LLVM OpenMP* Runtime Library
kmp.h
1 
2 /*
3  * kmp.h -- KPTS runtime header file.
4  */
5 
6 //===----------------------------------------------------------------------===//
7 //
8 // The LLVM Compiler Infrastructure
9 //
10 // This file is dual licensed under the MIT and the University of Illinois Open
11 // Source Licenses. See LICENSE.txt for details.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef KMP_H
16 #define KMP_H
17 
18 #include "kmp_config.h"
19 
20 /* #define BUILD_PARALLEL_ORDERED 1 */
21 
22 /* This fix replaces gettimeofday with clock_gettime for better scalability on
23  the Altix. Requires user code to be linked with -lrt. */
24 //#define FIX_SGI_CLOCK
25 
26 /* Defines for OpenMP 3.0 tasking and auto scheduling */
27 
28 #ifndef KMP_STATIC_STEAL_ENABLED
29 #define KMP_STATIC_STEAL_ENABLED 1
30 #endif
31 
32 #define TASK_CURRENT_NOT_QUEUED 0
33 #define TASK_CURRENT_QUEUED 1
34 
35 #ifdef BUILD_TIED_TASK_STACK
36 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
37 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
38 // Number of entries in each task stack array
39 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
40 // Mask for determining index into stack block
41 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
42 #endif // BUILD_TIED_TASK_STACK
43 
44 #define TASK_NOT_PUSHED 1
45 #define TASK_SUCCESSFULLY_PUSHED 0
46 #define TASK_TIED 1
47 #define TASK_UNTIED 0
48 #define TASK_EXPLICIT 1
49 #define TASK_IMPLICIT 0
50 #define TASK_PROXY 1
51 #define TASK_FULL 0
52 
53 #define KMP_CANCEL_THREADS
54 #define KMP_THREAD_ATTR
55 
56 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
57 // built on Android
58 #if defined(__ANDROID__)
59 #undef KMP_CANCEL_THREADS
60 #endif
61 
62 #include <signal.h>
63 #include <stdarg.h>
64 #include <stddef.h>
65 #include <stdio.h>
66 #include <stdlib.h>
67 #include <string.h>
68 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
69  Microsoft library. Some macros provided below to replace these functions */
70 #ifndef __ABSOFT_WIN
71 #include <sys/types.h>
72 #endif
73 #include <limits.h>
74 #include <time.h>
75 
76 #include <errno.h>
77 
78 #include "kmp_os.h"
79 
80 #include "kmp_safe_c_api.h"
81 
82 #if KMP_STATS_ENABLED
83 class kmp_stats_list;
84 #endif
85 
86 #if KMP_USE_HIER_SCHED
87 // Only include hierarchical scheduling if affinity is supported
88 #undef KMP_USE_HIER_SCHED
89 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
90 #endif
91 
92 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
93 #include "hwloc.h"
94 #ifndef HWLOC_OBJ_NUMANODE
95 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
96 #endif
97 #ifndef HWLOC_OBJ_PACKAGE
98 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
99 #endif
100 #endif
101 
102 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
103 #include <xmmintrin.h>
104 #endif
105 
106 #include "kmp_debug.h"
107 #include "kmp_lock.h"
108 #include "kmp_version.h"
109 #if USE_DEBUGGER
110 #include "kmp_debugger.h"
111 #endif
112 #include "kmp_i18n.h"
113 
114 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
115 
116 #include "kmp_wrapper_malloc.h"
117 #if KMP_OS_UNIX
118 #include <unistd.h>
119 #if !defined NSIG && defined _NSIG
120 #define NSIG _NSIG
121 #endif
122 #endif
123 
124 #if KMP_OS_LINUX
125 #pragma weak clock_gettime
126 #endif
127 
128 #if OMPT_SUPPORT
129 #include "ompt-internal.h"
130 #endif
131 
132 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
133 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
134 // free lists of limited size.
135 #ifndef USE_FAST_MEMORY
136 #define USE_FAST_MEMORY 3
137 #endif
138 
139 #ifndef KMP_NESTED_HOT_TEAMS
140 #define KMP_NESTED_HOT_TEAMS 0
141 #define USE_NESTED_HOT_ARG(x)
142 #else
143 #if KMP_NESTED_HOT_TEAMS
144 #if OMP_40_ENABLED
145 #define USE_NESTED_HOT_ARG(x) , x
146 #else
147 // Nested hot teams feature depends on omp 4.0, disable it for earlier versions
148 #undef KMP_NESTED_HOT_TEAMS
149 #define KMP_NESTED_HOT_TEAMS 0
150 #define USE_NESTED_HOT_ARG(x)
151 #endif
152 #else
153 #define USE_NESTED_HOT_ARG(x)
154 #endif
155 #endif
156 
157 // Assume using BGET compare_exchange instruction instead of lock by default.
158 #ifndef USE_CMP_XCHG_FOR_BGET
159 #define USE_CMP_XCHG_FOR_BGET 1
160 #endif
161 
162 // Test to see if queuing lock is better than bootstrap lock for bget
163 // #ifndef USE_QUEUING_LOCK_FOR_BGET
164 // #define USE_QUEUING_LOCK_FOR_BGET
165 // #endif
166 
167 #define KMP_NSEC_PER_SEC 1000000000L
168 #define KMP_USEC_PER_SEC 1000000L
169 
175 // FIXME DOXYGEN... need to group these flags somehow (Making them an anonymous
176 // enum would do it...)
181 #define KMP_IDENT_IMB 0x01
182 
183 #define KMP_IDENT_KMPC 0x02
184 /* 0x04 is no longer used */
186 #define KMP_IDENT_AUTOPAR 0x08
187 
188 #define KMP_IDENT_ATOMIC_REDUCE 0x10
189 
190 #define KMP_IDENT_BARRIER_EXPL 0x20
191 
192 #define KMP_IDENT_BARRIER_IMPL 0x0040
193 #define KMP_IDENT_BARRIER_IMPL_MASK 0x01C0
194 #define KMP_IDENT_BARRIER_IMPL_FOR 0x0040
195 #define KMP_IDENT_BARRIER_IMPL_SECTIONS 0x00C0
196 
197 #define KMP_IDENT_BARRIER_IMPL_SINGLE 0x0140
198 #define KMP_IDENT_BARRIER_IMPL_WORKSHARE 0x01C0
199 
200 #define KMP_IDENT_WORK_LOOP 0x200 // static loop
201 #define KMP_IDENT_WORK_SECTIONS 0x400 // sections
202 #define KMP_IDENT_WORK_DISTRIBUTE 0x800 // distribute
203 
207 typedef struct ident {
208  kmp_int32 reserved_1;
209  kmp_int32 flags;
211  kmp_int32 reserved_2;
212 #if USE_ITT_BUILD
213 /* but currently used for storing region-specific ITT */
214 /* contextual information. */
215 #endif /* USE_ITT_BUILD */
216  kmp_int32 reserved_3;
217  char const *psource;
221 } ident_t;
226 // Some forward declarations.
227 typedef union kmp_team kmp_team_t;
228 typedef struct kmp_taskdata kmp_taskdata_t;
229 typedef union kmp_task_team kmp_task_team_t;
230 typedef union kmp_team kmp_team_p;
231 typedef union kmp_info kmp_info_p;
232 typedef union kmp_root kmp_root_p;
233 
234 #ifdef __cplusplus
235 extern "C" {
236 #endif
237 
238 /* ------------------------------------------------------------------------ */
239 
240 /* Pack two 32-bit signed integers into a 64-bit signed integer */
241 /* ToDo: Fix word ordering for big-endian machines. */
242 #define KMP_PACK_64(HIGH_32, LOW_32) \
243  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
244 
245 // Generic string manipulation macros. Assume that _x is of type char *
246 #define SKIP_WS(_x) \
247  { \
248  while (*(_x) == ' ' || *(_x) == '\t') \
249  (_x)++; \
250  }
251 #define SKIP_DIGITS(_x) \
252  { \
253  while (*(_x) >= '0' && *(_x) <= '9') \
254  (_x)++; \
255  }
256 #define SKIP_TOKEN(_x) \
257  { \
258  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
259  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
260  (_x)++; \
261  }
262 #define SKIP_TO(_x, _c) \
263  { \
264  while (*(_x) != '\0' && *(_x) != (_c)) \
265  (_x)++; \
266  }
267 
268 /* ------------------------------------------------------------------------ */
269 
270 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
271 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
272 
273 /* ------------------------------------------------------------------------ */
274 /* Enumeration types */
275 
276 enum kmp_state_timer {
277  ts_stop,
278  ts_start,
279  ts_pause,
280 
281  ts_last_state
282 };
283 
284 enum dynamic_mode {
285  dynamic_default,
286 #ifdef USE_LOAD_BALANCE
287  dynamic_load_balance,
288 #endif /* USE_LOAD_BALANCE */
289  dynamic_random,
290  dynamic_thread_limit,
291  dynamic_max
292 };
293 
294 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
295  * not include it here */
296 #ifndef KMP_SCHED_TYPE_DEFINED
297 #define KMP_SCHED_TYPE_DEFINED
298 typedef enum kmp_sched {
299  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
300  // Note: need to adjust __kmp_sch_map global array in case enum is changed
301  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
302  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
303  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
304  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
305  kmp_sched_upper_std = 5, // upper bound for standard schedules
306  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
307  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
308 #if KMP_STATIC_STEAL_ENABLED
309  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
310 #endif
311  kmp_sched_upper,
312  kmp_sched_default = kmp_sched_static // default scheduling
313 } kmp_sched_t;
314 #endif
315 
320 enum sched_type : kmp_int32 {
322  kmp_sch_static_chunked = 33,
324  kmp_sch_dynamic_chunked = 35,
326  kmp_sch_runtime = 37,
328  kmp_sch_trapezoidal = 39,
329 
330  /* accessible only through KMP_SCHEDULE environment variable */
331  kmp_sch_static_greedy = 40,
332  kmp_sch_static_balanced = 41,
333  /* accessible only through KMP_SCHEDULE environment variable */
334  kmp_sch_guided_iterative_chunked = 42,
335  kmp_sch_guided_analytical_chunked = 43,
336  /* accessible only through KMP_SCHEDULE environment variable */
337  kmp_sch_static_steal = 44,
338 
339 #if OMP_45_ENABLED
340  /* static with chunk adjustment (e.g., simd) */
341  kmp_sch_static_balanced_chunked = 45,
342  kmp_sch_guided_simd = 46,
343  kmp_sch_runtime_simd = 47,
344 #endif
345 
346  /* accessible only through KMP_SCHEDULE environment variable */
350  kmp_ord_static_chunked = 65,
352  kmp_ord_dynamic_chunked = 67,
353  kmp_ord_guided_chunked = 68,
354  kmp_ord_runtime = 69,
356  kmp_ord_trapezoidal = 71,
359 #if OMP_40_ENABLED
360  /* Schedules for Distribute construct */
363 #endif
364 
365  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
366  single iteration/chunk, even if the loop is serialized. For the schedule
367  types listed above, the entire iteration vector is returned if the loop is
368  serialized. This doesn't work for gcc/gcomp sections. */
369  kmp_nm_lower = 160,
371  kmp_nm_static_chunked =
372  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
374  kmp_nm_dynamic_chunked = 163,
376  kmp_nm_runtime = 165,
377  kmp_nm_auto = 166,
378  kmp_nm_trapezoidal = 167,
379 
380  /* accessible only through KMP_SCHEDULE environment variable */
381  kmp_nm_static_greedy = 168,
382  kmp_nm_static_balanced = 169,
383  /* accessible only through KMP_SCHEDULE environment variable */
384  kmp_nm_guided_iterative_chunked = 170,
385  kmp_nm_guided_analytical_chunked = 171,
386  kmp_nm_static_steal =
387  172, /* accessible only through OMP_SCHEDULE environment variable */
388 
389  kmp_nm_ord_static_chunked = 193,
391  kmp_nm_ord_dynamic_chunked = 195,
392  kmp_nm_ord_guided_chunked = 196,
393  kmp_nm_ord_runtime = 197,
395  kmp_nm_ord_trapezoidal = 199,
398 #if OMP_45_ENABLED
399  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
400  we need to distinguish the three possible cases (no modifier, monotonic
401  modifier, nonmonotonic modifier), we need separate bits for each modifier.
402  The absence of monotonic does not imply nonmonotonic, especially since 4.5
403  says that the behaviour of the "no modifier" case is implementation defined
404  in 4.5, but will become "nonmonotonic" in 5.0.
405 
406  Since we're passing a full 32 bit value, we can use a couple of high bits
407  for these flags; out of paranoia we avoid the sign bit.
408 
409  These modifiers can be or-ed into non-static schedules by the compiler to
410  pass the additional information. They will be stripped early in the
411  processing in __kmp_dispatch_init when setting up schedules, so most of the
412  code won't ever see schedules with these bits set. */
413  kmp_sch_modifier_monotonic =
414  (1 << 29),
415  kmp_sch_modifier_nonmonotonic =
416  (1 << 30),
418 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
419  (enum sched_type)( \
420  (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))
421 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
422 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
423 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
424  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
425 #else
426 /* By doing this we hope to avoid multiple tests on OMP_45_ENABLED. Compilers
427  can now eliminate tests on compile time constants and dead code that results
428  from them, so we can leave code guarded by such an if in place. */
429 #define SCHEDULE_WITHOUT_MODIFIERS(s) (s)
430 #define SCHEDULE_HAS_MONOTONIC(s) false
431 #define SCHEDULE_HAS_NONMONOTONIC(s) false
432 #define SCHEDULE_HAS_NO_MODIFIERS(s) true
433 #endif
434 
436 };
437 
438 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
439 typedef union kmp_r_sched {
440  struct {
441  enum sched_type r_sched_type;
442  int chunk;
443  };
444  kmp_int64 sched;
445 } kmp_r_sched_t;
446 
447 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
448 // internal schedule types
449 
450 enum library_type {
451  library_none,
452  library_serial,
453  library_turnaround,
454  library_throughput
455 };
456 
457 #if KMP_OS_LINUX
458 enum clock_function_type {
459  clock_function_gettimeofday,
460  clock_function_clock_gettime
461 };
462 #endif /* KMP_OS_LINUX */
463 
464 #if KMP_MIC_SUPPORTED
465 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
466 #endif
467 
468 /* -- fast reduction stuff ------------------------------------------------ */
469 
470 #undef KMP_FAST_REDUCTION_BARRIER
471 #define KMP_FAST_REDUCTION_BARRIER 1
472 
473 #undef KMP_FAST_REDUCTION_CORE_DUO
474 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
475 #define KMP_FAST_REDUCTION_CORE_DUO 1
476 #endif
477 
478 enum _reduction_method {
479  reduction_method_not_defined = 0,
480  critical_reduce_block = (1 << 8),
481  atomic_reduce_block = (2 << 8),
482  tree_reduce_block = (3 << 8),
483  empty_reduce_block = (4 << 8)
484 };
485 
486 // Description of the packed_reduction_method variable:
487 // The packed_reduction_method variable consists of two enum types variables
488 // that are packed together into 0-th byte and 1-st byte:
489 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
490 // barrier that will be used in fast reduction: bs_plain_barrier or
491 // bs_reduction_barrier
492 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
493 // be used in fast reduction;
494 // Reduction method is of 'enum _reduction_method' type and it's defined the way
495 // so that the bits of 0-th byte are empty, so no need to execute a shift
496 // instruction while packing/unpacking
497 
498 #if KMP_FAST_REDUCTION_BARRIER
499 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
500  ((reduction_method) | (barrier_type))
501 
502 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
503  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
504 
505 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
506  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
507 #else
508 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
509  (reduction_method)
510 
511 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
512  (packed_reduction_method)
513 
514 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
515 #endif
516 
517 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
518  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
519  (which_reduction_block))
520 
521 #if KMP_FAST_REDUCTION_BARRIER
522 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
523  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
524 
525 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
526  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
527 #endif
528 
529 typedef int PACKED_REDUCTION_METHOD_T;
530 
531 /* -- end of fast reduction stuff ----------------------------------------- */
532 
533 #if KMP_OS_WINDOWS
534 #define USE_CBLKDATA
535 #pragma warning(push)
536 #pragma warning(disable : 271 310)
537 #include <windows.h>
538 #pragma warning(pop)
539 #endif
540 
541 #if KMP_OS_UNIX
542 #include <dlfcn.h>
543 #include <pthread.h>
544 #endif
545 
546 /* Only Linux* OS and Windows* OS support thread affinity. */
547 #if KMP_AFFINITY_SUPPORTED
548 
549 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
550 #if KMP_OS_WINDOWS
551 #if _MSC_VER < 1600
552 typedef struct GROUP_AFFINITY {
553  KAFFINITY Mask;
554  WORD Group;
555  WORD Reserved[3];
556 } GROUP_AFFINITY;
557 #endif /* _MSC_VER < 1600 */
558 #if KMP_GROUP_AFFINITY
559 extern int __kmp_num_proc_groups;
560 #else
561 static const int __kmp_num_proc_groups = 1;
562 #endif /* KMP_GROUP_AFFINITY */
563 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
564 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
565 
566 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
567 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
568 
569 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
570 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
571 
572 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
573  GROUP_AFFINITY *);
574 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
575 #endif /* KMP_OS_WINDOWS */
576 
577 #if KMP_USE_HWLOC
578 extern hwloc_topology_t __kmp_hwloc_topology;
579 extern int __kmp_hwloc_error;
580 extern int __kmp_numa_detected;
581 extern int __kmp_tile_depth;
582 #endif
583 
584 extern size_t __kmp_affin_mask_size;
585 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
586 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
587 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
588 #define KMP_CPU_SET_ITERATE(i, mask) \
589  for (i = (mask)->begin(); i != (mask)->end(); i = (mask)->next(i))
590 #define KMP_CPU_SET(i, mask) (mask)->set(i)
591 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
592 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
593 #define KMP_CPU_ZERO(mask) (mask)->zero()
594 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
595 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
596 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
597 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
598 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
599 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
600 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
601 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
602 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
603 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
604 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
605 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
606  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
607 #define KMP_CPU_FREE_ARRAY(arr, n) \
608  __kmp_affinity_dispatch->deallocate_mask_array(arr)
609 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
610 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
611 #define __kmp_get_system_affinity(mask, abort_bool) \
612  (mask)->get_system_affinity(abort_bool)
613 #define __kmp_set_system_affinity(mask, abort_bool) \
614  (mask)->set_system_affinity(abort_bool)
615 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
616 
617 class KMPAffinity {
618 public:
619  class Mask {
620  public:
621  void *operator new(size_t n);
622  void operator delete(void *p);
623  void *operator new[](size_t n);
624  void operator delete[](void *p);
625  virtual ~Mask() {}
626  // Set bit i to 1
627  virtual void set(int i) {}
628  // Return bit i
629  virtual bool is_set(int i) const { return false; }
630  // Set bit i to 0
631  virtual void clear(int i) {}
632  // Zero out entire mask
633  virtual void zero() {}
634  // Copy src into this mask
635  virtual void copy(const Mask *src) {}
636  // this &= rhs
637  virtual void bitwise_and(const Mask *rhs) {}
638  // this |= rhs
639  virtual void bitwise_or(const Mask *rhs) {}
640  // this = ~this
641  virtual void bitwise_not() {}
642  // API for iterating over an affinity mask
643  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
644  virtual int begin() const { return 0; }
645  virtual int end() const { return 0; }
646  virtual int next(int previous) const { return 0; }
647  // Set the system's affinity to this affinity mask's value
648  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
649  // Set this affinity mask to the current system affinity
650  virtual int get_system_affinity(bool abort_on_error) { return -1; }
651  // Only 1 DWORD in the mask should have any procs set.
