cg_max_nth;
3140 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3141 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3142                                       __kmp_root */
3143 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3144                                    region a la OMP_NUM_THREADS */
3145 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3146                                       initialization */
3147 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3148                                  used (fixed) */
3149 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3150                                (__kmpc_threadprivate_cached()) */
3151 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3152                                     blocking (env setting) */
3153 #if KMP_USE_MONITOR
3154 extern int
3155     __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3156 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3157                                   blocking */
3158 #endif
3159 #ifdef KMP_ADJUST_BLOCKTIME
3160 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3161 #endif /* KMP_ADJUST_BLOCKTIME */
3162 #ifdef KMP_DFLT_NTH_CORES
3163 extern int __kmp_ncores; /* Total number of cores for threads placement */
3164 #endif
3165 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3166 extern int __kmp_abort_delay;
3167 
3168 extern int __kmp_need_register_atfork_specified;
3169 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3170                                           to install fork handler */
3171 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3172                                0 - not set, will be set at runtime
3173                                1 - using stack search
3174                                2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3175                                    X*) or TlsGetValue(Windows* OS))
3176                                3 - static TLS (__declspec(thread) __kmp_gtid),
3177                                    Linux* OS .so only.  */
3178 extern int
3179     __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3180 #ifdef KMP_TDATA_GTID
3181 extern KMP_THREAD_LOCAL int __kmp_gtid;
3182 #endif
3183 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3184 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3185 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3186 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3187 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3188 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3189 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3190 
3191 // max_active_levels for nested parallelism enabled by default via
3192 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3193 extern int __kmp_dflt_max_active_levels;
3194 // Indicates whether value of __kmp_dflt_max_active_levels was already
3195 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3196 extern bool __kmp_dflt_max_active_levels_set;
3197 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3198                                           concurrent execution per team */
3199 #if KMP_NESTED_HOT_TEAMS
3200 extern int __kmp_hot_teams_mode;
3201 extern int __kmp_hot_teams_max_level;
3202 #endif
3203 
3204 #if KMP_OS_LINUX
3205 extern enum clock_function_type __kmp_clock_function;
3206 extern int __kmp_clock_function_param;
3207 #endif /* KMP_OS_LINUX */
3208 
3209 #if KMP_MIC_SUPPORTED
3210 extern enum mic_type __kmp_mic_type;
3211 #endif
3212 
3213 #ifdef USE_LOAD_BALANCE
3214 extern double __kmp_load_balance_interval; // load balance algorithm interval
3215 #endif /* USE_LOAD_BALANCE */
3216 
3217 // OpenMP 3.1 - Nested num threads array
3218 typedef struct kmp_nested_nthreads_t {
3219   int *nth;
3220   int size;
3221   int used;
3222 } kmp_nested_nthreads_t;
3223 
3224 extern kmp_nested_nthreads_t __kmp_nested_nth;
3225 
3226 #if KMP_USE_ADAPTIVE_LOCKS
3227 
3228 // Parameters for the speculative lock backoff system.
3229 struct kmp_adaptive_backoff_params_t {
3230   // Number of soft retries before it counts as a hard retry.
3231   kmp_uint32 max_soft_retries;
3232   // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3233   // the right
3234   kmp_uint32 max_badness;
3235 };
3236 
3237 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3238 
3239 #if KMP_DEBUG_ADAPTIVE_LOCKS
3240 extern const char *__kmp_speculative_statsfile;
3241 #endif
3242 
3243 #endif // KMP_USE_ADAPTIVE_LOCKS
3244 
3245 extern int __kmp_display_env; /* TRUE or FALSE */
3246 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3247 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3248 extern int __kmp_nteams;
3249 extern int __kmp_teams_thread_limit;
3250 
3251 /* ------------------------------------------------------------------------- */
3252 
3253 /* the following are protected by the fork/join lock */
3254 /* write: lock  read: anytime */
3255 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3256 /* read/write: lock */
3257 extern volatile kmp_team_t *__kmp_team_pool;
3258 extern volatile kmp_info_t *__kmp_thread_pool;
3259 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3260 
3261 // total num threads reachable from some root thread including all root threads
3262 extern volatile int __kmp_nth;
3263 /* total number of threads reachable from some root thread including all root
3264    threads, and those in the thread pool */
3265 extern volatile int __kmp_all_nth;
3266 extern std::atomic<int> __kmp_thread_pool_active_nth;
3267 
3268 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3269 /* end data protected by fork/join lock */
3270 /* ------------------------------------------------------------------------- */
3271 
3272 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3273 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3274 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3275 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3276 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3277 
3278 // AT: Which way is correct?
