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 #elif KMP_OS_KFREEBSD
1056 /* TODO: tune for KMP_OS_KFREEBSD */
1057 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1058 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1059 #endif
1060 
1061 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1062 typedef struct kmp_cpuid {
1063  kmp_uint32 eax;
1064  kmp_uint32 ebx;
1065  kmp_uint32 ecx;
1066  kmp_uint32 edx;
1067 } kmp_cpuid_t;
1068 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1069 #if KMP_ARCH_X86
1070 extern void __kmp_x86_pause(void);
1071 #elif KMP_MIC
1072 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1073 // regression after removal of extra PAUSE from KMP_YIELD_SPIN(). Changing
1074 // the delay from 100 to 300 showed even better performance than double PAUSE
1075 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1076 static void __kmp_x86_pause(void) { _mm_delay_32(300); }
1077 #else
1078 static void __kmp_x86_pause(void) { _mm_pause(); }
1079 #endif
1080 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1081 #elif KMP_ARCH_PPC64
1082 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1083 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1084 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1085 #define KMP_CPU_PAUSE() \
1086  do { \
1087  KMP_PPC64_PRI_LOW(); \
1088  KMP_PPC64_PRI_MED(); \
1089  KMP_PPC64_PRI_LOC_MB(); \
1090  } while (0)
1091 #else
1092 #define KMP_CPU_PAUSE() /* nothing to do */
1093 #endif
1094 
1095 #define KMP_INIT_YIELD(count) \
1096  { (count) = __kmp_yield_init; }
1097 
1098 #define KMP_YIELD(cond) \
1099  { \
1100  KMP_CPU_PAUSE(); \
1101  __kmp_yield((cond)); \
1102  }
1103 
1104 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1105 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1106 
1107 #define KMP_YIELD_WHEN(cond, count) \
1108  { \
1109  KMP_CPU_PAUSE(); \
1110  (count) -= 2; \
1111  if (!(count)) { \
1112  __kmp_yield(cond); \
1113  (count) = __kmp_yield_next; \
1114  } \
1115  }
1116 #define KMP_YIELD_SPIN(count) \
1117  { \
1118  KMP_CPU_PAUSE(); \
1119  (count) -= 2; \
1120  if (!(count)) { \
1121  __kmp_yield(1); \
1122  (count) = __kmp_yield_next; \
1123  } \
1124  }
1125 
1126 /* ------------------------------------------------------------------------ */
1127 /* Support datatypes for the orphaned construct nesting checks. */
1128 /* ------------------------------------------------------------------------ */
1129 
1130 enum cons_type {
1131  ct_none,
1132  ct_parallel,
1133  ct_pdo,
1134  ct_pdo_ordered,
1135  ct_psections,
1136  ct_psingle,
1137 
1138  /* the following must be left in order and not split up */
1139  ct_taskq,
1140  ct_task, // really task inside non-ordered taskq, considered worksharing type
1141  ct_task_ordered, /* really task inside ordered taskq, considered a worksharing
1142  type */
1143  /* the preceding must be left in order and not split up */
1144 
1145  ct_critical,
1146  ct_ordered_in_parallel,
1147  ct_ordered_in_pdo,
1148  ct_ordered_in_taskq,
1149  ct_master,
1150  ct_reduce,
1151  ct_barrier
1152 };
1153 
1154 /* test to see if we are in a taskq construct */
1155 #define IS_CONS_TYPE_TASKQ(ct) \
1156  (((int)(ct)) >= ((int)ct_taskq) && ((int)(ct)) <= ((int)ct_task_ordered))
1157 #define IS_CONS_TYPE_ORDERED(ct) \
1158  ((ct) == ct_pdo_ordered || (ct) == ct_task_ordered)
1159 
1160 struct cons_data {
1161  ident_t const *ident;
1162  enum cons_type type;
1163  int prev;
1164  kmp_user_lock_p
1165  name; /* address exclusively for critical section name comparison */
1166 };
1167 
1168 struct cons_header {
1169  int p_top, w_top, s_top;
1170  int stack_size, stack_top;
1171  struct cons_data *stack_data;
1172 };
1173 
1174 struct kmp_region_info {
1175  char *text;
1176  int offset[KMP_MAX_FIELDS];
1177  int length[KMP_MAX_FIELDS];
1178 };
1179 
1180 /* ---------------------------------------------------------------------- */
1181 /* ---------------------------------------------------------------------- */
1182 
1183 #if KMP_OS_WINDOWS
1184 typedef HANDLE kmp_thread_t;
1185 typedef DWORD kmp_key_t;
1186 #endif /* KMP_OS_WINDOWS */
1187 
1188 #if KMP_OS_UNIX
1189 typedef pthread_t kmp_thread_t;
1190 typedef pthread_key_t kmp_key_t;
1191 #endif
1192 
1193 extern kmp_key_t __kmp_gtid_threadprivate_key;
1194 
1195 typedef struct kmp_sys_info {
1196  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1197  long minflt; /* the number of page faults serviced without any I/O */
1198  long majflt; /* the number of page faults serviced that required I/O */
1199  long nswap; /* the number of times a process was "swapped" out of memory */
1200  long inblock; /* the number of times the file system had to perform input */
1201  long oublock; /* the number of times the file system had to perform output */
1202  long nvcsw; /* the number of times a context switch was voluntarily */
1203  long nivcsw; /* the number of times a context switch was forced */
1204 } kmp_sys_info_t;
1205 
1206 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1207 typedef struct kmp_cpuinfo {
1208  int initialized; // If 0, other fields are not initialized.