652  // Return the appropriate index, or -1 for an invalid mask.
653  virtual int get_proc_group() const { return -1; }
654  };
655  void *operator new(size_t n);
656  void operator delete(void *p);
657  // Need virtual destructor
658  virtual ~KMPAffinity() = default;
659  // Determine if affinity is capable
660  virtual void determine_capable(const char *env_var) {}
661  // Bind the current thread to os proc
662  virtual void bind_thread(int proc) {}
663  // Factory functions to allocate/deallocate a mask
664  virtual Mask *allocate_mask() { return nullptr; }
665  virtual void deallocate_mask(Mask *m) {}
666  virtual Mask *allocate_mask_array(int num) { return nullptr; }
667  virtual void deallocate_mask_array(Mask *m) {}
668  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
669  static void pick_api();
670  static void destroy_api();
671  enum api_type {
672  NATIVE_OS
673 #if KMP_USE_HWLOC
674  ,
675  HWLOC
676 #endif
677  };
678  virtual api_type get_api_type() const {
679  KMP_ASSERT(0);
680  return NATIVE_OS;
681  }
682 
683 private:
684  static bool picked_api;
685 };
686 
687 typedef KMPAffinity::Mask kmp_affin_mask_t;
688 extern KMPAffinity *__kmp_affinity_dispatch;
689 
690 // Declare local char buffers with this size for printing debug and info
691 // messages, using __kmp_affinity_print_mask().
692 #define KMP_AFFIN_MASK_PRINT_LEN 1024
693 
694 enum affinity_type {
695  affinity_none = 0,
696  affinity_physical,
697  affinity_logical,
698  affinity_compact,
699  affinity_scatter,
700  affinity_explicit,
701  affinity_balanced,
702  affinity_disabled, // not used outsize the env var parser
703  affinity_default
704 };
705 
706 enum affinity_gran {
707  affinity_gran_fine = 0,
708  affinity_gran_thread,
709  affinity_gran_core,
710  affinity_gran_tile,
711  affinity_gran_numa,
712  affinity_gran_package,
713  affinity_gran_node,
714 #if KMP_GROUP_AFFINITY
715  // The "group" granularity isn't necesssarily coarser than all of the
716  // other levels, but we put it last in the enum.
717  affinity_gran_group,
718 #endif /* KMP_GROUP_AFFINITY */
719  affinity_gran_default
720 };
721 
722 enum affinity_top_method {
723  affinity_top_method_all = 0, // try all (supported) methods, in order
724 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
725  affinity_top_method_apicid,
726  affinity_top_method_x2apicid,
727 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
728  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
729 #if KMP_GROUP_AFFINITY
730  affinity_top_method_group,
731 #endif /* KMP_GROUP_AFFINITY */
732  affinity_top_method_flat,
733 #if KMP_USE_HWLOC
734  affinity_top_method_hwloc,
735 #endif
736  affinity_top_method_default
737 };
738 
739 #define affinity_respect_mask_default (-1)
740 
741 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
742 extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */
743 extern int __kmp_affinity_gran_levels; /* corresponding int value */
744 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
745 extern enum affinity_top_method __kmp_affinity_top_method;
746 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
747 extern int __kmp_affinity_offset; /* Affinity offset value */
748 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
749 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
750 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
751 extern char *__kmp_affinity_proclist; /* proc ID list */
752 extern kmp_affin_mask_t *__kmp_affinity_masks;
753 extern unsigned __kmp_affinity_num_masks;
754 extern void __kmp_affinity_bind_thread(int which);
755 
756 extern kmp_affin_mask_t *__kmp_affin_fullMask;
757 extern char *__kmp_cpuinfo_file;
758 
759 #endif /* KMP_AFFINITY_SUPPORTED */
760 
761 #if OMP_40_ENABLED
762 
763 // This needs to be kept in sync with the values in omp.h !!!
764 typedef enum kmp_proc_bind_t {
765  proc_bind_false = 0,
766  proc_bind_true,
767  proc_bind_master,
768  proc_bind_close,
769  proc_bind_spread,
770  proc_bind_intel, // use KMP_AFFINITY interface
771  proc_bind_default
772 } kmp_proc_bind_t;
773 
774 typedef struct kmp_nested_proc_bind_t {
775  kmp_proc_bind_t *bind_types;
776  int size;
777  int used;
778 } kmp_nested_proc_bind_t;
779 
780 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
781 
782 #endif /* OMP_40_ENABLED */
783 
784 #if KMP_AFFINITY_SUPPORTED
785 #define KMP_PLACE_ALL (-1)
786 #define KMP_PLACE_UNDEFINED (-2)
787 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
788 #define KMP_AFFINITY_NON_PROC_BIND \
789  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
790  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
791  (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
792 #endif /* KMP_AFFINITY_SUPPORTED */
793 
794 extern int __kmp_affinity_num_places;
795 
796 #if OMP_40_ENABLED
797 typedef enum kmp_cancel_kind_t {
798  cancel_noreq = 0,
799  cancel_parallel = 1,
800  cancel_loop = 2,
801  cancel_sections = 3,
802  cancel_taskgroup = 4
803 } kmp_cancel_kind_t;
804 #endif // OMP_40_ENABLED
805 
806 // KMP_HW_SUBSET support:
807 typedef struct kmp_hws_item {
808  int num;
809  int offset;
810 } kmp_hws_item_t;
811 
812 extern kmp_hws_item_t __kmp_hws_socket;
813 extern kmp_hws_item_t __kmp_hws_node;
814 extern kmp_hws_item_t __kmp_hws_tile;
815 extern kmp_hws_item_t __kmp_hws_core;
816 extern kmp_hws_item_t __kmp_hws_proc;
817 extern int __kmp_hws_requested;
818 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
819 
820 /* ------------------------------------------------------------------------ */
821 
822 #define KMP_PAD(type, sz) \
823  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
824 
825 // We need to avoid using -1 as a GTID as +1 is added to the gtid
826 // when storing it in a lock, and the value 0 is reserved.
827 #define KMP_GTID_DNE (-2) /* Does not exist */
828 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
829 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
830 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
831 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
832 
833 #define __kmp_get_gtid() __kmp_get_global_thread_id()
834 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
835 
836 #define __kmp_tid_from_gtid(gtid) \
837  (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid)
838 
839 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
840 #define __kmp_gtid_from_tid(tid, team) \
841  (KMP_DEBUG_ASSERT((tid) >= 0 && (team) != NULL), \
842  team->t.t_threads[(tid)]->th.th_info.ds.ds_gtid)
843 
844 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
845 #define __kmp_team_from_gtid(gtid) \
846  (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)]->th.th_team)
847 
848 #define __kmp_thread_from_gtid(gtid) \
849  (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)])
850 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
851 
852 // Returns current thread (pointer to kmp_info_t). In contrast to
853 // __kmp_get_thread(), it works with registered and not-yet-registered threads.
854 #define __kmp_gtid_from_thread(thr) \
855  (KMP_DEBUG_ASSERT((thr) != NULL), (thr)->th.th_info.ds.ds_gtid)
856 
857 // AT: Which way is correct?
858 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
859 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
860 #define __kmp_get_team_num_threads(gtid) \
861  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
862 
863 /* ------------------------------------------------------------------------ */
864 
865 #define KMP_UINT64_MAX \
866  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
867 
868 #define KMP_MIN_NTH 1
869 
870 #ifndef KMP_MAX_NTH
871 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
872 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
873 #else
874 #define KMP_MAX_NTH INT_MAX
875 #endif
876 #endif /* KMP_MAX_NTH */
877 
878 #ifdef PTHREAD_STACK_MIN
879 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
880 #else
881 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
882 #endif
883 
884 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
885 
886 #if KMP_ARCH_X86
887 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
888 #elif KMP_ARCH_X86_64
889 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
890 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
891 #else
892 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
893 #endif
894 
895 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
896 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
897 #define KMP_MAX_MALLOC_POOL_INCR \
898  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
899 
900 #define KMP_MIN_STKOFFSET (0)
901 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
902 #if KMP_OS_DARWIN
903 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
904 #else
905 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
906 #endif
907 
908 #define KMP_MIN_STKPADDING (0)
909 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
910 
911 #define KMP_BLOCKTIME_MULTIPLIER \
912  (1000) /* number of blocktime units per second */
913 #define KMP_MIN_BLOCKTIME (0)
914 #define KMP_MAX_BLOCKTIME \
915  (INT_MAX) /* Must be this for "infinite" setting the work */
916 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
917 
918 #if KMP_USE_MONITOR
919 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
920 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
921 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
922 
923 /* Calculate new number of monitor wakeups for a specific block time based on
924  previous monitor_wakeups. Only allow increasing number of wakeups */
925 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
926  (((blocktime) == KMP_MAX_BLOCKTIME) \
927  ? (monitor_wakeups) \
928  : ((blocktime) == KMP_MIN_BLOCKTIME) \
929  ? KMP_MAX_MONITOR_WAKEUPS \
930  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
931  ? (monitor_wakeups) \
932  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
933 
934 /* Calculate number of intervals for a specific block time based on
935  monitor_wakeups */
936 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
937  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
938  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
939 #else
940 #define KMP_BLOCKTIME(team, tid) \
941  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
942 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
943 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
944 extern kmp_uint64 __kmp_ticks_per_msec;
945 #if KMP_COMPILER_ICC
946 #define KMP_NOW() ((kmp_uint64)_rdtsc())
947 #else
948 #define KMP_NOW() __kmp_hardware_timestamp()
949 #endif
950 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
951 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
952  (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
953 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
954 #else
955 // System time is retrieved sporadically while blocking.