3279 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3280 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3281 #define __kmp_get_team_num_threads(gtid)                                       \
3282   (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3283 
3284 static inline bool KMP_UBER_GTID(int gtid) {
3285   KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3286   KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3287   return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3288           __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3289 }
3290 
3291 static inline int __kmp_tid_from_gtid(int gtid) {
3292   KMP_DEBUG_ASSERT(gtid >= 0);
3293   return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3294 }
3295 
3296 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3297   KMP_DEBUG_ASSERT(tid >= 0 && team);
3298   return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3299 }
3300 
3301 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3302   KMP_DEBUG_ASSERT(thr);
3303   return thr->th.th_info.ds.ds_gtid;
3304 }
3305 
3306 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3307   KMP_DEBUG_ASSERT(gtid >= 0);
3308   return __kmp_threads[gtid];
3309 }
3310 
3311 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3312   KMP_DEBUG_ASSERT(gtid >= 0);
3313   return __kmp_threads[gtid]->th.th_team;
3314 }
3315 
3316 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3317   if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3318     KMP_FATAL(ThreadIdentInvalid);
3319 }
3320 
3321 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3322 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3323 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3324 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3325 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3326 #endif
3327 
3328 #if KMP_HAVE_UMWAIT
3329 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3330 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3331 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3332 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3333 #endif
3334 
3335 /* ------------------------------------------------------------------------- */
3336 
3337 extern kmp_global_t __kmp_global; /* global status */
3338 
3339 extern kmp_info_t __kmp_monitor;
3340 // For Debugging Support Library
3341 extern std::atomic<kmp_int32> __kmp_team_counter;
3342 // For Debugging Support Library
3343 extern std::atomic<kmp_int32> __kmp_task_counter;
3344 
3345 #if USE_DEBUGGER
3346 #define _KMP_GEN_ID(counter)                                                   \
3347   (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3348 #else
3349 #define _KMP_GEN_ID(counter) (~0)
3350 #endif /* USE_DEBUGGER */
3351 
3352 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3353 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3354 
3355 /* ------------------------------------------------------------------------ */
3356 
3357 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3358                                          size_t size, char const *format, ...);
3359 
3360 extern void __kmp_serial_initialize(void);
3361 extern void __kmp_middle_initialize(void);
3362 extern void __kmp_parallel_initialize(void);
3363 
3364 extern void __kmp_internal_begin(void);
3365 extern void __kmp_internal_end_library(int gtid);
3366 extern void __kmp_internal_end_thread(int gtid);
3367 extern void __kmp_internal_end_atexit(void);
3368 extern void __kmp_internal_end_dtor(void);
3369 extern void __kmp_internal_end_dest(void *);
3370 
3371 extern int __kmp_register_root(int initial_thread);
3372 extern void __kmp_unregister_root(int gtid);
3373 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3374 
3375 extern int __kmp_ignore_mppbeg(void);
3376 extern int __kmp_ignore_mppend(void);
3377 
3378 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3379 extern void __kmp_exit_single(int gtid);
3380 
3381 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3382 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3383 
3384 #ifdef USE_LOAD_BALANCE
3385 extern int __kmp_get_load_balance(int);
3386 #endif
3387 
3388 extern int __kmp_get_global_thread_id(void);
3389 extern int __kmp_get_global_thread_id_reg(void);
3390 extern void __kmp_exit_thread(int exit_status);
3391 extern void __kmp_abort(char const *format, ...);
3392 extern void __kmp_abort_thread(void);
3393 KMP_NORETURN extern void __kmp_abort_process(void);
3394 extern void __kmp_warn(char const *format, ...);
3395 
3396 extern void __kmp_set_num_threads(int new_nth, int gtid);
3397 
3398 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3399 // registered.
3400 static inline kmp_info_t *__kmp_entry_thread() {
3401   int gtid = __kmp_entry_gtid();
3402 
3403   return __kmp_threads[gtid];
3404 }
3405 
3406 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3407 extern int __kmp_get_max_active_levels(int gtid);
3408 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3409 extern int __kmp_get_team_size(int gtid, int level);
3410 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3411 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3412 
3413 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3414 extern void __kmp_init_random(kmp_info_t *thread);
3415 
3416 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3417 extern void __kmp_adjust_num_threads(int new_nproc);
3418 extern void __kmp_check_stksize(size_t *val);
3419 
3420 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3421 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3422 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3423 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3424 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3425 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3426 
3427 #if USE_FAST_MEMORY
3428 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3429                                   size_t size KMP_SRC_LOC_DECL);
3430 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3431 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3432 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3433 #define __kmp_fast_allocate(this_thr, size)                                    \
3434   ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3435 #define __kmp_fast_free(this_thr, ptr)                                         \
3436   ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3437 #endif
3438 
3439 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3440 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3441                                   size_t elsize KMP_SRC_LOC_DECL);
3442 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3443                                    size_t size KMP_SRC_LOC_DECL);
3444 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3445 #define __kmp_thread_malloc(th, size)                                          \
3446   ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3447 #define __kmp_thread_calloc(th, nelem, elsize)                                 \
3448   ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3449 #define __kmp_thread_realloc(th, ptr, size)                                    \
3450   ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3451 #define __kmp_thread_free(th, ptr)                                             \
3452   ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3453 
3454 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3455 #define KMP_INTERNAL_FREE(p) free(p)
3456 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3457 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3458 
3459 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3460 
3461 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3462                                  kmp_proc_bind_t proc_bind);
3463 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3464                                  int num_threads);
3465 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3466                                     int num_teams_ub, int num_threads);
3467 
3468 extern void __kmp_yield();
3469 