1209  int signature; // CPUID(1).EAX
1210  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1211  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1212  // Model << 4 ) + Model)
1213  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1214  int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1215  int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1216  int cpu_stackoffset;
1217  int apic_id;
1218  int physical_id;
1219  int logical_id;
1220  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1221  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1222 } kmp_cpuinfo_t;
1223 #endif
1224 
1225 #if USE_ITT_BUILD
1226 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1227 // required type here. Later we will check the type meets requirements.
1228 typedef int kmp_itt_mark_t;
1229 #define KMP_ITT_DEBUG 0
1230 #endif /* USE_ITT_BUILD */
1231 
1232 /* Taskq data structures */
1233 
1234 #define HIGH_WATER_MARK(nslots) (((nslots)*3) / 4)
1235 // num thunks that each thread can simultaneously execute from a task queue
1236 #define __KMP_TASKQ_THUNKS_PER_TH 1
1237 
1238 /* flags for taskq_global_flags, kmp_task_queue_t tq_flags, kmpc_thunk_t
1239  th_flags */
1240 
1241 #define TQF_IS_ORDERED 0x0001 // __kmpc_taskq interface, taskq ordered
1242 // __kmpc_taskq interface, taskq with lastprivate list
1243 #define TQF_IS_LASTPRIVATE 0x0002
1244 #define TQF_IS_NOWAIT 0x0004 // __kmpc_taskq interface, end taskq nowait
1245 // __kmpc_taskq interface, use heuristics to decide task queue size
1246 #define TQF_HEURISTICS 0x0008
1247 
1248 // __kmpc_taskq interface, reserved for future use
1249 #define TQF_INTERFACE_RESERVED1 0x0010
1250 // __kmpc_taskq interface, reserved for future use
1251 #define TQF_INTERFACE_RESERVED2 0x0020
1252 // __kmpc_taskq interface, reserved for future use
1253 #define TQF_INTERFACE_RESERVED3 0x0040
1254 // __kmpc_taskq interface, reserved for future use
1255 #define TQF_INTERFACE_RESERVED4 0x0080
1256 
1257 #define TQF_INTERFACE_FLAGS 0x00ff // all the __kmpc_taskq interface flags
1258 // internal/read by instrumentation; only used with TQF_IS_LASTPRIVATE
1259 #define TQF_IS_LAST_TASK 0x0100
1260 // internal use only; this thunk->th_task is the taskq_task
1261 #define TQF_TASKQ_TASK 0x0200
1262 // internal use only; must release worker threads once ANY queued task
1263 // exists (global)
1264 #define TQF_RELEASE_WORKERS 0x0400
1265 // internal use only; notify workers that master has finished enqueuing tasks
1266 #define TQF_ALL_TASKS_QUEUED 0x0800
1267 // internal use only: this queue encountered in parallel context: not serialized
1268 #define TQF_PARALLEL_CONTEXT 0x1000
1269 // internal use only; this queue is on the freelist and not in use