956 extern kmp_uint64 __kmp_now_nsec();
957 #define KMP_NOW() __kmp_now_nsec()
958 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
959 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
960  (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
961 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
962 #endif
963 #define KMP_YIELD_NOW() \
964  (KMP_NOW_MSEC() / KMP_MAX(__kmp_dflt_blocktime, 1) % \
965  (__kmp_yield_on_count + __kmp_yield_off_count) < \
966  (kmp_uint32)__kmp_yield_on_count)
967 #endif // KMP_USE_MONITOR
968 
969 #define KMP_MIN_STATSCOLS 40
970 #define KMP_MAX_STATSCOLS 4096
971 #define KMP_DEFAULT_STATSCOLS 80
972 
973 #define KMP_MIN_INTERVAL 0
974 #define KMP_MAX_INTERVAL (INT_MAX - 1)
975 #define KMP_DEFAULT_INTERVAL 0
976 
977 #define KMP_MIN_CHUNK 1
978 #define KMP_MAX_CHUNK (INT_MAX - 1)
979 #define KMP_DEFAULT_CHUNK 1
980 
981 #define KMP_MIN_INIT_WAIT 1
982 #define KMP_MAX_INIT_WAIT (INT_MAX / 2)
983 #define KMP_DEFAULT_INIT_WAIT 2048U
984 
985 #define KMP_MIN_NEXT_WAIT 1
986 #define KMP_MAX_NEXT_WAIT (INT_MAX / 2)
987 #define KMP_DEFAULT_NEXT_WAIT 1024U
988 
989 #define KMP_DFLT_DISP_NUM_BUFF 7
990 #define KMP_MAX_ORDERED 8
991 
992 #define KMP_MAX_FIELDS 32
993 
994 #define KMP_MAX_BRANCH_BITS 31
995 
996 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
997 
998 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
999 
1000 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1001 
1002 /* Minimum number of threads before switch to TLS gtid (experimentally
1003  determined) */
1004 /* josh TODO: what about OS X* tuning? */
1005 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1006 #define KMP_TLS_GTID_MIN 5
1007 #else
1008 #define KMP_TLS_GTID_MIN INT_MAX
1009 #endif
1010 
1011 #define KMP_MASTER_TID(tid) ((tid) == 0)
1012 #define KMP_WORKER_TID(tid) ((tid) != 0)
1013 
1014 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1015 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1016 #define KMP_UBER_GTID(gtid) \
1017  (KMP_DEBUG_ASSERT((gtid) >= KMP_GTID_MIN), \
1018  KMP_DEBUG_ASSERT((gtid) < __kmp_threads_capacity), \
1019  (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && \
1020  (__kmp_threads[(gtid)] == __kmp_root[(gtid)]->r.r_uber_thread))
1021 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1022 
1023 #ifndef TRUE
1024 #define FALSE 0
1025 #define TRUE (!FALSE)
1026 #endif
1027 
1028 /* NOTE: all of the following constants must be even */
1029 
1030 #if KMP_OS_WINDOWS
1031 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1032 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1033 #elif KMP_OS_CNK
1034 #define KMP_INIT_WAIT 16U /* initial number of spin-tests */
1035 #define KMP_NEXT_WAIT 8U /* susequent number of spin-tests */
1036 #elif KMP_OS_LINUX
1037 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1038 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1039 #elif KMP_OS_DARWIN
1040 /* TODO: tune for KMP_OS_DARWIN */
1041 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1042 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1043 #elif KMP_OS_FREEBSD
1044 /* TODO: tune for KMP_OS_FREEBSD */
1045 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1046 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1047 #elif KMP_OS_NETBSD
1048 /* TODO: tune for KMP_OS_NETBSD */
1049 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1050 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1051 #elif KMP_OS_HURD
1052 /* TODO: tune for KMP_OS_HURD */
1053 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1054 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1055 #endif
1056 
1057 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1058 typedef struct kmp_cpuid {
1059  kmp_uint32 eax;
1060  kmp_uint32 ebx;
1061  kmp_uint32 ecx;
1062  kmp_uint32 edx;
1063 } kmp_cpuid_t;
1064 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1065 #if KMP_ARCH_X86
1066 extern void __kmp_x86_pause(void);
1067 #elif KMP_MIC
1068 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1069 // regression after removal of extra PAUSE from KMP_YIELD_SPIN(). Changing
1070 // the delay from 100 to 300 showed even better performance than double PAUSE
1071 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1072 static void __kmp_x86_pause(void) { _mm_delay_32(300); }
1073 #else
1074 static void __kmp_x86_pause(void) { _mm_pause(); }
1075 #endif
1076 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1077 #elif KMP_ARCH_PPC64
1078 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1079 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1080 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1081 #define KMP_CPU_PAUSE() \
1082  do { \
1083  KMP_PPC64_PRI_LOW(); \
1084  KMP_PPC64_PRI_MED(); \
1085  KMP_PPC64_PRI_LOC_MB(); \
1086  } while (0)
1087 #else
1088 #define KMP_CPU_PAUSE() /* nothing to do */
1089 #endif
1090 
1091 #define KMP_INIT_YIELD(count) \
1092  { (count) = __kmp_yield_init; }
1093 
1094 #define KMP_YIELD(cond) \
1095  { \
1096  KMP_CPU_PAUSE(); \
1097  __kmp_yield((cond)); \
1098  }
1099 
1100 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1101 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1102 
1103 #define KMP_YIELD_WHEN(cond, count) \
1104  { \
1105  KMP_CPU_PAUSE(); \
1106  (count) -= 2; \
1107  if (!(count)) { \
1108  __kmp_yield(cond); \
1109  (count) = __kmp_yield_next; \
1110  } \
1111  }
1112 #define KMP_YIELD_SPIN(count) \
1113  { \
1114  KMP_CPU_PAUSE(); \
1115  (count) -= 2; \
1116  if (!(count)) { \
1117  __kmp_yield(1); \
1118  (count) = __kmp_yield_next; \
1119  } \
1120  }
1121 
1122 /* ------------------------------------------------------------------------ */
1123 /* Support datatypes for the orphaned construct nesting checks. */
1124 /* ------------------------------------------------------------------------ */
1125 
1126 enum cons_type {
1127  ct_none,
1128  ct_parallel,
1129  ct_pdo,
1130  ct_pdo_ordered,
1131  ct_psections,
1132  ct_psingle,
1133 
1134  /* the following must be left in order and not split up */
1135  ct_taskq,
1136  ct_task, // really task inside non-ordered taskq, considered worksharing type
1137  ct_task_ordered, /* really task inside ordered taskq, considered a worksharing
1138  type */
1139  /* the preceding must be left in order and not split up */
1140 
1141  ct_critical,
1142  ct_ordered_in_parallel,
1143  ct_ordered_in_pdo,
1144  ct_ordered_in_taskq,
1145  ct_master,
1146  ct_reduce,
1147  ct_barrier
1148 };
1149 
1150 /* test to see if we are in a taskq construct */
1151 #define IS_CONS_TYPE_TASKQ(ct) \
1152  (((int)(ct)) >= ((int)ct_taskq) && ((int)(ct)) <= ((int)ct_task_ordered))
1153 #define IS_CONS_TYPE_ORDERED(ct) \
1154  ((ct) == ct_pdo_ordered || (ct) == ct_task_ordered)
1155 
1156 struct cons_data {
1157  ident_t const *ident;
1158  enum cons_type type;
1159  int prev;
1160  kmp_user_lock_p
1161  name; /* address exclusively for critical section name comparison */
1162 };
1163 
1164 struct cons_header {
1165  int p_top, w_top, s_top;
1166  int stack_size, stack_top;
1167  struct cons_data *stack_data;
1168 };
1169 
1170 struct kmp_region_info {
1171  char *text;
1172  int offset[KMP_MAX_FIELDS];
1173  int length[KMP_MAX_FIELDS];
1174 };
1175 
1176 /* ---------------------------------------------------------------------- */
1177 /* ---------------------------------------------------------------------- */
1178 
1179 #if KMP_OS_WINDOWS
1180 typedef HANDLE kmp_thread_t;
1181 typedef DWORD kmp_key_t;
1182 #endif /* KMP_OS_WINDOWS */
1183 
1184 #if KMP_OS_UNIX
1185 typedef pthread_t kmp_thread_t;
1186 typedef pthread_key_t kmp_key_t;
1187 #endif
1188 
1189 extern kmp_key_t __kmp_gtid_threadprivate_key;
1190 
1191 typedef struct kmp_sys_info {
1192  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1193  long minflt; /* the number of page faults serviced without any I/O */
1194  long majflt; /* the number of page faults serviced that required I/O */
1195  long nswap; /* the number of times a process was "swapped" out of memory */
1196  long inblock; /* the number of times the file system had to perform input */
1197  long oublock; /* the number of times the file system had to perform output */
1198  long nvcsw; /* the number of times a context switch was voluntarily */
1199  long nivcsw; /* the number of times a context switch was forced */
1200 } kmp_sys_info_t;
1201 
1202 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1203 typedef struct kmp_cpuinfo {
1204  int initialized; // If 0, other fields are not initialized.
1205  int signature; // CPUID(1).EAX
1206  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1207  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1208  // Model << 4 ) + Model)
1209  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1210  int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1211  int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1212  int cpu_stackoffset;
1213  int apic_id;
1214  int physical_id;
1215  int logical_id;
1216  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1217  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1218 } kmp_cpuinfo_t;
1219 #endif
1220 
1221 #if USE_ITT_BUILD
1222 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1223 // required type here. Later we will check the type meets requirements.
1224 typedef int kmp_itt_mark_t;
1225 #define KMP_ITT_DEBUG 0
1226 #endif /* USE_ITT_BUILD */
1227 
1228 /* Taskq data structures */
1229 
1230 #define HIGH_WATER_MARK(nslots) (((nslots)*3) / 4)
1231 // num thunks that each thread can simultaneously execute from a task queue
1232 #define __KMP_TASKQ_THUNKS_PER_TH 1
1233 
1234 /* flags for taskq_global_flags, kmp_task_queue_t tq_flags, kmpc_thunk_t
1235  th_flags */
1236 
1237 #define TQF_IS_ORDERED 0x0001 // __kmpc_taskq interface, taskq ordered
1238 // __kmpc_taskq interface, taskq with lastprivate list
1239 #define TQF_IS_LASTPRIVATE 0x0002
1240 #define TQF_IS_NOWAIT 0x0004 // __kmpc_taskq interface, end taskq nowait
1241 // __kmpc_taskq interface, use heuristics to decide task queue size
1242 #define TQF_HEURISTICS 0x0008
1243 
1244 // __kmpc_taskq interface, reserved for future use
1245 #define TQF_INTERFACE_RESERVED1 0x0010
1246 // __kmpc_taskq interface, reserved for future use
1247 #define TQF_INTERFACE_RESERVED2 0x0020
1248 // __kmpc_taskq interface, reserved for future use
1249 #define TQF_INTERFACE_RESERVED3 0x0040
1250 // __kmpc_taskq interface, reserved for future use
1251 #define TQF_INTERFACE_RESERVED4 0x0080
1252 
1253 #define TQF_INTERFACE_FLAGS 0x00ff // all the __kmpc_taskq interface flags
1254 // internal/read by instrumentation; only used with TQF_IS_LASTPRIVATE
1255 #define TQF_IS_LAST_TASK 0x0100
1256 // internal use only; this thunk->th_task is the taskq_task
1257 #define TQF_TASKQ_TASK 0x0200
1258 // internal use only; must release worker threads once ANY queued task
1259 // exists (global)
1260 #define TQF_RELEASE_WORKERS 0x0400
1261 // internal use only; notify workers that master has finished enqueuing tasks
1262 #define TQF_ALL_TASKS_QUEUED 0x0800
1263 // internal use only: this queue encountered in parallel context: not serialized
1264 #define TQF_PARALLEL_CONTEXT 0x1000
1265 // internal use only; this queue is on the freelist and not in use
1266 #define TQF_DEALLOCATED 0x2000
1267 
1268 #define TQF_INTERNAL_FLAGS 0x3f00 // all the internal use only flags
1269 
1270 typedef struct KMP_ALIGN_CACHE kmpc_aligned_int32_t {
1271  kmp_int32 ai_data;
1272 } kmpc_aligned_int32_t;
1273 
1274 typedef struct KMP_ALIGN_CACHE kmpc_aligned_queue_slot_t {
1275  struct kmpc_thunk_t *qs_thunk;
1276 } kmpc_aligned_queue_slot_t;
1277 
1278 typedef struct kmpc_task_queue_t {
1279  /* task queue linkage fields for n-ary tree of queues (locked with global
1280  taskq_tree_lck) */
1281  kmp_lock_t tq_link_lck; /* lock for child link, child next/prev links and
1282  child ref counts */
1283  union {
1284  struct kmpc_task_queue_t *tq_parent; // pointer to parent taskq, not locked
1285  // for taskq internal freelists, locked with global taskq_freelist_lck
1286  struct kmpc_task_queue_t *tq_next_free;
1287  } tq;
1288  // pointer to linked-list of children, locked by tq's tq_link_lck
1289  volatile struct kmpc_task_queue_t *tq_first_child;
1290  // next child in linked-list, locked by parent tq's tq_link_lck
1291  struct kmpc_task_queue_t *tq_next_child;
1292  // previous child in linked-list, locked by parent tq's tq_link_lck
1293  struct kmpc_task_queue_t *tq_prev_child;
1294  // reference count of threads with access to this task queue
1295  volatile kmp_int32 tq_ref_count;
1296  /* (other than the thread executing the kmpc_end_taskq call) */
1297  /* locked by parent tq's tq_link_lck */
1298 
1299  /* shared data for task queue */
1300  /* per-thread array of pointers to shared variable structures */
1301  struct kmpc_aligned_shared_vars_t *tq_shareds;
1302  /* only one array element exists for all but outermost taskq */
1303 
1304  /* bookkeeping for ordered task queue */
1305  kmp_uint32 tq_tasknum_queuing; // ordered task # assigned while queuing tasks
1306  // ordered number of next task to be served (executed)
1307  volatile kmp_uint32 tq_tasknum_serving;
1308 
1309  /* thunk storage management for task queue */
1310  kmp_lock_t tq_free_thunks_lck; /* lock for thunk freelist manipulation */
1311  // thunk freelist, chained via th.th_next_free
1312  struct kmpc_thunk_t *tq_free_thunks;
1313  // space allocated for thunks for this task queue
1314  struct kmpc_thunk_t *tq_thunk_space;
1315 
1316  /* data fields for queue itself */
1317  kmp_lock_t tq_queue_lck; /* lock for [de]enqueue operations: tq_queue,
1318  tq_head, tq_tail, tq_nfull */
1319  /* array of queue slots to hold thunks for tasks */
1320  kmpc_aligned_queue_slot_t *tq_queue;
1321  volatile struct kmpc_thunk_t *tq_taskq_slot; /* special slot for taskq task
1322  thunk, occupied if not NULL */
1323  kmp_int32 tq_nslots; /* # of tq_thunk_space thunks alloc'd (not incl.
1324  tq_taskq_slot space) */
1325  kmp_int32 tq_head; // enqueue puts item here (index into tq_queue array)
1326  kmp_int32 tq_tail; // dequeue takes item from here (index into tq_queue array)
1327  volatile kmp_int32 tq_nfull; // # of occupied entries in task queue right now
1328  kmp_int32 tq_hiwat; /* high-water mark for tq_nfull and queue scheduling */
1329  volatile kmp_int32 tq_flags; /* TQF_xxx */
1330 
1331  /* bookkeeping for outstanding thunks */
1332 
1333  /* per-thread array for # of regular thunks currently being executed */
1334  struct kmpc_aligned_int32_t *tq_th_thunks;
1335  kmp_int32 tq_nproc; /* number of thunks in the th_thunks array */
1336 
1337  /* statistics library bookkeeping */
1338  ident_t *tq_loc; /* source location information for taskq directive */
1339 } kmpc_task_queue_t;
1340 
1341 typedef void (*kmpc_task_t)(kmp_int32 global_tid, struct kmpc_thunk_t *thunk);
1342 
1343 /* sizeof_shareds passed as arg to __kmpc_taskq call */
1344 typedef struct kmpc_shared_vars_t { /* aligned during dynamic allocation */
1345  kmpc_task_queue_t *sv_queue; /* (pointers to) shared vars */
1346 } kmpc_shared_vars_t;
1347 
1348 typedef struct KMP_ALIGN_CACHE kmpc_aligned_shared_vars_t {
1349  volatile struct kmpc_shared_vars_t *ai_data;
1350 } kmpc_aligned_shared_vars_t;
1351 
1352 /* sizeof_thunk passed as arg to kmpc_taskq call */
1353 typedef struct kmpc_thunk_t { /* aligned during dynamic allocation */
1354  union { /* field used for internal freelists too */
1355  kmpc_shared_vars_t *th_shareds;
1356  struct kmpc_thunk_t *th_next_free; /* freelist of individual thunks within
1357  queue, head at tq_free_thunks */
1358  } th;
1359  kmpc_task_t th_task; /* taskq_task if flags & TQF_TASKQ_TASK */
1360  struct kmpc_thunk_t *th_encl_thunk; /* pointer to dynamically enclosing thunk
1361  on this thread's call stack */
1362  // TQF_xxx(tq_flags interface plus possible internal flags)
1363  kmp_int32 th_flags;
1364 
1365  kmp_int32 th_status;
1366  kmp_uint32 th_tasknum; /* task number assigned in order of queuing, used for
1367  ordered sections */
1368  /* private vars */
1369 } kmpc_thunk_t;
1370 
1371 typedef struct KMP_ALIGN_CACHE kmp_taskq {
1372  int tq_curr_thunk_capacity;
1373 
1374  kmpc_task_queue_t *tq_root;
1375  kmp_int32 tq_global_flags;
1376 
1377  kmp_lock_t tq_freelist_lck;
1378  kmpc_task_queue_t *tq_freelist;
1379 
1380  kmpc_thunk_t **tq_curr_thunk;
1381 } kmp_taskq_t;
1382 
1383 /* END Taskq data structures */
1384 
1385 typedef kmp_int32 kmp_critical_name[8];
1386 
1396 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1397 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1398  ...);
1399 
1404 /* ---------------------------------------------------------------------------
1405  */
1406 /* Threadprivate initialization/finalization function declarations */
1407 
1408 /* for non-array objects: __kmpc_threadprivate_register() */
1409 
1414 typedef void *(*kmpc_ctor)(void *);
1415 
1420 typedef void (*kmpc_dtor)(
1421  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1422  compiler */
1427 typedef void *(*kmpc_cctor)(void *, void *);
1428 
1429 /* for array objects: __kmpc_threadprivate_register_vec() */
1430 /* First arg: "this" pointer */
1431 /* Last arg: number of array elements */
1437 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1443 typedef void (*kmpc_dtor_vec)(void *, size_t);
1449 typedef void *(*kmpc_cctor_vec)(void *, void *,
1450  size_t); /* function unused by compiler */
1451 
1456 /* keeps tracked of threadprivate cache allocations for cleanup later */
1457 typedef struct kmp_cached_addr {
1458  void **addr; /* address of allocated cache */
1459  void ***compiler_cache; /* pointer to compiler's cache */
1460  void *data; /* pointer to global data */
1461  struct kmp_cached_addr *next; /* pointer to next cached address */
1462 } kmp_cached_addr_t;
1463 
1464 struct private_data {
1465  struct private_data *next; /* The next descriptor in the list */
1466  void *data; /* The data buffer for this descriptor */
1467  int more; /* The repeat count for this descriptor */
1468  size_t size; /* The data size for this descriptor */
1469 };
1470 
1471 struct private_common {
1472  struct private_common *next;
1473  struct private_common *link;
1474  void *gbl_addr;
1475  void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1476  size_t cmn_size;
1477 };
1478 
1479 struct shared_common {
1480  struct shared_common *next;
1481  struct private_data *pod_init;
1482  void *obj_init;
1483  void *gbl_addr;
1484  union {
1485  kmpc_ctor ctor;
1486  kmpc_ctor_vec ctorv;
1487  } ct;
1488  union {
1489  kmpc_cctor cctor;
1490  kmpc_cctor_vec cctorv;
1491  } cct;
1492  union {
1493  kmpc_dtor dtor;
1494  kmpc_dtor_vec dtorv;
1495  } dt;
1496  size_t vec_len;
1497  int is_vec;
1498  size_t cmn_size;
1499 };
1500 
1501 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1502 #define KMP_HASH_TABLE_SIZE \
1503  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1504 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1505 #define KMP_HASH(x) \
1506  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1507 
1508 struct common_table {
1509  struct private_common *data[KMP_HASH_TABLE_SIZE];
1510 };
1511 
1512 struct shared_table {
1513  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1514 };
1515 
1516 /* ------------------------------------------------------------------------ */
1517 
1518 #if KMP_USE_HIER_SCHED
1519 // Shared barrier data that exists inside a single unit of the scheduling
1520 // hierarchy
1521 typedef struct kmp_hier_private_bdata_t {
1522  kmp_int32 num_active;
1523  kmp_uint64 index;
1524  kmp_uint64 wait_val[2];
1525 } kmp_hier_private_bdata_t;
1526 #endif
1527 
1528 typedef struct kmp_sched_flags {
1529  unsigned ordered : 1;
1530  unsigned nomerge : 1;
1531  unsigned contains_last : 1;
1532 #if KMP_USE_HIER_SCHED
1533  unsigned use_hier : 1;
1534  unsigned unused : 28;
1535 #else
1536  unsigned unused : 29;
1537 #endif
1538 } kmp_sched_flags_t;
1539 
1540 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1541 
1542 #if KMP_STATIC_STEAL_ENABLED
1543 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1544  kmp_int32 count;
1545  kmp_int32 ub;
1546  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1547  kmp_int32 lb;
1548  kmp_int32 st;
1549  kmp_int32 tc;
1550  kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1551  after ub */
1552 
1553  // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1554  // a) parm3 is properly aligned and
1555  // b) all parm1-4 are in the same cache line.