3470 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3471                                    enum sched_type schedule, kmp_int32 lb,
3472                                    kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3473 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3474                                     enum sched_type schedule, kmp_uint32 lb,
3475                                     kmp_uint32 ub, kmp_int32 st,
3476                                     kmp_int32 chunk);
3477 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3478                                    enum sched_type schedule, kmp_int64 lb,
3479                                    kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3480 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3481                                     enum sched_type schedule, kmp_uint64 lb,
3482                                     kmp_uint64 ub, kmp_int64 st,
3483                                     kmp_int64 chunk);
3484 
3485 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3486                                   kmp_int32 *p_last, kmp_int32 *p_lb,
3487                                   kmp_int32 *p_ub, kmp_int32 *p_st);
3488 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3489                                    kmp_int32 *p_last, kmp_uint32 *p_lb,
3490                                    kmp_uint32 *p_ub, kmp_int32 *p_st);
3491 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3492                                   kmp_int32 *p_last, kmp_int64 *p_lb,
3493                                   kmp_int64 *p_ub, kmp_int64 *p_st);
3494 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3495                                    kmp_int32 *p_last, kmp_uint64 *p_lb,
3496                                    kmp_uint64 *p_ub, kmp_int64 *p_st);
3497 
3498 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3499 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3500 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3501 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3502 
3503 #ifdef KMP_GOMP_COMPAT
3504 
3505 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3506                                       enum sched_type schedule, kmp_int32 lb,
3507                                       kmp_int32 ub, kmp_int32 st,
3508                                       kmp_int32 chunk, int push_ws);
3509 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3510                                        enum sched_type schedule, kmp_uint32 lb,
3511                                        kmp_uint32 ub, kmp_int32 st,
3512                                        kmp_int32 chunk, int push_ws);
3513 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3514                                       enum sched_type schedule, kmp_int64 lb,
3515                                       kmp_int64 ub, kmp_int64 st,
3516                                       kmp_int64 chunk, int push_ws);
3517 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3518                                        enum sched_type schedule, kmp_uint64 lb,
3519                                        kmp_uint64 ub, kmp_int64 st,
3520                                        kmp_int64 chunk, int push_ws);
3521 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3522 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3523 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3524 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3525 
3526 #endif /* KMP_GOMP_COMPAT */
3527 
3528 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3529 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3530 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3531 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3532 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3533 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3534                                kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3535                                void *obj);
3536 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3537                              kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3538 
3539 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3540                           int final_spin
3541 #if USE_ITT_BUILD
3542                           ,
3543                           void *itt_sync_obj
3544 #endif
3545 );
3546 extern void __kmp_release_64(kmp_flag_64<> *flag);
3547 
3548 extern void __kmp_infinite_loop(void);
3549 
3550 extern void __kmp_cleanup(void);
3551 
3552 #if KMP_HANDLE_SIGNALS
3553 extern int __kmp_handle_signals;
3554 extern void __kmp_install_signals(int parallel_init);
3555 extern void __kmp_remove_signals(void);
3556 #endif
3557 
3558 extern void __kmp_clear_system_time(void);
3559 extern void __kmp_read_system_time(double *delta);
3560 
3561 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3562 
3563 extern void __kmp_expand_host_name(char *buffer, size_t size);
3564 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3565 
3566 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64)
3567 extern void
3568 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3569 #endif
3570 
3571 extern void
3572 __kmp_runtime_initialize(void); /* machine specific initialization */
3573 extern void __kmp_runtime_destroy(void);
3574 
3575 #if KMP_AFFINITY_SUPPORTED
3576 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3577                                        kmp_affin_mask_t *mask);
3578 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3579                                                   kmp_affin_mask_t *mask);
3580 extern void __kmp_affinity_initialize(void);
3581 extern void __kmp_affinity_uninitialize(void);
3582 extern void __kmp_affinity_set_init_mask(
3583     int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3584 extern void __kmp_affinity_set_place(int gtid);
3585 extern void __kmp_affinity_determine_capable(const char *env_var);
3586 extern int __kmp_aux_set_affinity(void **mask);
3587 extern int __kmp_aux_get_affinity(void **mask);
3588 extern int __kmp_aux_get_affinity_max_proc();
3589 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3590 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3591 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3592 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3593 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3594 extern int kmp_set_thread_affinity_mask_initial(void);
3595 #endif
3596 static inline void __kmp_assign_root_init_mask() {
3597   int gtid = __kmp_entry_gtid();
3598   kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3599   if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3600     __kmp_affinity_set_init_mask(gtid, TRUE);
3601     r->r.r_affinity_assigned = TRUE;
3602   }
3603 }
3604 #else /* KMP_AFFINITY_SUPPORTED */
3605 #define __kmp_assign_root_init_mask() /* Nothing */
3606 #endif /* KMP_AFFINITY_SUPPORTED */
3607 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3608 // format string is for affinity, so platforms that do not support
3609 // affinity can still use the other fields, e.g., %n for num_threads
3610 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3611                                          kmp_str_buf_t *buffer);
3612 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3613 
3614 extern void __kmp_cleanup_hierarchy();
3615 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3616 
3617 #if KMP_USE_FUTEX
3618 
3619 extern int __kmp_futex_determine_capable(void);
3620 
3621 #endif // KMP_USE_FUTEX
3622 
3623 extern void __kmp_gtid_set_specific(int gtid);
3624 extern int __kmp_gtid_get_specific(void);
3625 
3626 extern double __kmp_read_cpu_time(void);
3627 
3628 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3629 
3630 #if KMP_USE_MONITOR
3631 extern void __kmp_create_monitor(kmp_info_t *th);
3632 #endif
3633 
3634 extern void *__kmp_launch_thread(kmp_info_t *thr);
3635 
3636 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3637 
3638 #if KMP_OS_WINDOWS
3639 extern int __kmp_still_running(kmp_info_t *th);