1270 #define TQF_DEALLOCATED 0x2000
1271 
1272 #define TQF_INTERNAL_FLAGS 0x3f00 // all the internal use only flags
1273 
1274 typedef struct KMP_ALIGN_CACHE kmpc_aligned_int32_t {
1275  kmp_int32 ai_data;
1276 } kmpc_aligned_int32_t;
1277 
1278 typedef struct KMP_ALIGN_CACHE kmpc_aligned_queue_slot_t {
1279  struct kmpc_thunk_t *qs_thunk;
1280 } kmpc_aligned_queue_slot_t;
1281 
1282 typedef struct kmpc_task_queue_t {
1283  /* task queue linkage fields for n-ary tree of queues (locked with global
1284  taskq_tree_lck) */
1285  kmp_lock_t tq_link_lck; /* lock for child link, child next/prev links and
1286  child ref counts */
1287  union {
1288  struct kmpc_task_queue_t *tq_parent; // pointer to parent taskq, not locked
1289  // for taskq internal freelists, locked with global taskq_freelist_lck
1290  struct kmpc_task_queue_t *tq_next_free;
1291  } tq;
1292  // pointer to linked-list of children, locked by tq's tq_link_lck
1293  volatile struct kmpc_task_queue_t *tq_first_child;
1294  // next child in linked-list, locked by parent tq's tq_link_lck
1295  struct kmpc_task_queue_t *tq_next_child;
1296  // previous child in linked-list, locked by parent tq's tq_link_lck
1297  struct kmpc_task_queue_t *tq_prev_child;
1298  // reference count of threads with access to this task queue
1299  volatile kmp_int32 tq_ref_count;
1300  /* (other than the thread executing the kmpc_end_taskq call) */
1301  /* locked by parent tq's tq_link_lck */
1302 
1303  /* shared data for task queue */
1304  /* per-thread array of pointers to shared variable structures */
1305  struct kmpc_aligned_shared_vars_t *tq_shareds;
1306  /* only one array element exists for all but outermost taskq */
1307 
1308  /* bookkeeping for ordered task queue */
1309  kmp_uint32 tq_tasknum_queuing; // ordered task # assigned while queuing tasks
1310  // ordered number of next task to be served (executed)
1311  volatile kmp_uint32 tq_tasknum_serving;
1312 
1313  /* thunk storage management for task queue */
1314  kmp_lock_t tq_free_thunks_lck; /* lock for thunk freelist manipulation */
1315  // thunk freelist, chained via th.th_next_free
1316  struct kmpc_thunk_t *tq_free_thunks;
1317  // space allocated for thunks for this task queue
1318  struct kmpc_thunk_t *tq_thunk_space;
1319 
1320  /* data fields for queue itself */
1321  kmp_lock_t tq_queue_lck; /* lock for [de]enqueue operations: tq_queue,
1322  tq_head, tq_tail, tq_nfull */
1323  /* array of queue slots to hold thunks for tasks */
1324  kmpc_aligned_queue_slot_t *tq_queue;
1325  volatile struct kmpc_thunk_t *tq_taskq_slot; /* special slot for taskq task
1326  thunk, occupied if not NULL */
1327  kmp_int32 tq_nslots; /* # of tq_thunk_space thunks alloc'd (not incl.