1556  // Because of parm1-4 are used together, performance seems to be better
1557  // if they are in the same line (not measured though).
1558 
1559  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1560  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1561  kmp_int32 parm2; // make no real change at least while padding is off.
1562  kmp_int32 parm3;
1563  kmp_int32 parm4;
1564  };
1565 
1566  kmp_uint32 ordered_lower;
1567  kmp_uint32 ordered_upper;
1568 #if KMP_OS_WINDOWS
1569  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1570  // 'static_steal_counter'. It would be nice to measure execution times.
1571  // Conditional if/endif can be removed at all.
1572  kmp_int32 last_upper;
1573 #endif /* KMP_OS_WINDOWS */
1574 } dispatch_private_info32_t;
1575 
1576 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1577  kmp_int64 count; // current chunk number for static & static-steal scheduling
1578  kmp_int64 ub; /* upper-bound */
1579  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1580  kmp_int64 lb; /* lower-bound */
1581  kmp_int64 st; /* stride */
1582  kmp_int64 tc; /* trip count (number of iterations) */
1583  kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1584  after ub */
1585 
1586  /* parm[1-4] are used in different ways by different scheduling algorithms */
1587 
1588  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1589  // a) parm3 is properly aligned and
1590  // b) all parm1-4 are in the same cache line.
1591  // Because of parm1-4 are used together, performance seems to be better
1592  // if they are in the same line (not measured though).
1593 
1594  struct KMP_ALIGN(32) {
1595  kmp_int64 parm1;
1596  kmp_int64 parm2;
1597  kmp_int64 parm3;
1598  kmp_int64 parm4;
1599  };
1600 
1601  kmp_uint64 ordered_lower;
1602  kmp_uint64 ordered_upper;
1603 #if KMP_OS_WINDOWS
1604  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1605  // 'static_steal_counter'. It would be nice to measure execution times.
1606  // Conditional if/endif can be removed at all.
1607  kmp_int64 last_upper;
1608 #endif /* KMP_OS_WINDOWS */
1609 } dispatch_private_info64_t;
1610 #else /* KMP_STATIC_STEAL_ENABLED */
1611 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1612  kmp_int32 lb;
1613  kmp_int32 ub;
1614  kmp_int32 st;
1615  kmp_int32 tc;
1616 
1617  kmp_int32 parm1;
1618  kmp_int32 parm2;
1619  kmp_int32 parm3;
1620  kmp_int32 parm4;
1621 
1622  kmp_int32 count;
1623 
1624  kmp_uint32 ordered_lower;
1625  kmp_uint32 ordered_upper;
1626 #if KMP_OS_WINDOWS
1627  kmp_int32 last_upper;
1628 #endif /* KMP_OS_WINDOWS */
1629 } dispatch_private_info32_t;
1630 
1631 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1632  kmp_int64 lb; /* lower-bound */
1633  kmp_int64 ub; /* upper-bound */
1634  kmp_int64 st; /* stride */
1635  kmp_int64 tc; /* trip count (number of iterations) */
1636 
1637  /* parm[1-4] are used in different ways by different scheduling algorithms */
1638  kmp_int64 parm1;
1639  kmp_int64 parm2;
1640  kmp_int64 parm3;
1641  kmp_int64 parm4;
1642 
1643  kmp_int64 count; /* current chunk number for static scheduling */
1644 
1645  kmp_uint64 ordered_lower;
1646  kmp_uint64 ordered_upper;
1647 #if KMP_OS_WINDOWS
1648  kmp_int64 last_upper;
1649 #endif /* KMP_OS_WINDOWS */
1650 } dispatch_private_info64_t;
1651 #endif /* KMP_STATIC_STEAL_ENABLED */
1652 
1653 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1654  union private_info {
1655  dispatch_private_info32_t p32;
1656  dispatch_private_info64_t p64;
1657  } u;
1658  enum sched_type schedule; /* scheduling algorithm */
1659  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1660  kmp_int32 ordered_bumped;
1661  // To retain the structure size after making ordered_iteration scalar
1662  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1663  // Stack of buffers for nest of serial regions
1664  struct dispatch_private_info *next;
1665  kmp_int32 type_size; /* the size of types in private_info */
1666 #if KMP_USE_HIER_SCHED
1667  kmp_int32 hier_id;
1668  void *parent; /* hierarchical scheduling parent pointer */
1669 #endif
1670  enum cons_type pushed_ws;
1671 } dispatch_private_info_t;
1672 
1673 typedef struct dispatch_shared_info32 {
1674  /* chunk index under dynamic, number of idle threads under static-steal;
1675  iteration index otherwise */
1676  volatile kmp_uint32 iteration;
1677  volatile kmp_uint32 num_done;
1678  volatile kmp_uint32 ordered_iteration;
1679  // Dummy to retain the structure size after making ordered_iteration scalar
1680  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1681 } dispatch_shared_info32_t;
1682 
1683 typedef struct dispatch_shared_info64 {
1684  /* chunk index under dynamic, number of idle threads under static-steal;
1685  iteration index otherwise */
1686  volatile kmp_uint64 iteration;
1687  volatile kmp_uint64 num_done;
1688  volatile kmp_uint64 ordered_iteration;
1689  // Dummy to retain the structure size after making ordered_iteration scalar
1690  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1691 } dispatch_shared_info64_t;
1692 
1693 typedef struct dispatch_shared_info {
1694  union shared_info {
1695  dispatch_shared_info32_t s32;
1696  dispatch_shared_info64_t s64;
1697  } u;
1698  volatile kmp_uint32 buffer_index;
1699 #if OMP_45_ENABLED
1700  volatile kmp_int32 doacross_buf_idx; // teamwise index
1701  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1702  kmp_int32 doacross_num_done; // count finished threads
1703 #endif
1704 #if KMP_USE_HIER_SCHED
1705  void *hier;
1706 #endif
1707 #if KMP_USE_HWLOC
1708  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1709  // machines (> 48 cores). Performance analysis showed that a cache thrash
1710  // was occurring and this padding helps alleviate the problem.
1711  char padding[64];
1712 #endif
1713 } dispatch_shared_info_t;
1714 
1715 typedef struct kmp_disp {
1716  /* Vector for ORDERED SECTION */
1717  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1718  /* Vector for END ORDERED SECTION */
1719  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1720 
1721  dispatch_shared_info_t *th_dispatch_sh_current;
1722  dispatch_private_info_t *th_dispatch_pr_current;
1723 
1724  dispatch_private_info_t *th_disp_buffer;
1725  kmp_int32 th_disp_index;
1726 #if OMP_45_ENABLED
1727  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1728  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1729  union { // we can use union here because doacross cannot be used in
1730  // nonmonotonic loops
1731  kmp_int64 *th_doacross_info; // info on loop bounds
1732  kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1733  };
1734 #else
1735 #if KMP_STATIC_STEAL_ENABLED
1736  kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1737  void *dummy_padding[1]; // make it 64 bytes on Intel(R) 64
1738 #else
1739  void *dummy_padding[2]; // make it 64 bytes on Intel(R) 64
1740 #endif
1741 #endif
1742 #if KMP_USE_INTERNODE_ALIGNMENT
1743  char more_padding[INTERNODE_CACHE_LINE];
1744 #endif
1745 } kmp_disp_t;
1746 
1747 /* ------------------------------------------------------------------------ */
1748 /* Barrier stuff */
1749 
1750 /* constants for barrier state update */
1751 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1752 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1753 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1754 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1755 
1756 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1757 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1758 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1759 
1760 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1761 #error "Barrier sleep bit must be smaller than barrier bump bit"
1762 #endif
1763 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1764 #error "Barrier unused bit must be smaller than barrier bump bit"
1765 #endif
1766 
1767 // Constants for release barrier wait state: currently, hierarchical only
1768 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1769 #define KMP_BARRIER_OWN_FLAG \
1770  1 // Normal state; worker waiting on own b_go flag in release
1771 #define KMP_BARRIER_PARENT_FLAG \
1772  2 // Special state; worker waiting on parent's b_go flag in release
1773 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1774  3 // Special state; tells worker to shift from parent to own b_go
1775 #define KMP_BARRIER_SWITCHING \
1776  4 // Special state; worker resets appropriate flag on wake-up
1777 
1778 #define KMP_NOT_SAFE_TO_REAP \
1779  0 // Thread th_reap_state: not safe to reap (tasking)
1780 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1781 
1782 enum barrier_type {
1783  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1784  barriers if enabled) */
1785  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1786 #if KMP_FAST_REDUCTION_BARRIER
1787  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1788 #endif // KMP_FAST_REDUCTION_BARRIER
1789  bs_last_barrier /* Just a placeholder to mark the end */
1790 };
1791 
1792 // to work with reduction barriers just like with plain barriers
1793 #if !KMP_FAST_REDUCTION_BARRIER
1794 #define bs_reduction_barrier bs_plain_barrier
1795 #endif // KMP_FAST_REDUCTION_BARRIER
1796 
1797 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1798  bp_linear_bar =
1799  0, /* Single level (degenerate) tree */
1800  bp_tree_bar =
1801  1, /* Balanced tree with branching factor 2^n */
1802  bp_hyper_bar =
1803  2, /* Hypercube-embedded tree with min branching
1804  factor 2^n */
1805  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1806  bp_last_bar /* Placeholder to mark the end */
1807 } kmp_bar_pat_e;
1808 
1809 #define KMP_BARRIER_ICV_PUSH 1
1810 
1811 /* Record for holding the values of the internal controls stack records */
1812 typedef struct kmp_internal_control {
1813  int serial_nesting_level; /* corresponds to the value of the
1814  th_team_serialized field */
1815  kmp_int8 nested; /* internal control for nested parallelism (per thread) */
1816  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1817  thread) */
1818  kmp_int8
1819  bt_set; /* internal control for whether blocktime is explicitly set */
1820  int blocktime; /* internal control for blocktime */
1821 #if KMP_USE_MONITOR
1822  int bt_intervals; /* internal control for blocktime intervals */
1823 #endif
1824  int nproc; /* internal control for #threads for next parallel region (per
1825  thread) */
1826  int max_active_levels; /* internal control for max_active_levels */
1827  kmp_r_sched_t
1828  sched; /* internal control for runtime schedule {sched,chunk} pair */
1829 #if OMP_40_ENABLED
1830  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1831  kmp_int32 default_device; /* internal control for default device */
1832 #endif // OMP_40_ENABLED
1833  struct kmp_internal_control *next;
1834 } kmp_internal_control_t;
1835 
1836 static inline void copy_icvs(kmp_internal_control_t *dst,
1837  kmp_internal_control_t *src) {
1838  *dst = *src;
1839 }
1840 
1841 /* Thread barrier needs volatile barrier fields */
1842 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1843  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1844  // uses of it). It is not explicitly aligned below, because we *don't* want
1845  // it to be padded -- instead, we fit b_go into the same cache line with
1846  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1847  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1848  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1849  // same NGO store
1850  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1851  KMP_ALIGN_CACHE volatile kmp_uint64
1852  b_arrived; // STATE => task reached synch point.
1853  kmp_uint32 *skip_per_level;
1854  kmp_uint32 my_level;
1855  kmp_int32 parent_tid;
1856  kmp_int32 old_tid;
1857  kmp_uint32 depth;
1858  struct kmp_bstate *parent_bar;
1859  kmp_team_t *team;
1860  kmp_uint64 leaf_state;
1861  kmp_uint32 nproc;
1862  kmp_uint8 base_leaf_kids;
1863  kmp_uint8 leaf_kids;
1864  kmp_uint8 offset;
1865  kmp_uint8 wait_flag;
1866  kmp_uint8 use_oncore_barrier;
1867 #if USE_DEBUGGER
1868  // The following field is intended for the debugger solely. Only the worker
1869  // thread itself accesses this field: the worker increases it by 1 when it
1870  // arrives to a barrier.
1871  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1872 #endif /* USE_DEBUGGER */
1873 } kmp_bstate_t;
1874 
1875 union KMP_ALIGN_CACHE kmp_barrier_union {
1876  double b_align; /* use worst case alignment */
1877  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1878  kmp_bstate_t bb;
1879 };
1880 
1881 typedef union kmp_barrier_union kmp_balign_t;
1882 
1883 /* Team barrier needs only non-volatile arrived counter */
1884 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1885  double b_align; /* use worst case alignment */
1886  char b_pad[CACHE_LINE];
1887  struct {
1888  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1889 #if USE_DEBUGGER
1890  // The following two fields are indended for the debugger solely. Only
1891  // master of the team accesses these fields: the first one is increased by
1892  // 1 when master arrives to a barrier, the second one is increased by one
1893  // when all the threads arrived.