3640 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3641 extern void __kmp_free_handle(kmp_thread_t tHandle);
3642 #endif
3643 
3644 #if KMP_USE_MONITOR
3645 extern void __kmp_reap_monitor(kmp_info_t *th);
3646 #endif
3647 extern void __kmp_reap_worker(kmp_info_t *th);
3648 extern void __kmp_terminate_thread(int gtid);
3649 
3650 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3651 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3652 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3653 
3654 extern void __kmp_elapsed(double *);
3655 extern void __kmp_elapsed_tick(double *);
3656 
3657 extern void __kmp_enable(int old_state);
3658 extern void __kmp_disable(int *old_state);
3659 
3660 extern void __kmp_thread_sleep(int millis);
3661 
3662 extern void __kmp_common_initialize(void);
3663 extern void __kmp_common_destroy(void);
3664 extern void __kmp_common_destroy_gtid(int gtid);
3665 
3666 #if KMP_OS_UNIX
3667 extern void __kmp_register_atfork(void);
3668 #endif
3669 extern void __kmp_suspend_initialize(void);
3670 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3671 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3672 
3673 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3674                                          int tid);
3675 extern kmp_team_t *
3676 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3677 #if OMPT_SUPPORT
3678                     ompt_data_t ompt_parallel_data,
3679 #endif
3680                     kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3681                     int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3682 extern void __kmp_free_thread(kmp_info_t *);
3683 extern void __kmp_free_team(kmp_root_t *,
3684                             kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3685 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3686 
3687 /* ------------------------------------------------------------------------ */
3688 
3689 extern void __kmp_initialize_bget(kmp_info_t *th);
3690 extern void __kmp_finalize_bget(kmp_info_t *th);
3691 
3692 KMP_EXPORT void *kmpc_malloc(size_t size);
3693 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3694 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3695 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3696 KMP_EXPORT void kmpc_free(void *ptr);
3697 
3698 /* declarations for internal use */
3699 
3700 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3701                          size_t reduce_size, void *reduce_data,
3702                          void (*reduce)(void *, void *));
3703 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3704 extern int __kmp_barrier_gomp_cancel(int gtid);
3705 
3706 /*!
3707  * Tell the fork call which compiler generated the fork call, and therefore how
3708  * to deal with the call.
3709  */
3710 enum fork_context_e {
3711   fork_context_gnu, /**< Called from GNU generated code, so must not invoke the
3712                        microtask internally. */
3713   fork_context_intel, /**< Called from Intel generated code.  */
3714   fork_context_last
3715 };
3716 extern int __kmp_fork_call(ident_t *loc, int gtid,
3717                            enum fork_context_e fork_context, kmp_int32 argc,
3718                            microtask_t microtask, launch_t invoker,
3719                            kmp_va_list ap);
3720 
3721 extern void __kmp_join_call(ident_t *loc, int gtid
3722 #if OMPT_SUPPORT
3723                             ,
3724                             enum fork_context_e fork_context
3725 #endif
3726                             ,
3727                             int exit_teams = 0);
3728 
3729 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3730 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3731 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3732 extern int __kmp_invoke_task_func(int gtid);
3733 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3734                                           kmp_info_t *this_thr,
3735                                           kmp_team_t *team);
3736 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3737                                          kmp_info_t *this_thr,
3738                                          kmp_team_t *team);
3739 
3740 // should never have been exported
3741 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3742 extern int __kmp_invoke_teams_master(int gtid);
3743 extern void __kmp_teams_master(int gtid);
3744 extern int __kmp_aux_get_team_num();
3745 extern int __kmp_aux_get_num_teams();
3746 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3747 extern void __kmp_user_set_library(enum library_type arg);
3748 extern void __kmp_aux_set_library(enum library_type arg);
3749 extern void __kmp_aux_set_stacksize(size_t arg);
3750 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3751 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3752 
3753 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3754 void kmpc_set_blocktime(int arg);
3755 void ompc_set_nested(int flag);
3756 void ompc_set_dynamic(int flag);
3757 void ompc_set_num_threads(int arg);
3758 
3759 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3760                                               kmp_team_t *team, int tid);
3761 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3762 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3763                                     kmp_tasking_flags_t *flags,
3764                                     size_t sizeof_kmp_task_t,
3765                                     size_t sizeof_shareds,
3766                                     kmp_routine_entry_t task_entry);
3767 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3768                                      kmp_team_t *team, int tid,
3769                                      int set_curr_task);
3770 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3771 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3772 
3773 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3774                                                        int gtid,
3775                                                        kmp_task_t *task);
3776 extern void __kmp_fulfill_event(kmp_event_t *event);
3777 
3778 extern void __kmp_free_task_team(kmp_info_t *thread,
3779                                  kmp_task_team_t *task_team);
3780 extern void __kmp_reap_task_teams(void);
3781 extern void __kmp_wait_to_unref_task_teams(void);
3782 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3783                                   int always);
3784 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3785 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3786 #if USE_ITT_BUILD
3787                                  ,
3788                                  void *itt_sync_obj
3789 #endif /* USE_ITT_BUILD */
3790                                  ,
3791                                  int wait = 1);
3792 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3793                                   int gtid);
3794 
3795 extern int __kmp_is_address_mapped(void *addr);
3796 extern kmp_uint64 __kmp_hardware_timestamp(void);
3797 
3798 #if KMP_OS_UNIX
3799 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3800 #endif
3801 
3802 /* ------------------------------------------------------------------------ */
3803 //
3804 // Assembly routines that have no compiler intrinsic replacement
3805 //
3806 
3807 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3808                                   void *argv[]
3809 #if OMPT_SUPPORT
3810                                   ,
3811                                   void **exit_frame_ptr
3812 #endif
3813 );
3814 
3815 /* ------------------------------------------------------------------------ */
3816 
3817 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3818 KMP_EXPORT void __kmpc_end(ident_t *);
3819 
3820 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3821                                                   kmpc_ctor_vec ctor,
3822                                                   kmpc_cctor_vec cctor,
3823                                                   kmpc_dtor_vec dtor,