1328  tq_taskq_slot space) */
1329  kmp_int32 tq_head; // enqueue puts item here (index into tq_queue array)
1330  kmp_int32 tq_tail; // dequeue takes item from here (index into tq_queue array)
1331  volatile kmp_int32 tq_nfull; // # of occupied entries in task queue right now
1332  kmp_int32 tq_hiwat; /* high-water mark for tq_nfull and queue scheduling */
1333  volatile kmp_int32 tq_flags; /* TQF_xxx */
1334 
1335  /* bookkeeping for outstanding thunks */
1336 
1337  /* per-thread array for # of regular thunks currently being executed */
1338  struct kmpc_aligned_int32_t *tq_th_thunks;
1339  kmp_int32 tq_nproc; /* number of thunks in the th_thunks array */
1340 
1341  /* statistics library bookkeeping */
1342  ident_t *tq_loc; /* source location information for taskq directive */
1343 } kmpc_task_queue_t;
1344 
1345 typedef void (*kmpc_task_t)(kmp_int32 global_tid, struct kmpc_thunk_t *thunk);
1346 
1347 /* sizeof_shareds passed as arg to __kmpc_taskq call */
1348 typedef struct kmpc_shared_vars_t { /* aligned during dynamic allocation */
1349  kmpc_task_queue_t *sv_queue; /* (pointers to) shared vars */
1350 } kmpc_shared_vars_t;
1351 
1352 typedef struct KMP_ALIGN_CACHE kmpc_aligned_shared_vars_t {
1353  volatile struct kmpc_shared_vars_t *ai_data;
1354 } kmpc_aligned_shared_vars_t;
1355 
1356 /* sizeof_thunk passed as arg to kmpc_taskq call */
1357 typedef struct kmpc_thunk_t { /* aligned during dynamic allocation */
1358  union { /* field used for internal freelists too */
1359  kmpc_shared_vars_t *th_shareds;
1360  struct kmpc_thunk_t *th_next_free; /* freelist of individual thunks within
1361  queue, head at tq_free_thunks */
1362  } th;
1363  kmpc_task_t th_task; /* taskq_task if flags & TQF_TASKQ_TASK */
1364  struct kmpc_thunk_t *th_encl_thunk; /* pointer to dynamically enclosing thunk
1365  on this thread's call stack */
1366  // TQF_xxx(tq_flags interface plus possible internal flags)
1367  kmp_int32 th_flags;
1368 
1369  kmp_int32 th_status;
1370  kmp_uint32 th_tasknum; /* task number assigned in order of queuing, used for
1371  ordered sections */
1372  /* private vars */
1373 } kmpc_thunk_t;
1374 
1375 typedef struct KMP_ALIGN_CACHE kmp_taskq {
1376  int tq_curr_thunk_capacity;
1377 
1378  kmpc_task_queue_t *tq_root;
1379  kmp_int32 tq_global_flags;
1380 
1381  kmp_lock_t tq_freelist_lck;
1382  kmpc_task_queue_t *tq_freelist;
1383 
1384  kmpc_thunk_t **tq_curr_thunk;
1385 } kmp_taskq_t;
1386 
1387 /* END Taskq data structures */
1388 
1389 typedef kmp_int32 kmp_critical_name[8];
1390 
1400 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1401 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1402  ...);
1403 
1408 /* ---------------------------------------------------------------------------
1409  */
1410 /* Threadprivate initialization/finalization function declarations */
1411 
1412 /* for non-array objects: __kmpc_threadprivate_register() */
1413 
1418 typedef void *(*kmpc_ctor)(void *);
1419 
1424 typedef void (*kmpc_dtor)(
1425  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1426  compiler */
1431 typedef void *(*kmpc_cctor)(void *, void *);
1432 
1433 /* for array objects: __kmpc_threadprivate_register_vec() */
1434 /* First arg: "this" pointer */
1435 /* Last arg: number of array elements */
1441 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1447 typedef void (*kmpc_dtor_vec)(void *, size_t);
1453 typedef void *(*kmpc_cctor_vec)(void *, void *,
1454  size_t); /* function unused by compiler */
1455 
1460 /* keeps tracked of threadprivate cache allocations for cleanup later */
1461 typedef struct kmp_cached_addr {
1462  void **addr; /* address of allocated cache */
1463  void ***compiler_cache; /* pointer to compiler's cache */
1464  void *data; /* pointer to global data */
1465  struct kmp_cached_addr *next; /* pointer to next cached address */
1466 } kmp_cached_addr_t;
1467 
1468 struct private_data {
1469  struct private_data *next; /* The next descriptor in the list */
1470  void *data; /* The data buffer for this descriptor */
1471  int more; /* The repeat count for this descriptor */
1472  size_t size; /* The data size for this descriptor */
1473 };
1474 
1475 struct private_common {
1476  struct private_common *next;
1477  struct private_common *link;
1478  void *gbl_addr;
1479  void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1480  size_t cmn_size;
1481 };
1482 
1483 struct shared_common {
1484  struct shared_common *next;
1485  struct private_data *pod_init;
1486  void *obj_init;
1487  void *gbl_addr;
1488  union {