1894  kmp_uint b_master_arrived;
1895  kmp_uint b_team_arrived;
1896 #endif
1897  };
1898 };
1899 
1900 typedef union kmp_barrier_team_union kmp_balign_team_t;
1901 
1902 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
1903  threads when a condition changes. This is to workaround an NPTL bug where
1904  padding was added to pthread_cond_t which caused the initialization routine
1905  to write outside of the structure if compiled on pre-NPTL threads. */
1906 #if KMP_OS_WINDOWS
1907 typedef struct kmp_win32_mutex {
1908  /* The Lock */
1909  CRITICAL_SECTION cs;
1910 } kmp_win32_mutex_t;
1911 
1912 typedef struct kmp_win32_cond {
1913  /* Count of the number of waiters. */
1914  int waiters_count_;
1915 
1916  /* Serialize access to <waiters_count_> */
1917  kmp_win32_mutex_t waiters_count_lock_;
1918 
1919  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
1920  int release_count_;
1921 
1922  /* Keeps track of the current "generation" so that we don't allow */
1923  /* one thread to steal all the "releases" from the broadcast. */
1924  int wait_generation_count_;
1925 
1926  /* A manual-reset event that's used to block and release waiting threads. */
1927  HANDLE event_;
1928 } kmp_win32_cond_t;
1929 #endif
1930 
1931 #if KMP_OS_UNIX
1932 
1933 union KMP_ALIGN_CACHE kmp_cond_union {
1934  double c_align;
1935  char c_pad[CACHE_LINE];
1936  pthread_cond_t c_cond;
1937 };
1938 
1939 typedef union kmp_cond_union kmp_cond_align_t;
1940 
1941 union KMP_ALIGN_CACHE kmp_mutex_union {
1942  double m_align;
1943  char m_pad[CACHE_LINE];
1944  pthread_mutex_t m_mutex;
1945 };
1946 
1947 typedef union kmp_mutex_union kmp_mutex_align_t;
1948 
1949 #endif /* KMP_OS_UNIX */
1950 
1951 typedef struct kmp_desc_base {
1952  void *ds_stackbase;
1953  size_t ds_stacksize;
1954  int ds_stackgrow;
1955  kmp_thread_t ds_thread;
1956  volatile int ds_tid;
1957  int ds_gtid;
1958 #if KMP_OS_WINDOWS
1959  volatile int ds_alive;
1960  DWORD ds_thread_id;
1961 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
1962  However, debugger support (libomp_db) cannot work with handles, because they
1963  uncomparable. For example, debugger requests info about thread with handle h.
1964  h is valid within debugger process, and meaningless within debugee process.
1965  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
1966  within debugee process, but it is a *new* handle which does *not* equal to
1967  any other handle in debugee... The only way to compare handles is convert
1968  them to system-wide ids. GetThreadId() function is available only in
1969  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
1970  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
1971  thread id by call to GetCurrentThreadId() from within the thread and save it
1972  to let libomp_db identify threads. */
1973 #endif /* KMP_OS_WINDOWS */
1974 } kmp_desc_base_t;
1975 
1976 typedef union KMP_ALIGN_CACHE kmp_desc {
1977  double ds_align; /* use worst case alignment */
1978  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1979  kmp_desc_base_t ds;
1980 } kmp_desc_t;
1981 
1982 typedef struct kmp_local {
1983  volatile int this_construct; /* count of single's encountered by thread */
1984  void *reduce_data;
1985 #if KMP_USE_BGET
1986  void *bget_data;
1987  void *bget_list;
1988 #if !USE_CMP_XCHG_FOR_BGET
1989 #ifdef USE_QUEUING_LOCK_FOR_BGET
1990  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
1991 #else
1992  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
1993 // bootstrap lock so we can use it at library
1994 // shutdown.
1995 #endif /* USE_LOCK_FOR_BGET */
1996 #endif /* ! USE_CMP_XCHG_FOR_BGET */
1997 #endif /* KMP_USE_BGET */
1998 
1999  PACKED_REDUCTION_METHOD_T
2000  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2001  __kmpc_end_reduce*() */
2002 
2003 } kmp_local_t;
2004 
2005 #define KMP_CHECK_UPDATE(a, b) \
2006  if ((a) != (b)) \
2007  (a) = (b)
2008 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2009  if ((a) != (b)) \
2010  TCW_SYNC_PTR((a), (b))
2011 
2012 #define get__blocktime(xteam, xtid) \
2013  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2014 #define get__bt_set(xteam, xtid) \
2015  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2016 #if KMP_USE_MONITOR
2017 #define get__bt_intervals(xteam, xtid) \
2018  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2019 #endif
2020 
2021 #define get__nested_2(xteam, xtid) \
2022  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nested)
2023 #define get__dynamic_2(xteam, xtid) \
2024  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2025 #define get__nproc_2(xteam, xtid) \
2026  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2027 #define get__sched_2(xteam, xtid) \
2028  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2029 
2030 #define set__blocktime_team(xteam, xtid, xval) \
2031  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2032  (xval))
2033 
2034 #if KMP_USE_MONITOR
2035 #define set__bt_intervals_team(xteam, xtid, xval) \
2036  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2037  (xval))
2038 #endif
2039 
2040 #define set__bt_set_team(xteam, xtid, xval) \
2041  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2042 
2043 #define set__nested(xthread, xval) \
2044  (((xthread)->th.th_current_task->td_icvs.nested) = (xval))
2045 #define get__nested(xthread) \
2046  (((xthread)->th.th_current_task->td_icvs.nested) ? (FTN_TRUE) : (FTN_FALSE))
2047 
2048 #define set__dynamic(xthread, xval) \
2049  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2050 #define get__dynamic(xthread) \
2051  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2052 
2053 #define set__nproc(xthread, xval) \
2054  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2055 
2056 #define set__max_active_levels(xthread, xval) \
2057  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2058 
2059 #define set__sched(xthread, xval) \
2060  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2061 
2062 #if OMP_40_ENABLED
2063 
2064 #define set__proc_bind(xthread, xval) \
2065  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2066 #define get__proc_bind(xthread) \
2067  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2068 
2069 #endif /* OMP_40_ENABLED */
2070 
2071 // OpenMP tasking data structures
2072 
2073 typedef enum kmp_tasking_mode {
2074  tskm_immediate_exec = 0,
2075  tskm_extra_barrier = 1,
2076  tskm_task_teams = 2,
2077  tskm_max = 2
2078 } kmp_tasking_mode_t;
2079 
2080 extern kmp_tasking_mode_t
2081  __kmp_tasking_mode; /* determines how/when to execute tasks */
2082 extern int __kmp_task_stealing_constraint;
2083 #if OMP_40_ENABLED
2084 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2085 // specified, defaults to 0 otherwise
2086 #endif
2087 #if OMP_45_ENABLED
2088 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2089 extern kmp_int32 __kmp_max_task_priority;
2090 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2091 extern kmp_uint64 __kmp_taskloop_min_tasks;
2092 #endif
2093 
2094 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2095  taskdata first */
2096 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2097 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2098 
2099 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2100 // were spawned and queued since the previous barrier release.
2101 #define KMP_TASKING_ENABLED(task_team) \
2102  (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2103 
2110 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2111 
2112 #if OMP_40_ENABLED || OMP_45_ENABLED
2113 typedef union kmp_cmplrdata {
2114 #if OMP_45_ENABLED
2115  kmp_int32 priority;
2116 #endif // OMP_45_ENABLED
2117 #if OMP_40_ENABLED
2118  kmp_routine_entry_t
2119  destructors; /* pointer to function to invoke deconstructors of
2120  firstprivate C++ objects */
2121 #endif // OMP_40_ENABLED
2122  /* future data */
2123 } kmp_cmplrdata_t;
2124 #endif
2125 
2126 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2129 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2130  void *shareds;
2131  kmp_routine_entry_t
2132  routine;
2133  kmp_int32 part_id;
2134 #if OMP_40_ENABLED || OMP_45_ENABLED
2135  kmp_cmplrdata_t
2136  data1; /* Two known optional additions: destructors and priority */
2137  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2138 /* future data */
2139 #endif
2140  /* private vars */
2141 } kmp_task_t;
2142 
2147 #if OMP_40_ENABLED
2148 typedef struct kmp_taskgroup {
2149  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2150  std::atomic<kmp_int32>
2151  cancel_request; // request for cancellation of this taskgroup
2152  struct kmp_taskgroup *parent; // parent taskgroup
2153 // TODO: change to OMP_50_ENABLED, need to change build tools for this to work
2154 #if OMP_45_ENABLED
2155  // Block of data to perform task reduction
2156  void *reduce_data; // reduction related info
2157  kmp_int32 reduce_num_data; // number of data items to reduce
2158 #endif
2159 } kmp_taskgroup_t;
2160 
2161 // forward declarations
2162 typedef union kmp_depnode kmp_depnode_t;
2163 typedef struct kmp_depnode_list kmp_depnode_list_t;
2164 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2165 
2166 typedef struct kmp_depend_info {
2167  kmp_intptr_t base_addr;
2168  size_t len;
2169  struct {
2170  bool in : 1;
2171  bool out : 1;
2172  } flags;
2173 } kmp_depend_info_t;
2174 
2175 struct kmp_depnode_list {
2176  kmp_depnode_t *node;
2177  kmp_depnode_list_t *next;
2178 };
2179 
2180 typedef struct kmp_base_depnode {
2181  kmp_depnode_list_t *successors;
2182  kmp_task_t *task;
2183 
2184  kmp_lock_t lock;
2185 
2186 #if KMP_SUPPORT_GRAPH_OUTPUT
2187  kmp_uint32 id;
2188 #endif
2189 
2190  std::atomic<kmp_int32> npredecessors;
2191  std::atomic<kmp_int32> nrefs;
2192 } kmp_base_depnode_t;
2193 
2194 union KMP_ALIGN_CACHE kmp_depnode {
2195  double dn_align; /* use worst case alignment */
2196  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2197  kmp_base_depnode_t dn;
2198 };
2199 
2200 struct kmp_dephash_entry {
2201  kmp_intptr_t addr;
2202  kmp_depnode_t *last_out;
2203  kmp_depnode_list_t *last_ins;
2204  kmp_dephash_entry_t *next_in_bucket;
2205 };
2206 
2207 typedef struct kmp_dephash {
2208  kmp_dephash_entry_t **buckets;
2209  size_t size;
2210 #ifdef KMP_DEBUG
2211  kmp_uint32 nelements;
2212  kmp_uint32 nconflicts;
2213 #endif
2214 } kmp_dephash_t;
2215 
2216 #endif
2217 
2218 #ifdef BUILD_TIED_TASK_STACK
2219 
2220 /* Tied Task stack definitions */
2221 typedef struct kmp_stack_block {
2222  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2223  struct kmp_stack_block *sb_next;
2224  struct kmp_stack_block *sb_prev;
2225 } kmp_stack_block_t;
2226 
2227 typedef struct kmp_task_stack {
2228  kmp_stack_block_t ts_first_block; // first block of stack entries
2229  kmp_taskdata_t **ts_top; // pointer to the top of stack
2230  kmp_int32 ts_entries; // number of entries on the stack
2231 } kmp_task_stack_t;
2232 
2233 #endif // BUILD_TIED_TASK_STACK
2234 
2235 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2236  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2237  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2238  unsigned final : 1; /* task is final(1) so execute immediately */
2239  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2240  code path */
2241 #if OMP_40_ENABLED
2242  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2243  invoke destructors from the runtime */
2244 #if OMP_45_ENABLED
2245  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2246  context of the RTL) */
2247  unsigned priority_specified : 1; /* set if the compiler provides priority
2248  setting for the task */
2249  unsigned reserved : 10; /* reserved for compiler use */
2250 #else
2251  unsigned reserved : 12; /* reserved for compiler use */
2252 #endif
2253 #else // OMP_40_ENABLED
2254  unsigned reserved : 13; /* reserved for compiler use */
2255 #endif // OMP_40_ENABLED
2256 
2257  /* Library flags */ /* Total library flags must be 16 bits */
2258  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2259  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2260  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2261  // (1) or may be deferred (0)
2262  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2263  // (0) [>= 2 threads]
2264  /* If either team_serial or tasking_ser is set, task team may be NULL */
2265  /* Task State Flags: */
2266  unsigned started : 1; /* 1==started, 0==not started */
2267  unsigned executing : 1; /* 1==executing, 0==not executing */
2268  unsigned complete : 1; /* 1==complete, 0==not complete */
2269  unsigned freed : 1; /* 1==freed, 0==allocateed */
2270  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2271  unsigned reserved31 : 7; /* reserved for library use */
2272 
2273 } kmp_tasking_flags_t;
2274 
2275 struct kmp_taskdata { /* aligned during dynamic allocation */
2276  kmp_int32 td_task_id; /* id, assigned by debugger */
2277  kmp_tasking_flags_t td_flags; /* task flags */
2278  kmp_team_t *td_team; /* team for this task */
2279  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2280  /* Currently not used except for perhaps IDB */
2281  kmp_taskdata_t *td_parent; /* parent task */
2282  kmp_int32 td_level; /* task nesting level */
2283  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2284  ident_t *td_ident; /* task identifier */
2285  // Taskwait data.
2286  ident_t *td_taskwait_ident;
2287  kmp_uint32 td_taskwait_counter;
2288  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2289  KMP_ALIGN_CACHE kmp_internal_control_t
2290  td_icvs; /* Internal control variables for the task */
2291  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2292  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2293  deallocated */
2294  std::atomic<kmp_int32>
2295  td_incomplete_child_tasks; /* Child tasks not yet complete */
2296 #if OMP_40_ENABLED
2297  kmp_taskgroup_t
2298  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2299  kmp_dephash_t
2300  *td_dephash; // Dependencies for children tasks are tracked from here
2301  kmp_depnode_t
2302  *td_depnode; // Pointer to graph node if this task has dependencies
2303 #endif // OMP_40_ENABLED
2304 #if OMP_45_ENABLED
2305  kmp_task_team_t *td_task_team;
2306  kmp_int32 td_size_alloc; // The size of task structure, including shareds etc.
2307 #if defined(KMP_GOMP_COMPAT)
2308  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2309  kmp_int32 td_size_loop_bounds;
2310 #endif
2311 #endif // OMP_45_ENABLED
2312  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2313 #if defined(KMP_GOMP_COMPAT) && OMP_45_ENABLED
2314  // GOMP sends in a copy function for copy constructors
2315  void (*td_copy_func)(void *, void *);
2316 #endif
2317 #if OMPT_SUPPORT
2318  ompt_task_info_t ompt_task_info;
2319 #endif
2320 }; // struct kmp_taskdata
2321 
2322 // Make sure padding above worked
2323 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2324 
2325 // Data for task team but per thread
2326 typedef struct kmp_base_thread_data {
2327  kmp_info_p *td_thr; // Pointer back to thread info
2328  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2329  // queued?
2330  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2331  kmp_taskdata_t *
2332  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2333  kmp_int32 td_deque_size; // Size of deck
2334  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2335  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2336  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2337  // GEH: shouldn't this be volatile since used in while-spin?
2338  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2339 #ifdef BUILD_TIED_TASK_STACK
2340  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2341 // scheduling constraint
2342 #endif // BUILD_TIED_TASK_STACK
2343 } kmp_base_thread_data_t;
2344 
2345 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2346 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2347 
2348 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2349 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2350 
2351 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2352  kmp_base_thread_data_t td;
2353  double td_align; /* use worst case alignment */
2354  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2355 } kmp_thread_data_t;
2356 
2357 // Data for task teams which are used when tasking is enabled for the team
2358 typedef struct kmp_base_task_team {
2359  kmp_bootstrap_lock_t
2360  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2361  /* must be bootstrap lock since used at library shutdown*/
2362  kmp_task_team_t *tt_next; /* For linking the task team free list */
2363  kmp_thread_data_t
2364  *tt_threads_data; /* Array of per-thread structures for task team */
2365  /* Data survives task team deallocation */
2366  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2367  executing this team? */
2368  /* TRUE means tt_threads_data is set up and initialized */
2369  kmp_int32 tt_nproc; /* #threads in team */
2370  kmp_int32
2371  tt_max_threads; /* number of entries allocated for threads_data array */
2372 #if OMP_45_ENABLED
2373  kmp_int32
2374  tt_found_proxy_tasks; /* Have we found proxy tasks since last barrier */
2375 #endif
2376  kmp_int32 tt_untied_task_encountered;
2377 
2378  KMP_ALIGN_CACHE
2379  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2380 
2381  KMP_ALIGN_CACHE
2382  volatile kmp_uint32
2383  tt_active; /* is the team still actively executing tasks */
2384 } kmp_base_task_team_t;
2385 
2386 union KMP_ALIGN_CACHE kmp_task_team {
2387  kmp_base_task_team_t tt;
2388  double tt_align; /* use worst case alignment */
2389  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2390 };
2391 
2392 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2393 // Free lists keep same-size free memory slots for fast memory allocation
2394 // routines
2395 typedef struct kmp_free_list {
2396  void *th_free_list_self; // Self-allocated tasks free list
2397  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2398  // threads
2399  void *th_free_list_other; // Non-self free list (to be returned to owner's
2400  // sync list)
2401 } kmp_free_list_t;
2402 #endif
2403 #if KMP_NESTED_HOT_TEAMS
2404 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2405 // are not put in teams pool, and they don't put threads in threads pool.