3824                                                   size_t vector_length);
3825 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3826                                               kmpc_ctor ctor, kmpc_cctor cctor,
3827                                               kmpc_dtor dtor);
3828 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3829                                       void *data, size_t size);
3830 
3831 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3832 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3833 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3834 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3835 
3836 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3837 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3838                                  kmpc_micro microtask, ...);
3839 
3840 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3841 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3842 
3843 KMP_EXPORT void __kmpc_flush(ident_t *);
3844 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3845 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3846 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3847 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3848                                    kmp_int32 filter);
3849 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3850 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3851 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3852 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3853                                 kmp_critical_name *);
3854 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3855                                     kmp_critical_name *);
3856 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3857                                           kmp_critical_name *, uint32_t hint);
3858 
3859 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3860 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3861 
3862 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3863                                                   kmp_int32 global_tid);
3864 
3865 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3866 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3867 
3868 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3869                                      kmp_int32 schedtype, kmp_int32 *plastiter,
3870                                      kmp_int *plower, kmp_int *pupper,
3871                                      kmp_int *pstride, kmp_int incr,
3872                                      kmp_int chunk);
3873 
3874 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3875 
3876 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3877                                    size_t cpy_size, void *cpy_data,
3878                                    void (*cpy_func)(void *, void *),
3879                                    kmp_int32 didit);
3880 
3881 extern void KMPC_SET_NUM_THREADS(int arg);
3882 extern void KMPC_SET_DYNAMIC(int flag);
3883 extern void KMPC_SET_NESTED(int flag);
3884 
3885 /* OMP 3.0 tasking interface routines */
3886 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3887                                      kmp_task_t *new_task);
3888 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3889                                              kmp_int32 flags,
3890                                              size_t sizeof_kmp_task_t,
3891                                              size_t sizeof_shareds,
3892                                              kmp_routine_entry_t task_entry);
3893 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
3894     ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
3895     size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
3896 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3897                                           kmp_task_t *task);
3898 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3899                                              kmp_task_t *task);
3900 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3901                                            kmp_task_t *new_task);
3902 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3903 
3904 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3905                                           int end_part);
3906 
3907 #if TASK_UNUSED
3908 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3909 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3910                               kmp_task_t *task);
3911 #endif // TASK_UNUSED
3912 
3913 /* ------------------------------------------------------------------------ */
3914 
3915 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3916 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3917 
3918 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3919     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3920     kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3921     kmp_depend_info_t *noalias_dep_list);
3922 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3923                                      kmp_int32 ndeps,
3924                                      kmp_depend_info_t *dep_list,
3925                                      kmp_int32 ndeps_noalias,
3926                                      kmp_depend_info_t *noalias_dep_list);
3927 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3928                                 bool serialize_immediate);
3929 
3930 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3931                                    kmp_int32 cncl_kind);
3932 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3933                                               kmp_int32 cncl_kind);
3934 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3935 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3936 
3937 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3938 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3939 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3940                                 kmp_int32 if_val, kmp_uint64 *lb,
3941                                 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3942                                 kmp_int32 sched, kmp_uint64 grainsize,
3943                                 void *task_dup);
3944 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3945                                   kmp_task_t *task, kmp_int32 if_val,
3946                                   kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3947                                   kmp_int32 nogroup, kmp_int32 sched,
3948                                   kmp_uint64 grainsize, kmp_int32 modifier,
3949                                   void *task_dup);
3950 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3951 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3952 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3953 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3954                                                      int is_ws, int num,
3955                                                      void *data);
3956 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3957                                               int num, void *data);
3958 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3959                                                     int is_ws);
3960 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3961     ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3962     kmp_task_affinity_info_t *affin_list);
3963 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
3964 KMP_EXPORT int __kmp_get_max_teams(void);
3965 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
3966 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
3967 
3968 /* Lock interface routines (fast versions with gtid passed in) */
3969 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3970                                  void **user_lock);
3971 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3972                                       void **user_lock);
3973 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3974                                     void **user_lock);