2406 typedef struct kmp_hot_team_ptr {
2407  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2408  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2409 } kmp_hot_team_ptr_t;
2410 #endif
2411 #if OMP_40_ENABLED
2412 typedef struct kmp_teams_size {
2413  kmp_int32 nteams; // number of teams in a league
2414  kmp_int32 nth; // number of threads in each team of the league
2415 } kmp_teams_size_t;
2416 #endif
2417 
2418 // OpenMP thread data structures
2419 
2420 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2421  /* Start with the readonly data which is cache aligned and padded. This is
2422  written before the thread starts working by the master. Uber masters may
2423  update themselves later. Usage does not consider serialized regions. */
2424  kmp_desc_t th_info;
2425  kmp_team_p *th_team; /* team we belong to */
2426  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2427  kmp_info_p *th_next_pool; /* next available thread in the pool */
2428  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2429  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2430 
2431  /* The following are cached from the team info structure */
2432  /* TODO use these in more places as determined to be needed via profiling */
2433  int th_team_nproc; /* number of threads in a team */
2434  kmp_info_p *th_team_master; /* the team's master thread */
2435  int th_team_serialized; /* team is serialized */
2436 #if OMP_40_ENABLED
2437  microtask_t th_teams_microtask; /* save entry address for teams construct */
2438  int th_teams_level; /* save initial level of teams construct */
2439 /* it is 0 on device but may be any on host */
2440 #endif
2441 
2442 /* The blocktime info is copied from the team struct to the thread sruct */
2443 /* at the start of a barrier, and the values stored in the team are used */
2444 /* at points in the code where the team struct is no longer guaranteed */
2445 /* to exist (from the POV of worker threads). */
2446 #if KMP_USE_MONITOR
2447  int th_team_bt_intervals;
2448  int th_team_bt_set;
2449 #else
2450  kmp_uint64 th_team_bt_intervals;
2451 #endif
2452 
2453 #if KMP_AFFINITY_SUPPORTED
2454  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2455 #endif
2456 
2457  /* The data set by the master at reinit, then R/W by the worker */
2458  KMP_ALIGN_CACHE int
2459  th_set_nproc; /* if > 0, then only use this request for the next fork */
2460 #if KMP_NESTED_HOT_TEAMS
2461  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2462 #endif
2463 #if OMP_40_ENABLED
2464  kmp_proc_bind_t
2465  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2466  kmp_teams_size_t
2467  th_teams_size; /* number of teams/threads in teams construct */
2468 #if KMP_AFFINITY_SUPPORTED
2469  int th_current_place; /* place currently bound to */
2470  int th_new_place; /* place to bind to in par reg */
2471  int th_first_place; /* first place in partition */
2472  int th_last_place; /* last place in partition */
2473 #endif
2474 #endif
2475 #if USE_ITT_BUILD
2476  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2477  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2478  kmp_uint64 th_frame_time; /* frame timestamp */
2479 #endif /* USE_ITT_BUILD */
2480  kmp_local_t th_local;
2481  struct private_common *th_pri_head;
2482 
2483  /* Now the data only used by the worker (after initial allocation) */
2484  /* TODO the first serial team should actually be stored in the info_t
2485  structure. this will help reduce initial allocation overhead */
2486  KMP_ALIGN_CACHE kmp_team_p
2487  *th_serial_team; /*serialized team held in reserve*/
2488 
2489 #if OMPT_SUPPORT
2490  ompt_thread_info_t ompt_thread_info;
2491 #endif
2492 
2493  /* The following are also read by the master during reinit */
2494  struct common_table *th_pri_common;
2495 
2496  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2497  /* while awaiting queuing lock acquire */
2498 
2499  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2500 
2501  ident_t *th_ident;
2502  unsigned th_x; // Random number generator data
2503  unsigned th_a; // Random number generator data
2504 
2505  /* Tasking-related data for the thread */
2506  kmp_task_team_t *th_task_team; // Task team struct
2507  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2508  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2509  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2510  // at nested levels
2511  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2512  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2513  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2514  // tasking, thus safe to reap
2515 
2516  /* More stuff for keeping track of active/sleeping threads (this part is
2517  written by the worker thread) */
2518  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2519  int th_active; // ! sleeping; 32 bits for TCR/TCW
2520  struct cons_header *th_cons; // used for consistency check
2521 #if KMP_USE_HIER_SCHED
2522  // used for hierarchical scheduling
2523  kmp_hier_private_bdata_t *th_hier_bar_data;
2524 #endif
2525 
2526  /* Add the syncronizing data which is cache aligned and padded. */
2527  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2528 
2529  KMP_ALIGN_CACHE volatile kmp_int32
2530  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2531 
2532 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2533 #define NUM_LISTS 4
2534  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2535 // allocation routines
2536 #endif
2537 
2538 #if KMP_OS_WINDOWS
2539  kmp_win32_cond_t th_suspend_cv;
2540  kmp_win32_mutex_t th_suspend_mx;
2541  int th_suspend_init;
2542 #endif
2543 #if KMP_OS_UNIX
2544  kmp_cond_align_t th_suspend_cv;
2545  kmp_mutex_align_t th_suspend_mx;
2546  int th_suspend_init_count;
2547 #endif
2548 
2549 #if USE_ITT_BUILD
2550  kmp_itt_mark_t th_itt_mark_single;
2551 // alignment ???
2552 #endif /* USE_ITT_BUILD */
2553 #if KMP_STATS_ENABLED
2554  kmp_stats_list *th_stats;
2555 #endif
2556 #if KMP_OS_UNIX
2557  std::atomic<bool> th_blocking;
2558 #endif
2559 } kmp_base_info_t;
2560 
2561 typedef union KMP_ALIGN_CACHE kmp_info {
2562  double th_align; /* use worst case alignment */
2563  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2564  kmp_base_info_t th;
2565 } kmp_info_t;
2566 
2567 // OpenMP thread team data structures
2568 
2569 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2570 
2571 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2572  double dt_align; /* use worst case alignment */
2573  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2574  kmp_base_data_t dt;
2575 } kmp_sleep_team_t;
2576 
2577 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2578  double dt_align; /* use worst case alignment */
2579  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2580  kmp_base_data_t dt;
2581 } kmp_ordered_team_t;
2582 
2583 typedef int (*launch_t)(int gtid);
2584 
2585 /* Minimum number of ARGV entries to malloc if necessary */
2586 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2587 
2588 // Set up how many argv pointers will fit in cache lines containing
2589 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2590 // larger value for more space between the master write/worker read section and
2591 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2592 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2593 #define KMP_INLINE_ARGV_BYTES \
2594  (4 * CACHE_LINE - \
2595  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2596  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2597  CACHE_LINE))
2598 #else
2599 #define KMP_INLINE_ARGV_BYTES \
2600  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2601 #endif
2602 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2603 
2604 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2605  // Synchronization Data
2606  // ---------------------------------------------------------------------------
2607  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2608  kmp_balign_team_t t_bar[bs_last_barrier];
2609  std::atomic<int> t_construct; // count of single directive encountered by team
2610  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2611 
2612  // Master only
2613  // ---------------------------------------------------------------------------
2614  KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2615  int t_master_this_cons; // "this_construct" single counter of master in parent
2616  // team
2617  ident_t *t_ident; // if volatile, have to change too much other crud to
2618  // volatile too
2619  kmp_team_p *t_parent; // parent team
2620  kmp_team_p *t_next_pool; // next free team in the team pool
2621  kmp_disp_t *t_dispatch; // thread's dispatch data
2622  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2623 #if OMP_40_ENABLED
2624  kmp_proc_bind_t t_proc_bind; // bind type for par region
2625 #endif // OMP_40_ENABLED
2626 #if USE_ITT_BUILD
2627  kmp_uint64 t_region_time; // region begin timestamp
2628 #endif /* USE_ITT_BUILD */
2629 
2630  // Master write, workers read
2631  // --------------------------------------------------------------------------
2632  KMP_ALIGN_CACHE void **t_argv;
2633  int t_argc;
2634  int t_nproc; // number of threads in team
2635  microtask_t t_pkfn;
2636  launch_t t_invoke; // procedure to launch the microtask
2637 
2638 #if OMPT_SUPPORT
2639  ompt_team_info_t ompt_team_info;
2640  ompt_lw_taskteam_t *ompt_serialized_team_info;
2641 #endif
2642 
2643 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2644  kmp_int8 t_fp_control_saved;
2645  kmp_int8 t_pad2b;
2646  kmp_int16 t_x87_fpu_control_word; // FP control regs
2647  kmp_uint32 t_mxcsr;
2648 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2649 
2650  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2651 
2652  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2653  kmp_taskdata_t
2654  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2655  int t_level; // nested parallel level
2656 
2657  KMP_ALIGN_CACHE int t_max_argc;
2658  int t_max_nproc; // max threads this team can handle (dynamicly expandable)
2659  int t_serialized; // levels deep of serialized teams
2660  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2661  int t_id; // team's id, assigned by debugger.
2662  int t_active_level; // nested active parallel level
2663  kmp_r_sched_t t_sched; // run-time schedule for the team
2664 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2665  int t_first_place; // first & last place in parent thread's partition.
2666  int t_last_place; // Restore these values to master after par region.
2667 #endif // OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2668  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2669 // omp_set_num_threads() call
2670 
2671 // Read/write by workers as well
2672 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2673  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2674  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2675  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2676  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2677  char dummy_padding[1024];
2678 #endif
2679  // Internal control stack for additional nested teams.
2680  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2681 // for SERIALIZED teams nested 2 or more levels deep
2682 #if OMP_40_ENABLED
2683  // typed flag to store request state of cancellation
2684  std::atomic<kmp_int32> t_cancel_request;
2685 #endif
2686  int t_master_active; // save on fork, restore on join
2687  kmp_taskq_t t_taskq; // this team's task queue
2688  void *t_copypriv_data; // team specific pointer to copyprivate data array
2689 #if KMP_OS_WINDOWS
2690  std::atomic<kmp_uint32> t_copyin_counter;
2691 #endif
2692 #if USE_ITT_BUILD
2693  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2694 #endif /* USE_ITT_BUILD */
2695 } kmp_base_team_t;
2696 
2697 union KMP_ALIGN_CACHE kmp_team {
2698  kmp_base_team_t t;
2699  double t_align; /* use worst case alignment */
2700  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2701 };
2702 
2703 typedef union KMP_ALIGN_CACHE kmp_time_global {
2704  double dt_align; /* use worst case alignment */
2705  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2706  kmp_base_data_t dt;
2707 } kmp_time_global_t;
2708 
2709 typedef struct kmp_base_global {
2710  /* cache-aligned */
2711  kmp_time_global_t g_time;
2712 
2713  /* non cache-aligned */
2714  volatile int g_abort;
2715  volatile int g_done;
2716 
2717  int g_dynamic;
2718  enum dynamic_mode g_dynamic_mode;
2719 } kmp_base_global_t;
2720 
2721 typedef union KMP_ALIGN_CACHE kmp_global {
2722  kmp_base_global_t g;
2723  double g_align; /* use worst case alignment */
2724  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2725 } kmp_global_t;
2726 
2727 typedef struct kmp_base_root {
2728  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2729  // (r_in_parallel>= 0)
2730  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2731  // the synch overhead or keeping r_active
2732  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2733  // GEH: This is misnamed, should be r_in_parallel
2734  volatile int r_nested; // TODO: GEH - This is unused, just remove it entirely.
2735  // keeps a count of active parallel regions per root
2736  std::atomic<int> r_in_parallel;
2737  // GEH: This is misnamed, should be r_active_levels
2738  kmp_team_t *r_root_team;
2739  kmp_team_t *r_hot_team;
2740  kmp_info_t *r_uber_thread;
2741  kmp_lock_t r_begin_lock;
2742  volatile int r_begin;
2743  int r_blocktime; /* blocktime for this root and descendants */
2744  int r_cg_nthreads; // count of active threads in a contention group
2745 } kmp_base_root_t;
2746 
2747 typedef union KMP_ALIGN_CACHE kmp_root {
2748  kmp_base_root_t r;
2749  double r_align; /* use worst case alignment */
2750  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2751 } kmp_root_t;
2752 
2753 struct fortran_inx_info {
2754  kmp_int32 data;
2755 };
2756 
2757 /* ------------------------------------------------------------------------ */
2758 
2759 extern int __kmp_settings;
2760 extern int __kmp_duplicate_library_ok;
2761 #if USE_ITT_BUILD
2762 extern int __kmp_forkjoin_frames;
2763 extern int __kmp_forkjoin_frames_mode;
2764 #endif
2765 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2766 extern int __kmp_determ_red;
2767 
2768 #ifdef KMP_DEBUG
2769 extern int kmp_a_debug;
2770 extern int kmp_b_debug;
2771 extern int kmp_c_debug;
2772 extern int kmp_d_debug;
2773 extern int kmp_e_debug;
2774 extern int kmp_f_debug;
2775 #endif /* KMP_DEBUG */
2776 
2777 /* For debug information logging using rotating buffer */
2778 #define KMP_DEBUG_BUF_LINES_INIT 512
2779 #define KMP_DEBUG_BUF_LINES_MIN 1
2780 
2781 #define KMP_DEBUG_BUF_CHARS_INIT 128
2782 #define KMP_DEBUG_BUF_CHARS_MIN 2
2783 
2784 extern int
2785  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2786 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2787 extern int
2788  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2789 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2790  entry pointer */
2791 
2792 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2793 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2794  printed in buffer so far */
2795 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2796  recommended in warnings */
2797 /* end rotating debug buffer */
2798 
2799 #ifdef KMP_DEBUG
2800 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2801 
2802 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2803 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2804 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2805 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2806 extern int __kmp_par_range_lb;
2807 extern int __kmp_par_range_ub;
2808 #endif
2809 
2810 /* For printing out dynamic storage map for threads and teams */
2811 extern int
2812  __kmp_storage_map; /* True means print storage map for threads and teams */
2813 extern int __kmp_storage_map_verbose; /* True means storage map includes
2814  placement info */
2815 extern int __kmp_storage_map_verbose_specified;
2816 
2817 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2818 extern kmp_cpuinfo_t __kmp_cpuinfo;
2819 #endif
2820 
2821 extern volatile int __kmp_init_serial;
2822 extern volatile int __kmp_init_gtid;
2823 extern volatile int __kmp_init_common;
2824 extern volatile int __kmp_init_middle;
2825 extern volatile int __kmp_init_parallel;
2826 #if KMP_USE_MONITOR
2827 extern volatile int __kmp_init_monitor;
2828 #endif
2829 extern volatile int __kmp_init_user_locks;
2830 extern int __kmp_init_counter;
2831 extern int __kmp_root_counter;
2832 extern int __kmp_version;
2833 
2834 /* list of address of allocated caches for commons */
2835 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2836 
2837 /* Barrier algorithm types and options */
2838 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2839 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2840 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2841 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2842 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2843 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2844 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2845 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2846 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2847 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2848 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2849 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2850 
2851 /* Global Locks */
2852 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2853 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2854 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2855 extern kmp_bootstrap_lock_t
2856  __kmp_exit_lock; /* exit() is not always thread-safe */
2857 #if KMP_USE_MONITOR
2858 extern kmp_bootstrap_lock_t
2859  __kmp_monitor_lock; /* control monitor thread creation */
2860 #endif
2861 extern kmp_bootstrap_lock_t
2862  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2863  __kmp_threads expansion to co-exist */
2864 
2865 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
2866 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
2867 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2868 
2869 /* used for yielding spin-waits */
2870 extern unsigned int __kmp_init_wait; /* initial number of spin-tests */
2871 extern unsigned int __kmp_next_wait; /* susequent number of spin-tests */
2872 
2873 extern enum library_type __kmp_library;
2874 
2875 extern enum sched_type __kmp_sched; /* default runtime scheduling */
2876 extern enum sched_type __kmp_static; /* default static scheduling method */
2877 extern enum sched_type __kmp_guided; /* default guided scheduling method */
2878 extern enum sched_type __kmp_auto; /* default auto scheduling method */
2879 extern int __kmp_chunk; /* default runtime chunk size */
2880 
2881 extern size_t __kmp_stksize; /* stack size per thread */
2882 #if KMP_USE_MONITOR
2883 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
2884 #endif
2885 extern size_t __kmp_stkoffset; /* stack offset per thread */
2886 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
2887 
2888 extern size_t
2889  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
2890 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
2891 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
2892 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
2893 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
2894 extern int __kmp_generate_warnings; /* should we issue warnings? */
2895 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
2896 
2897 #ifdef DEBUG_SUSPEND
2898 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
2899 #endif
2900 
2901 extern kmp_uint32 __kmp_yield_init;
2902 extern kmp_uint32 __kmp_yield_next;
2903 
2904 #if KMP_USE_MONITOR
2905 extern kmp_uint32 __kmp_yielding_on;
2906 #endif
2907 extern kmp_uint32 __kmp_yield_cycle;
2908 extern kmp_int32 __kmp_yield_on_count;
2909 extern kmp_int32 __kmp_yield_off_count;
2910 
2911 /* ------------------------------------------------------------------------- */
2912 extern int __kmp_allThreadsSpecified;
2913 
2914 extern size_t __kmp_align_alloc;
2915 /* following data protected by initialization routines */
2916 extern int __kmp_xproc; /* number of processors in the system */
2917 extern int __kmp_avail_proc; /* number of processors available to the process */
2918 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
2919 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
2920 // maximum total number of concurrently-existing threads on device
2921 extern int __kmp_max_nth;
2922 // maximum total number of concurrently-existing threads in a contention group
2923 extern int __kmp_cg_max_nth;
2924 extern int __kmp_teams_max_nth; // max threads used in a teams construct
2925 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
2926  __kmp_root */
2927 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
2928  region a la OMP_NUM_THREADS */
2929 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
2930  initialization */
2931 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
2932  used (fixed) */
2933 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
2934  (__kmpc_threadprivate_cached()) */
2935 extern int __kmp_dflt_nested; /* nested parallelism enabled by default a la
2936  OMP_NESTED */
2937 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
2938  blocking (env setting) */
2939 #if KMP_USE_MONITOR
2940 extern int
2941  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
2942 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
2943  blocking */
2944 #endif
2945 #ifdef KMP_ADJUST_BLOCKTIME
2946 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
2947 #endif /* KMP_ADJUST_BLOCKTIME */
2948 #ifdef KMP_DFLT_NTH_CORES
2949 extern int __kmp_ncores; /* Total number of cores for threads placement */
2950 #endif
2951 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
2952 extern int __kmp_abort_delay;
2953 
2954 extern int __kmp_need_register_atfork_specified;
2955 extern int
2956  __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
2957  install fork handler */
2958 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
2959  0 - not set, will be set at runtime
2960  1 - using stack search
2961  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
2962  X*) or TlsGetValue(Windows* OS))
2963  3 - static TLS (__declspec(thread) __kmp_gtid),
2964  Linux* OS .so only. */
2965 extern int
2966  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
2967 #ifdef KMP_TDATA_GTID
2968 extern KMP_THREAD_LOCAL int __kmp_gtid;
2969 #endif
2970 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
2971 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
2972 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2973 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
2974 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
2975 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
2976 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2977 
2978 extern int __kmp_dflt_max_active_levels; /* max_active_levels for nested
2979  parallelism enabled by default via
2980  OMP_MAX_ACTIVE_LEVELS */
2981 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
2982  concurrent execution per team */
2983 #if KMP_NESTED_HOT_TEAMS
2984 extern int __kmp_hot_teams_mode;
2985 extern int __kmp_hot_teams_max_level;
2986 #endif
2987 
2988 #if KMP_OS_LINUX
2989 extern enum clock_function_type __kmp_clock_function;
2990 extern int __kmp_clock_function_param;
2991 #endif /* KMP_OS_LINUX */
2992 
2993 #if KMP_MIC_SUPPORTED
2994 extern enum mic_type __kmp_mic_type;
2995 #endif
2996 
2997 #ifdef USE_LOAD_BALANCE
2998 extern double __kmp_load_balance_interval; // load balance algorithm interval
2999 #endif /* USE_LOAD_BALANCE */
3000 
3001 // OpenMP 3.1 - Nested num threads array
3002 typedef struct kmp_nested_nthreads_t {
3003  int *nth;
3004  int size;
3005  int used;
3006 } kmp_nested_nthreads_t;
3007 
3008 extern kmp_nested_nthreads_t __kmp_nested_nth;
3009 
3010 #if KMP_USE_ADAPTIVE_LOCKS
3011 
3012 // Parameters for the speculative lock backoff system.