3975 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3976                                          void **user_lock);
3977 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3978 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3979                                      void **user_lock);
3980 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3981                                   void **user_lock);
3982 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3983                                        void **user_lock);
3984 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3985 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3986                                      void **user_lock);
3987 
3988 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3989                                            void **user_lock, uintptr_t hint);
3990 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3991                                                 void **user_lock,
3992                                                 uintptr_t hint);
3993 
3994 /* Interface to fast scalable reduce methods routines */
3995 
3996 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3997     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3998     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3999     kmp_critical_name *lck);
4000 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4001                                          kmp_critical_name *lck);
4002 KMP_EXPORT kmp_int32 __kmpc_reduce(
4003     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4004     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4005     kmp_critical_name *lck);
4006 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4007                                   kmp_critical_name *lck);
4008 
4009 /* Internal fast reduction routines */
4010 
4011 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4012     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4013     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4014     kmp_critical_name *lck);
4015 
4016 // this function is for testing set/get/determine reduce method
4017 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4018 
4019 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4020 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4021 
4022 // C++ port
4023 // missing 'extern "C"' declarations
4024 
4025 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4026 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4027 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4028                                         kmp_int32 num_threads);
4029 
4030 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4031                                       int proc_bind);
4032 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4033                                       kmp_int32 num_teams,
4034                                       kmp_int32 num_threads);
4035 /* Function for OpenMP 5.1 num_teams clause */
4036 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4037                                          kmp_int32 num_teams_lb,
4038                                          kmp_int32 num_teams_ub,
4039                                          kmp_int32 num_threads);
4040 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4041                                   kmpc_micro microtask, ...);
4042 struct kmp_dim { // loop bounds info casted to kmp_int64
4043   kmp_int64 lo; // lower
4044   kmp_int64 up; // upper
4045   kmp_int64 st; // stride
4046 };
4047 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4048                                      kmp_int32 num_dims,
4049                                      const struct kmp_dim *dims);
4050 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4051                                      const kmp_int64 *vec);
4052 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4053                                      const kmp_int64 *vec);
4054 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4055 
4056 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4057                                              void *data, size_t size,
4058                                              void ***cache);
4059 
4060 // Symbols for MS mutual detection.
4061 extern int _You_must_link_with_exactly_one_OpenMP_library;
4062 extern int _You_must_link_with_Intel_OpenMP_library;
4063 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
4064 extern int _You_must_link_with_Microsoft_OpenMP_library;
4065 #endif
4066 
4067 // The routines below are not exported.
4068 // Consider making them 'static' in corresponding source files.
4069 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4070                                            void *data_addr, size_t pc_size);
4071 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4072                                                 void *data_addr,
4073                                                 size_t pc_size);
4074 void __kmp_threadprivate_resize_cache(int newCapacity);
4075 void __kmp_cleanup_threadprivate_caches();
4076 
4077 // ompc_, kmpc_ entries moved from omp.h.
4078 #if KMP_OS_WINDOWS
4079 #define KMPC_CONVENTION __cdecl
4080 #else
4081 #define KMPC_CONVENTION
4082 #endif
4083 
4084 #ifndef __OMP_H
4085 typedef enum omp_sched_t {
4086   omp_sched_static = 1,
4087   omp_sched_dynamic = 2,
4088   omp_sched_guided = 3,
4089   omp_sched_auto = 4
4090 } omp_sched_t;
4091 typedef void *kmp_affinity_mask_t;
4092 #endif
4093 
4094 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4095 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4096 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4097 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4098 KMP_EXPORT int KMPC_CONVENTION
4099 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4100 KMP_EXPORT int KMPC_CONVENTION
4101 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4102 KMP_EXPORT int KMPC_CONVENTION
4103 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4104 
4105 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4106 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4107 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4108 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4109 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4110 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4111 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4112 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4113 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4114                                               char const *format);
4115 
4116 enum kmp_target_offload_kind {
4117   tgt_disabled = 0,
4118   tgt_default = 1,
4119   tgt_mandatory = 2
4120 };
4121 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4122 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4123 extern kmp_target_offload_kind_t __kmp_target_offload;
4124 extern int __kmpc_get_target_offload();
4125 
4126 // Constants used in libomptarget
4127 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4128 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4129 
4130 // OMP Pause Resource
4131 
4132 // The following enum is used both to set the status in __kmp_pause_status, and
4133 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4134 typedef enum kmp_pause_status_t {
4135   kmp_not_paused = 0, // status is not paused, or, requesting resume
4136   kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4137   kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4138 } kmp_pause_status_t;
4139 
4140 // This stores the pause state of the runtime
4141 extern kmp_pause_status_t __kmp_pause_status;
4142 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4143 extern int __kmp_pause_resource(kmp_pause_status_t level);
4144 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4145 extern void __kmp_resume_if_soft_paused();
4146 // Hard resume simply resets the status to not paused. Library will appear to
4147 // be uninitialized after hard pause. Let OMP constructs trigger required
4148 // initializations.
4149 static inline void __kmp_resume_if_hard_paused() {
4150   if (__kmp_pause_status == kmp_hard_paused) {
4151     __kmp_pause_status = kmp_not_paused;
4152   }
4153 }
4154 
4155 extern void __kmp_omp_display_env(int verbose);
4156 
4157 // 1: it is initializing hidden helper team
4158 extern volatile int __kmp_init_hidden_helper;
4159 // 1: the hidden helper team is done
4160 extern volatile int __kmp_hidden_helper_team_done;
4161 // 1: enable hidden helper task
4162 extern kmp_int32 __kmp_enable_hidden_helper;
4163 // Main thread of hidden helper team
4164 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4165 // Descriptors for the hidden helper threads
4166 extern kmp_info_t **__kmp_hidden_helper_threads;
4167 // Number of hidden helper threads
4168 extern kmp_int32 __kmp_hidden_helper_threads_num;
4169 // Number of hidden helper tasks that have not been executed yet
4170 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4171 
4172 extern void __kmp_hidden_helper_initialize();
4173 extern void __kmp_hidden_helper_threads_initz_routine();
4174 extern void __kmp_do_initialize_hidden_helper_threads();
4175 extern void __kmp_hidden_helper_threads_initz_wait();
4176 extern void __kmp_hidden_helper_initz_release();
4177 extern void __kmp_hidden_helper_threads_deinitz_wait();
4178 extern void __kmp_hidden_helper_threads_deinitz_release();
4179 extern void __kmp_hidden_helper_main_thread_wait();
4180 extern void __kmp_hidden_helper_worker_thread_wait();
4181 extern void __kmp_hidden_helper_worker_thread_signal();
4182 extern void __kmp_hidden_helper_main_thread_release();
4183 
4184 // Check whether a given thread is a hidden helper thread
4185 #define KMP_HIDDEN_HELPER_THREAD(gtid)                                         \
4186   ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4187 
4188 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid)                                  \
4189   ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4190 
4191 #define KMP_HIDDEN_HELPER_TEAM(team)                                           \
4192   (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4193 
4194 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4195 // main thread, is skipped.
4196 #define KMP_GTID_TO_SHADOW_GTID(gtid)                                          \
4197   ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4198 
4199 // Return the adjusted gtid value by subtracting from gtid the number
4200 // of hidden helper threads. This adjusted value is the gtid the thread would
4201 // have received if there were no hidden helper threads.
4202 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4203   int adjusted_gtid = gtid;
4204   if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4205       gtid - __kmp_hidden_helper_threads_num >= 0) {
4206     adjusted_gtid -= __kmp_hidden_helper_threads_num;
4207   }
4208   return adjusted_gtid;
4209 }
4210 
4211 // Support for error directive
4212 typedef enum kmp_severity_t {
4213   severity_warning = 1,
4214   severity_fatal = 2
4215 } kmp_severity_t;
4216 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4217 
4218 // Support for scope directive
4219 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4220 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4221 
4222 #ifdef __cplusplus
4223 }
4224 #endif
4225 
4226 template <bool C, bool S>
4227 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4228 template <bool C, bool S>
4229 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4230 template <bool C, bool S>
4231 extern void __kmp_atomic_suspend_64(int th_gtid,
4232                                     kmp_atomic_flag_64<C, S> *flag);
4233 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4234 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4235 template <bool C, bool S>
4236 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4237 template <bool C, bool S>
4238 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4239 template <bool C, bool S>
4240 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4241 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4242 #endif
4243 template <bool C, bool S>
4244 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4245 template <bool C, bool S>
4246 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4247 template <bool C, bool S>
4248 extern void __kmp_atomic_resume_64(int target_gtid,
4249                                    kmp_atomic_flag_64<C, S> *flag);
4250 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4251 
4252 template <bool C, bool S>
4253 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4254                            kmp_flag_32<C, S> *flag, int final_spin,
4255                            int *thread_finished,
4256 #if USE_ITT_BUILD
4257                            void *itt_sync_obj,
4258 #endif /* USE_ITT_BUILD */
4259                            kmp_int32 is_constrained);
4260 template <bool C, bool S>
4261 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4262                            kmp_flag_64<C, S> *flag, int final_spin,
4263                            int *thread_finished,
4264 #if USE_ITT_BUILD
4265                            void *itt_sync_obj,
4266 #endif /* USE_ITT_BUILD */
4267                            kmp_int32 is_constrained);
4268 template <bool C, bool S>
4269 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4270                                   kmp_atomic_flag_64<C, S> *flag,
4271                                   int final_spin, int *thread_finished,
4272 #if USE_ITT_BUILD
4273                                   void *itt_sync_obj,
4274 #endif /* USE_ITT_BUILD */
4275                                   kmp_int32 is_constrained);
4276 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4277                                kmp_flag_oncore *flag, int final_spin,
4278                                int *thread_finished,
4279 #if USE_ITT_BUILD
4280                                void *itt_sync_obj,
4281 #endif /* USE_ITT_BUILD */
4282                                kmp_int32 is_constrained);
4283 
4284 extern int __kmp_nesting_mode;
4285 extern int __kmp_nesting_mode_nlevels;
4286 extern int *__kmp_nesting_nth_level;
4287 extern void __kmp_init_nesting_mode();
4288 extern void __kmp_set_nesting_mode_threads();
4289 
4290 /// This class safely opens and closes a C-style FILE* object using RAII
4291 /// semantics. There are also methods which allow using stdout or stderr as
4292 /// the underlying FILE* object. With the implicit conversion operator to
4293 /// FILE*, an object with this type can be used in any function which takes
4294 /// a FILE* object e.g., fprintf().
4295 /// No close method is needed at use sites.
4296 class kmp_safe_raii_file_t {
4297   FILE *f;
4298 
4299   void close() {
4300     if (f && f != stdout && f != stderr) {
4301       fclose(f);
4302       f = nullptr;
4303     }
4304   }
4305 
4306 public:
4307   kmp_safe_raii_file_t() : f(nullptr) {}
4308   kmp_safe_raii_file_t(const char *filename, const char *mode,
4309                        const char *env_var = nullptr)
4310       : f(nullptr) {
4311     open(filename, mode, env_var);
4312   }
4313   ~kmp_safe_raii_file_t() { close(); }
4314 
4315   /// Open filename using mode. This is automatically closed in the destructor.