3013 struct kmp_adaptive_backoff_params_t {
3014  // Number of soft retries before it counts as a hard retry.
3015  kmp_uint32 max_soft_retries;
3016  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3017  // the right
3018  kmp_uint32 max_badness;
3019 };
3020 
3021 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3022 
3023 #if KMP_DEBUG_ADAPTIVE_LOCKS
3024 extern const char *__kmp_speculative_statsfile;
3025 #endif
3026 
3027 #endif // KMP_USE_ADAPTIVE_LOCKS
3028 
3029 #if OMP_40_ENABLED
3030 extern int __kmp_display_env; /* TRUE or FALSE */
3031 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3032 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3033 #endif
3034 
3035 /* ------------------------------------------------------------------------- */
3036 
3037 /* the following are protected by the fork/join lock */
3038 /* write: lock read: anytime */
3039 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3040 /* read/write: lock */
3041 extern volatile kmp_team_t *__kmp_team_pool;
3042 extern volatile kmp_info_t *__kmp_thread_pool;
3043 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3044 
3045 // total num threads reachable from some root thread including all root threads
3046 extern volatile int __kmp_nth;
3047 /* total number of threads reachable from some root thread including all root
3048  threads, and those in the thread pool */
3049 extern volatile int __kmp_all_nth;
3050 extern int __kmp_thread_pool_nth;
3051 extern std::atomic<int> __kmp_thread_pool_active_nth;
3052 
3053 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3054 /* end data protected by fork/join lock */
3055 /* ------------------------------------------------------------------------- */
3056 
3057 extern kmp_global_t __kmp_global; /* global status */
3058 
3059 extern kmp_info_t __kmp_monitor;
3060 // For Debugging Support Library
3061 extern std::atomic<kmp_uint32> __kmp_team_counter;
3062 // For Debugging Support Library
3063 extern std::atomic<kmp_uint32> __kmp_task_counter;
3064 
3065 #if USE_DEBUGGER
3066 #define _KMP_GEN_ID(counter) \
3067  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3068 #else
3069 #define _KMP_GEN_ID(counter) (~0)
3070 #endif /* USE_DEBUGGER */
3071 
3072 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3073 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3074 
3075 /* ------------------------------------------------------------------------ */
3076 
3077 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3078  size_t size, char const *format, ...);
3079 
3080 extern void __kmp_serial_initialize(void);
3081 extern void __kmp_middle_initialize(void);
3082 extern void __kmp_parallel_initialize(void);
3083 
3084 extern void __kmp_internal_begin(void);
3085 extern void __kmp_internal_end_library(int gtid);
3086 extern void __kmp_internal_end_thread(int gtid);
3087 extern void __kmp_internal_end_atexit(void);
3088 extern void __kmp_internal_end_fini(void);
3089 extern void __kmp_internal_end_dtor(void);
3090 extern void __kmp_internal_end_dest(void *);
3091 
3092 extern int __kmp_register_root(int initial_thread);
3093 extern void __kmp_unregister_root(int gtid);
3094 
3095 extern int __kmp_ignore_mppbeg(void);
3096 extern int __kmp_ignore_mppend(void);
3097 
3098 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3099 extern void __kmp_exit_single(int gtid);
3100 
3101 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3102 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3103 
3104 #ifdef USE_LOAD_BALANCE
3105 extern int __kmp_get_load_balance(int);
3106 #endif
3107 
3108 extern int __kmp_get_global_thread_id(void);
3109 extern int __kmp_get_global_thread_id_reg(void);
3110 extern void __kmp_exit_thread(int exit_status);
3111 extern void __kmp_abort(char const *format, ...);
3112 extern void __kmp_abort_thread(void);
3113 KMP_NORETURN extern void __kmp_abort_process(void);
3114 extern void __kmp_warn(char const *format, ...);
3115 
3116 extern void __kmp_set_num_threads(int new_nth, int gtid);
3117 
3118 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3119 // registered.
3120 static inline kmp_info_t *__kmp_entry_thread() {
3121  int gtid = __kmp_entry_gtid();
3122 
3123  return __kmp_threads[gtid];
3124 }
3125 
3126 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3127 extern int __kmp_get_max_active_levels(int gtid);
3128 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3129 extern int __kmp_get_team_size(int gtid, int level);
3130 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3131 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3132 
3133 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3134 extern void __kmp_init_random(kmp_info_t *thread);
3135 
3136 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3137 extern void __kmp_adjust_num_threads(int new_nproc);
3138 
3139 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3140 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3141 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3142 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3143 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3144 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3145 
3146 #if USE_FAST_MEMORY
3147 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3148  size_t size KMP_SRC_LOC_DECL);
3149 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3150 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3151 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3152 #define __kmp_fast_allocate(this_thr, size) \
3153  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3154 #define __kmp_fast_free(this_thr, ptr) \
3155  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3156 #endif
3157 
3158 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3159 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3160  size_t elsize KMP_SRC_LOC_DECL);
3161 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3162  size_t size KMP_SRC_LOC_DECL);
3163 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3164 #define __kmp_thread_malloc(th, size) \
3165  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3166 #define __kmp_thread_calloc(th, nelem, elsize) \
3167  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3168 #define __kmp_thread_realloc(th, ptr, size) \
3169  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3170 #define __kmp_thread_free(th, ptr) \
3171  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3172 
3173 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3174 #define KMP_INTERNAL_FREE(p) free(p)
3175 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3176 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3177 
3178 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3179 
3180 #if OMP_40_ENABLED
3181 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3182  kmp_proc_bind_t proc_bind);
3183 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3184  int num_threads);
3185 #endif
3186 
3187 extern void __kmp_yield(int cond);
3188 
3189 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3190  enum sched_type schedule, kmp_int32 lb,
3191  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3192 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3193  enum sched_type schedule, kmp_uint32 lb,
3194  kmp_uint32 ub, kmp_int32 st,
3195  kmp_int32 chunk);
3196 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3197  enum sched_type schedule, kmp_int64 lb,
3198  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3199 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3200  enum sched_type schedule, kmp_uint64 lb,
3201  kmp_uint64 ub, kmp_int64 st,
3202  kmp_int64 chunk);
3203 
3204 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3205  kmp_int32 *p_last, kmp_int32 *p_lb,
3206  kmp_int32 *p_ub, kmp_int32 *p_st);
3207 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3208  kmp_int32 *p_last, kmp_uint32 *p_lb,
3209  kmp_uint32 *p_ub, kmp_int32 *p_st);
3210 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3211  kmp_int32 *p_last, kmp_int64 *p_lb,
3212  kmp_int64 *p_ub, kmp_int64 *p_st);
3213 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3214  kmp_int32 *p_last, kmp_uint64 *p_lb,
3215  kmp_uint64 *p_ub, kmp_int64 *p_st);
3216 
3217 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3218 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3219 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3220 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3221 
3222 #ifdef KMP_GOMP_COMPAT
3223 
3224 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3225  enum sched_type schedule, kmp_int32 lb,
3226  kmp_int32 ub, kmp_int32 st,
3227  kmp_int32 chunk, int push_ws);
3228 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3229  enum sched_type schedule, kmp_uint32 lb,
3230  kmp_uint32 ub, kmp_int32 st,
3231  kmp_int32 chunk, int push_ws);
3232 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3233  enum sched_type schedule, kmp_int64 lb,
3234  kmp_int64 ub, kmp_int64 st,
3235  kmp_int64 chunk, int push_ws);
3236 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3237  enum sched_type schedule, kmp_uint64 lb,
3238  kmp_uint64 ub, kmp_int64 st,
3239  kmp_int64 chunk, int push_ws);
3240 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3241 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3242 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3243 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3244 
3245 #endif /* KMP_GOMP_COMPAT */
3246 
3247 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3248 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3249 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3250 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3251 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3252 extern kmp_uint32 __kmp_wait_yield_4(kmp_uint32 volatile *spinner,
3253  kmp_uint32 checker,
3254  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3255  void *obj);
3256 extern void __kmp_wait_yield_4_ptr(void *spinner, kmp_uint32 checker,
3257  kmp_uint32 (*pred)(void *, kmp_uint32),
3258  void *obj);
3259 
3260 class kmp_flag_32;
3261 class kmp_flag_64;
3262 class kmp_flag_oncore;
3263 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3264  int final_spin
3265 #if USE_ITT_BUILD
3266  ,
3267  void *itt_sync_obj
3268 #endif
3269  );
3270 extern void __kmp_release_64(kmp_flag_64 *flag);
3271 
3272 extern void __kmp_infinite_loop(void);
3273 
3274 extern void __kmp_cleanup(void);
3275 
3276 #if KMP_HANDLE_SIGNALS
3277 extern int __kmp_handle_signals;
3278 extern void __kmp_install_signals(int parallel_init);
3279 extern void __kmp_remove_signals(void);
3280 #endif
3281 
3282 extern void __kmp_clear_system_time(void);
3283 extern void __kmp_read_system_time(double *delta);
3284 
3285 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3286 
3287 extern void __kmp_expand_host_name(char *buffer, size_t size);
3288 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3289 
3290 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3291 extern void
3292 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3293 #endif
3294 
3295 extern void
3296 __kmp_runtime_initialize(void); /* machine specific initialization */
3297 extern void __kmp_runtime_destroy(void);
3298 
3299 #if KMP_AFFINITY_SUPPORTED
3300 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3301  kmp_affin_mask_t *mask);
3302 extern void __kmp_affinity_initialize(void);
3303 extern void __kmp_affinity_uninitialize(void);
3304 extern void __kmp_affinity_set_init_mask(
3305  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3306 #if OMP_40_ENABLED
3307 extern void __kmp_affinity_set_place(int gtid);
3308 #endif
3309 extern void __kmp_affinity_determine_capable(const char *env_var);
3310 extern int __kmp_aux_set_affinity(void **mask);
3311 extern int __kmp_aux_get_affinity(void **mask);
3312 extern int __kmp_aux_get_affinity_max_proc();
3313 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3314 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3315 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3316 extern void __kmp_balanced_affinity(int tid, int team_size);
3317 #if KMP_OS_LINUX
3318 extern int kmp_set_thread_affinity_mask_initial(void);
3319 #endif
3320 #endif /* KMP_AFFINITY_SUPPORTED */
3321 
3322 extern void __kmp_cleanup_hierarchy();
3323 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3324 
3325 #if KMP_USE_FUTEX
3326 
3327 extern int __kmp_futex_determine_capable(void);
3328 
3329 #endif // KMP_USE_FUTEX
3330 
3331 extern void __kmp_gtid_set_specific(int gtid);
3332 extern int __kmp_gtid_get_specific(void);
3333 
3334 extern double __kmp_read_cpu_time(void);
3335 
3336 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3337 
3338 #if KMP_USE_MONITOR
3339 extern void __kmp_create_monitor(kmp_info_t *th);
3340 #endif
3341 
3342 extern void *__kmp_launch_thread(kmp_info_t *thr);
3343 
3344 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3345 
3346 #if KMP_OS_WINDOWS
3347 extern int __kmp_still_running(kmp_info_t *th);
3348 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3349 extern void __kmp_free_handle(kmp_thread_t tHandle);
3350 #endif
3351 
3352 #if KMP_USE_MONITOR
3353 extern void __kmp_reap_monitor(kmp_info_t *th);
3354 #endif
3355 extern void __kmp_reap_worker(kmp_info_t *th);
3356 extern void __kmp_terminate_thread(int gtid);
3357 
3358 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag);
3359 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag);
3360 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
3361 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag);
3362 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag);
3363 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
3364 
3365 extern void __kmp_elapsed(double *);
3366 extern void __kmp_elapsed_tick(double *);
3367 
3368 extern void __kmp_enable(int old_state);
3369 extern void __kmp_disable(int *old_state);
3370 
3371 extern void __kmp_thread_sleep(int millis);
3372 
3373 extern void __kmp_common_initialize(void);
3374 extern void __kmp_common_destroy(void);
3375 extern void __kmp_common_destroy_gtid(int gtid);
3376 
3377 #if KMP_OS_UNIX
3378 extern void __kmp_register_atfork(void);
3379 #endif
3380 extern void __kmp_suspend_initialize(void);
3381 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3382 
3383 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3384  int tid);
3385 #if OMP_40_ENABLED
3386 extern kmp_team_t *
3387 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3388 #if OMPT_SUPPORT
3389  ompt_data_t ompt_parallel_data,
3390 #endif
3391  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3392  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3393 #else
3394 extern kmp_team_t *
3395 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3396 #if OMPT_SUPPORT
3397  ompt_id_t ompt_parallel_id,
3398 #endif
3399  kmp_internal_control_t *new_icvs,
3400  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3401 #endif // OMP_40_ENABLED
3402 extern void __kmp_free_thread(kmp_info_t *);
3403 extern void __kmp_free_team(kmp_root_t *,
3404  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3405 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3406 
3407 /* ------------------------------------------------------------------------ */
3408 
3409 extern void __kmp_initialize_bget(kmp_info_t *th);
3410 extern void __kmp_finalize_bget(kmp_info_t *th);
3411 
3412 KMP_EXPORT void *kmpc_malloc(size_t size);
3413 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3414 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3415 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3416 KMP_EXPORT void kmpc_free(void *ptr);
3417 
3418 /* declarations for internal use */
3419 
3420 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3421  size_t reduce_size, void *reduce_data,
3422  void (*reduce)(void *, void *));
3423 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3424 
3429 enum fork_context_e {
3430  fork_context_gnu,
3432  fork_context_intel,
3433  fork_context_last
3434 };
3435 extern int __kmp_fork_call(ident_t *loc, int gtid,
3436  enum fork_context_e fork_context, kmp_int32 argc,
3437  microtask_t microtask, launch_t invoker,
3438 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
3439 #if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX
3440  va_list *ap
3441 #else
3442  va_list ap
3443 #endif
3444  );
3445 
3446 extern void __kmp_join_call(ident_t *loc, int gtid
3447 #if OMPT_SUPPORT
3448  ,
3449  enum fork_context_e fork_context
3450 #endif
3451 #if OMP_40_ENABLED
3452  ,
3453  int exit_teams = 0
3454 #endif
3455  );
3456 
3457 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3458 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3459 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3460 extern int __kmp_invoke_task_func(int gtid);
3461 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3462  kmp_info_t *this_thr,
3463  kmp_team_t *team);
3464 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3465  kmp_info_t *this_thr,
3466  kmp_team_t *team);
3467 
3468 // should never have been exported
3469 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3470 #if OMP_40_ENABLED
3471 extern int __kmp_invoke_teams_master(int gtid);
3472 extern void __kmp_teams_master(int gtid);
3473 #endif
3474 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3475 extern void __kmp_user_set_library(enum library_type arg);
3476 extern void __kmp_aux_set_library(enum library_type arg);
3477 extern void __kmp_aux_set_stacksize(size_t arg);
3478 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3479 extern void __kmp_aux_set_defaults(char const *str, int len);
3480 