4316   /// The env_var parameter indicates the environment variable the filename
4317   /// came from if != nullptr.
4318   void open(const char *filename, const char *mode,
4319             const char *env_var = nullptr) {
4320     KMP_ASSERT(!f);
4321     f = fopen(filename, mode);
4322     if (!f) {
4323       int code = errno;
4324       if (env_var) {
4325         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4326                     KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4327       } else {
4328         __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4329                     __kmp_msg_null);
4330       }
4331     }
4332   }
4333   /// Instead of erroring out, return non-zero when
4334   /// unsuccessful fopen() for any reason
4335   int try_open(const char *filename, const char *mode) {
4336     KMP_ASSERT(!f);
4337     f = fopen(filename, mode);
4338     if (!f)
4339       return errno;
4340     return 0;
4341   }
4342   /// Set the FILE* object to stdout and output there
4343   /// No open call should happen before this call.
4344   void set_stdout() {
4345     KMP_ASSERT(!f);
4346     f = stdout;
4347   }
4348   /// Set the FILE* object to stderr and output there
4349   /// No open call should happen before this call.
4350   void set_stderr() {
4351     KMP_ASSERT(!f);
4352     f = stderr;
4353   }
4354   operator bool() { return bool(f); }
4355   operator FILE *() { return f; }
4356 };
4357 
4358 template <typename SourceType, typename TargetType,
4359           bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4360           bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4361           bool isSourceSigned = std::is_signed<SourceType>::value,
4362           bool isTargetSigned = std::is_signed<TargetType>::value>
4363 struct kmp_convert {};
4364 
4365 // Both types are signed; Source smaller
4366 template <typename SourceType, typename TargetType>
4367 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4368   static TargetType to(SourceType src) { return (TargetType)src; }
4369 };
4370 // Source equal
4371 template <typename SourceType, typename TargetType>
4372 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4373   static TargetType to(SourceType src) { return src; }
4374 };
4375 // Source bigger
4376 template <typename SourceType, typename TargetType>
4377 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4378   static TargetType to(SourceType src) {
4379     KMP_ASSERT(src <= static_cast<SourceType>(
4380                           (std::numeric_limits<TargetType>::max)()));
4381     KMP_ASSERT(src >= static_cast<SourceType>(
4382                           (std::numeric_limits<TargetType>::min)()));
4383     return (TargetType)src;
4384   }
4385 };
4386 
4387 // Source signed, Target unsigned
4388 // Source smaller
4389 template <typename SourceType, typename TargetType>
4390 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4391   static TargetType to(SourceType src) {
4392     KMP_ASSERT(src >= 0);
4393     return (TargetType)src;
4394   }
4395 };
4396 // Source equal
4397 template <typename SourceType, typename TargetType>
4398 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4399   static TargetType to(SourceType src) {
4400     KMP_ASSERT(src >= 0);
4401     return (TargetType)src;
4402   }
4403 };
4404 // Source bigger
4405 template <typename SourceType, typename TargetType>
4406 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4407   static TargetType to(SourceType src) {
4408     KMP_ASSERT(src >= 0);
4409     KMP_ASSERT(src <= static_cast<SourceType>(
4410                           (std::numeric_limits<TargetType>::max)()));
4411     return (TargetType)src;
4412   }
4413 };
4414 
4415 // Source unsigned, Target signed
4416 // Source smaller
4417 template <typename SourceType, typename TargetType>
4418 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4419   static TargetType to(SourceType src) { return (TargetType)src; }
4420 };
4421 // Source equal
4422 template <typename SourceType, typename TargetType>
4423 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4424   static TargetType to(SourceType src) {
4425     KMP_ASSERT(src <= static_cast<SourceType>(
4426                           (std::numeric_limits<TargetType>::max)()));
4427     return (TargetType)src;
4428   }
4429 };
4430 // Source bigger
4431 template <typename SourceType, typename TargetType>
4432 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4433   static TargetType to(SourceType src) {
4434     KMP_ASSERT(src <= static_cast<SourceType>(
4435                           (std::numeric_limits<TargetType>::max)()));
4436     return (TargetType)src;
4437   }
4438 };
4439 
4440 // Source unsigned, Target unsigned
4441 // Source smaller
4442 template <typename SourceType, typename TargetType>
4443 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4444   static TargetType to(SourceType src) { return (TargetType)src; }
4445 };
4446 // Source equal
4447 template <typename SourceType, typename TargetType>
4448 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4449   static TargetType to(SourceType src) { return src; }
4450 };
4451 // Source bigger
4452 template <typename SourceType, typename TargetType>
4453 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4454   static TargetType to(SourceType src) {
4455     KMP_ASSERT(src <= static_cast<SourceType>(
4456                           (std::numeric_limits<TargetType>::max)()));
4457     return (TargetType)src;
4458   }
4459 };
4460 
4461 template <typename T1, typename T2>
4462 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4463   *dest = kmp_convert<T1, T2>::to(src);
4464 }
4465 
4466 #endif /* KMP_H */
4467