3481 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3482 void kmpc_set_blocktime(int arg);
3483 void ompc_set_nested(int flag);
3484 void ompc_set_dynamic(int flag);
3485 void ompc_set_num_threads(int arg);
3486 
3487 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3488  kmp_team_t *team, int tid);
3489 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3490 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3491  kmp_tasking_flags_t *flags,
3492  size_t sizeof_kmp_task_t,
3493  size_t sizeof_shareds,
3494  kmp_routine_entry_t task_entry);
3495 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3496  kmp_team_t *team, int tid,
3497  int set_curr_task);
3498 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3499 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3500 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3501  kmp_flag_32 *flag, int final_spin,
3502  int *thread_finished,
3503 #if USE_ITT_BUILD
3504  void *itt_sync_obj,
3505 #endif /* USE_ITT_BUILD */
3506  kmp_int32 is_constrained);
3507 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3508  kmp_flag_64 *flag, int final_spin,
3509  int *thread_finished,
3510 #if USE_ITT_BUILD
3511  void *itt_sync_obj,
3512 #endif /* USE_ITT_BUILD */
3513  kmp_int32 is_constrained);
3514 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3515  kmp_flag_oncore *flag, int final_spin,
3516  int *thread_finished,
3517 #if USE_ITT_BUILD
3518  void *itt_sync_obj,
3519 #endif /* USE_ITT_BUILD */
3520  kmp_int32 is_constrained);
3521 
3522 extern void __kmp_free_task_team(kmp_info_t *thread,
3523  kmp_task_team_t *task_team);
3524 extern void __kmp_reap_task_teams(void);
3525 extern void __kmp_wait_to_unref_task_teams(void);
3526 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3527  int always);
3528 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3529 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3530 #if USE_ITT_BUILD
3531  ,
3532  void *itt_sync_obj
3533 #endif /* USE_ITT_BUILD */
3534  ,
3535  int wait = 1);
3536 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3537  int gtid);
3538 
3539 extern int __kmp_is_address_mapped(void *addr);
3540 extern kmp_uint64 __kmp_hardware_timestamp(void);
3541 
3542 #if KMP_OS_UNIX
3543 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3544 #endif
3545 
3546 /* ------------------------------------------------------------------------ */
3547 //
3548 // Assembly routines that have no compiler intrinsic replacement
3549 //
3550 
3551 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3552 
3553 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
3554 
3555 #define __kmp_load_mxcsr(p) _mm_setcsr(*(p))
3556 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
3557 
3558 extern void __kmp_load_x87_fpu_control_word(kmp_int16 *p);
3559 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
3560 extern void __kmp_clear_x87_fpu_status_word();
3561 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
3562 
3563 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3564 
3565 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3566  void *argv[]
3567 #if OMPT_SUPPORT
3568  ,
3569  void **exit_frame_ptr
3570 #endif
3571  );
3572 
3573 /* ------------------------------------------------------------------------ */
3574 
3575 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3576 KMP_EXPORT void __kmpc_end(ident_t *);
3577 
3578 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3579  kmpc_ctor_vec ctor,
3580  kmpc_cctor_vec cctor,
3581  kmpc_dtor_vec dtor,
3582  size_t vector_length);
3583 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3584  kmpc_ctor ctor, kmpc_cctor cctor,
3585  kmpc_dtor dtor);
3586 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3587  void *data, size_t size);
3588 
3589 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3590 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3591 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3592 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3593 
3594 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3595 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3596  kmpc_micro microtask, ...);
3597 
3598 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3599 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3600 
3601 KMP_EXPORT void __kmpc_flush(ident_t *);
3602 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3603 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3604 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3605 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3606 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3607 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3608  kmp_critical_name *);
3609 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3610  kmp_critical_name *);
3611 
3612 #if OMP_45_ENABLED
3613 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3614  kmp_critical_name *, uintptr_t hint);
3615 #endif
3616 
3617 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3618 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3619 
3620 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3621  kmp_int32 global_tid);
3622 
3623 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3624 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3625 
3626 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3627  kmp_int32 schedtype, kmp_int32 *plastiter,
3628  kmp_int *plower, kmp_int *pupper,
3629  kmp_int *pstride, kmp_int incr,
3630  kmp_int chunk);
3631 
3632 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3633 
3634 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3635  size_t cpy_size, void *cpy_data,
3636  void (*cpy_func)(void *, void *),
3637  kmp_int32 didit);
3638 
3639 extern void KMPC_SET_NUM_THREADS(int arg);
3640 extern void KMPC_SET_DYNAMIC(int flag);
3641 extern void KMPC_SET_NESTED(int flag);
3642 
3643 /* Taskq interface routines */
3644 KMP_EXPORT kmpc_thunk_t *__kmpc_taskq(ident_t *loc, kmp_int32 global_tid,
3645  kmpc_task_t taskq_task,
3646  size_t sizeof_thunk,
3647  size_t sizeof_shareds, kmp_int32 flags,
3648  kmpc_shared_vars_t **shareds);
3649 KMP_EXPORT void __kmpc_end_taskq(ident_t *loc, kmp_int32 global_tid,
3650  kmpc_thunk_t *thunk);
3651 KMP_EXPORT kmp_int32 __kmpc_task(ident_t *loc, kmp_int32 global_tid,
3652  kmpc_thunk_t *thunk);
3653 KMP_EXPORT void __kmpc_taskq_task(ident_t *loc, kmp_int32 global_tid,
3654  kmpc_thunk_t *thunk, kmp_int32 status);
3655 KMP_EXPORT void __kmpc_end_taskq_task(ident_t *loc, kmp_int32 global_tid,
3656  kmpc_thunk_t *thunk);
3657 KMP_EXPORT kmpc_thunk_t *__kmpc_task_buffer(ident_t *loc, kmp_int32 global_tid,
3658  kmpc_thunk_t *taskq_thunk,
3659  kmpc_task_t task);
3660 
3661 /* OMP 3.0 tasking interface routines */
3662 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3663  kmp_task_t *new_task);
3664 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3665  kmp_int32 flags,
3666  size_t sizeof_kmp_task_t,
3667  size_t sizeof_shareds,
3668  kmp_routine_entry_t task_entry);
3669 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3670  kmp_task_t *task);
3671 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3672  kmp_task_t *task);
3673 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3674  kmp_task_t *new_task);
3675 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3676 
3677 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3678  int end_part);
3679 
3680 #if TASK_UNUSED
3681 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3682 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3683  kmp_task_t *task);
3684 #endif // TASK_UNUSED
3685 
3686 /* ------------------------------------------------------------------------ */
3687 
3688 #if OMP_40_ENABLED
3689 
3690 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3691 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3692 
3693 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3694  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3695  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3696  kmp_depend_info_t *noalias_dep_list);
3697 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3698  kmp_int32 ndeps,
3699  kmp_depend_info_t *dep_list,
3700  kmp_int32 ndeps_noalias,
3701  kmp_depend_info_t *noalias_dep_list);
3702 extern void __kmp_release_deps(kmp_int32 gtid, kmp_taskdata_t *task);
3703 extern void __kmp_dephash_free_entries(kmp_info_t *thread, kmp_dephash_t *h);
3704 extern void __kmp_dephash_free(kmp_info_t *thread, kmp_dephash_t *h);
3705 
3706 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3707  bool serialize_immediate);
3708 
3709 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3710  kmp_int32 cncl_kind);
3711 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3712  kmp_int32 cncl_kind);
3713 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3714 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3715 
3716 #if OMP_45_ENABLED
3717 
3718 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3719 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3720 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3721  kmp_int32 if_val, kmp_uint64 *lb,
3722  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3723  kmp_int32 sched, kmp_uint64 grainsize,
3724  void *task_dup);
3725 #endif
3726 // TODO: change to OMP_50_ENABLED, need to change build tools for this to work
3727 #if OMP_45_ENABLED
3728 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3729 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3730 #endif
3731 
3732 #endif
3733 
3734 /* Lock interface routines (fast versions with gtid passed in) */
3735 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3736  void **user_lock);
3737 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3738  void **user_lock);
3739 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3740  void **user_lock);
3741 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3742  void **user_lock);
3743 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3744 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3745  void **user_lock);
3746 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3747  void **user_lock);
3748 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3749  void **user_lock);
3750 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3751 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3752  void **user_lock);
3753 
3754 #if OMP_45_ENABLED
3755 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3756  void **user_lock, uintptr_t hint);
3757 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3758  void **user_lock,
3759  uintptr_t hint);
3760 #endif
3761 
3762 /* Interface to fast scalable reduce methods routines */
3763 
3764 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3765  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3766  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3767  kmp_critical_name *lck);
3768 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3769  kmp_critical_name *lck);
3770 KMP_EXPORT kmp_int32 __kmpc_reduce(
3771  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3772  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3773  kmp_critical_name *lck);
3774 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3775  kmp_critical_name *lck);
3776 
3777 /* Internal fast reduction routines */
3778 
3779 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3780  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3781  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3782  kmp_critical_name *lck);
3783 
3784 // this function is for testing set/get/determine reduce method
3785 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3786 
3787 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3788 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3789 
3790 // C++ port
3791 // missing 'extern "C"' declarations
3792 
3793 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3794 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3795 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3796  kmp_int32 num_threads);
3797 
3798 #if OMP_40_ENABLED
3799 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3800  int proc_bind);
3801 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3802  kmp_int32 num_teams,
3803  kmp_int32 num_threads);
3804 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3805  kmpc_micro microtask, ...);
3806 #endif
3807 #if OMP_45_ENABLED
3808 struct kmp_dim { // loop bounds info casted to kmp_int64
3809  kmp_int64 lo; // lower
3810  kmp_int64 up; // upper
3811  kmp_int64 st; // stride
3812 };
3813 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3814  kmp_int32 num_dims,
3815  const struct kmp_dim *dims);
3816 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3817  const kmp_int64 *vec);
3818 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3819  const kmp_int64 *vec);
3820 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3821 #endif
3822 
3823 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3824  void *data, size_t size,
3825  void ***cache);
3826 
3827 // Symbols for MS mutual detection.
3828 extern int _You_must_link_with_exactly_one_OpenMP_library;
3829 extern int _You_must_link_with_Intel_OpenMP_library;
3830 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3831 extern int _You_must_link_with_Microsoft_OpenMP_library;
3832 #endif
3833 
3834 // The routines below are not exported.
3835 // Consider making them 'static' in corresponding source files.
3836 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3837  void *data_addr, size_t pc_size);
3838 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3839  void *data_addr,
3840  size_t pc_size);
3841 void __kmp_threadprivate_resize_cache(int newCapacity);
3842 void __kmp_cleanup_threadprivate_caches();
3843 
3844 // ompc_, kmpc_ entries moved from omp.h.
3845 #if KMP_OS_WINDOWS
3846 #define KMPC_CONVENTION __cdecl
3847 #else
3848 #define KMPC_CONVENTION
3849 #endif
3850 
3851 #ifndef __OMP_H
3852 typedef enum omp_sched_t {
3853  omp_sched_static = 1,
3854  omp_sched_dynamic = 2,
3855  omp_sched_guided = 3,
3856  omp_sched_auto = 4
3857 } omp_sched_t;
3858 typedef void *kmp_affinity_mask_t;
3859 #endif
3860 
3861 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3862 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3863 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3864 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3865 KMP_EXPORT int KMPC_CONVENTION
3866 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3867 KMP_EXPORT int KMPC_CONVENTION
3868 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3869 KMP_EXPORT int KMPC_CONVENTION
3870 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3871 
3872 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3873 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3874 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3875 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
3876 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
3877 
3878 #if OMP_50_ENABLED
3879 enum kmp_target_offload_kind {
3880  tgt_disabled = 0,
3881  tgt_default = 1,
3882  tgt_mandatory = 2
3883 };
3884 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3885 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
3886 extern kmp_target_offload_kind_t __kmp_target_offload;
3887 extern int __kmpc_get_target_offload();
3888 #endif
3889 
3890 #ifdef __cplusplus
3891 }
3892 #endif
3893 
3894 #endif /* KMP_H */
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
kmp_int32 reserved_2
Definition: kmp.h:211
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
void(* kmpc_dtor)(void *)
Definition: kmp.h:1420
void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st, kmp_int32 chunk)
KMP_EXPORT kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint32 *p_lb, kmp_uint32 *p_ub, kmp_int32 *p_st)
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1443
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
kmp_int32 reserved_1
Definition: kmp.h:208
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1437
KMP_EXPORT void * __kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, void *data, size_t size, void ***cache)
kmp_int32 reserved_3
Definition: kmp.h:216
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition: kmp.h:1449
KMP_EXPORT void __kmpc_flush(ident_t *)
void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint64 lb, kmp_uint64 ub, kmp_int64 st, kmp_int64 chunk)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int32 *p_lb, kmp_int32 *p_ub, kmp_int32 *p_st)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1427
KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor)
KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int64 *p_lb, kmp_int64 *p_ub, kmp_int64 *p_st)
KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *), kmp_int32 didit)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
sched_type
Definition: kmp.h:320
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
Definition: kmp.h:207
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1414
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint32 lb, kmp_uint32 ub, kmp_int32 st, kmp_int32 chunk)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1396
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, kmp_int32 global_tid)
int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint64 *p_lb, kmp_uint64 *p_ub, kmp_int64 *p_st)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length)
char const * psource
Definition: kmp.h:217
void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int64 lb, kmp_int64 ub, kmp_int64 st, kmp_int64 chunk)
kmp_int32 flags
Definition: kmp.h:209
struct ident ident_t