1489  kmpc_ctor ctor;
1490  kmpc_ctor_vec ctorv;
1491  } ct;
1492  union {
1493  kmpc_cctor cctor;
1494  kmpc_cctor_vec cctorv;
1495  } cct;
1496  union {
1497  kmpc_dtor dtor;
1498  kmpc_dtor_vec dtorv;
1499  } dt;
1500  size_t vec_len;
1501  int is_vec;
1502  size_t cmn_size;
1503 };
1504 
1505 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1506 #define KMP_HASH_TABLE_SIZE \
1507  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1508 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1509 #define KMP_HASH(x) \
1510  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1511 
1512 struct common_table {
1513  struct private_common *data[KMP_HASH_TABLE_SIZE];
1514 };
1515 
1516 struct shared_table {
1517  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1518 };
1519 
1520 /* ------------------------------------------------------------------------ */
1521 
1522 #if KMP_USE_HIER_SCHED
1523 // Shared barrier data that exists inside a single unit of the scheduling
1524 // hierarchy
1525 typedef struct kmp_hier_private_bdata_t {
1526  kmp_int32 num_active;
1527  kmp_uint64 index;
1528  kmp_uint64 wait_val[2];
1529 } kmp_hier_private_bdata_t;
1530 #endif
1531 
1532 typedef struct kmp_sched_flags {
1533  unsigned ordered : 1;
1534  unsigned nomerge : 1;
1535  unsigned contains_last : 1;
1536 #if KMP_USE_HIER_SCHED
1537  unsigned use_hier : 1;
1538  unsigned unused : 28;
1539 #else
1540  unsigned unused : 29;
1541 #endif
1542 } kmp_sched_flags_t;
1543 
1544 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1545 
1546 #if KMP_STATIC_STEAL_ENABLED
1547 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1548  kmp_int32 count;
1549  kmp_int32 ub;
1550  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1551  kmp_int32 lb;
1552  kmp_int32 st;
1553  kmp_int32 tc;
1554  kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1555  after ub */
1556 
1557  // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1558  // a) parm3 is properly aligned and
1559  // b) all parm1-4 are in the same cache line.
1560  // Because of parm1-4 are used together, performance seems to be better
1561  // if they are in the same line (not measured though).
1562 
1563  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1564  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1565  kmp_int32 parm2; // make no real change at least while padding is off.
1566  kmp_int32 parm3;
1567  kmp_int32 parm4;
1568  };
1569 
1570  kmp_uint32 ordered_lower;
1571  kmp_uint32 ordered_upper;
1572 #if KMP_OS_WINDOWS
1573  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1574  // 'static_steal_counter'. It would be nice to measure execution times.
1575  // Conditional if/endif can be removed at all.
1576  kmp_int32 last_upper;
1577 #endif /* KMP_OS_WINDOWS */
1578 } dispatch_private_info32_t;
1579 
1580 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1581  kmp_int64 count; // current chunk number for static & static-steal scheduling
1582  kmp_int64 ub; /* upper-bound */
1583  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1584  kmp_int64 lb; /* lower-bound */
1585  kmp_int64 st; /* stride */
1586  kmp_int64 tc; /* trip count (number of iterations) */
1587  kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1588  after ub */
1589 
1590  /* parm[1-4] are used in different ways by different scheduling algorithms */
1591 
1592  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1593  // a) parm3 is properly aligned and
1594  // b) all parm1-4 are in the same cache line.
1595  // Because of parm1-4 are used together, performance seems to be better
1596  // if they are in the same line (not measured though).
1597 
1598  struct KMP_ALIGN(32) {
1599  kmp_int64 parm1;
1600  kmp_int64 parm2;
1601  kmp_int64 parm3;
1602  kmp_int64 parm4;
1603  };
1604 
1605  kmp_uint64 ordered_lower;
1606  kmp_uint64 ordered_upper;
1607 #if KMP_OS_WINDOWS
1608  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1609  // 'static_steal_counter'. It would be nice to measure execution times.
1610  // Conditional if/endif can be removed at all.
1611  kmp_int64 last_upper;
1612 #endif /* KMP_OS_WINDOWS */
1613 } dispatch_private_info64_t;
1614 #else /* KMP_STATIC_STEAL_ENABLED */
1615 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1616  kmp_int32 lb;
1617  kmp_int32 ub;
1618  kmp_int32 st;
1619  kmp_int32 tc;
1620 
1621  kmp_int32 parm1;
1622  kmp_int32 parm2;
1623  kmp_int32 parm3;
1624  kmp_int32 parm4;
1625 
1626  kmp_int32 count;
1627 
1628  kmp_uint32 ordered_lower;
1629  kmp_uint32 ordered_upper;
1630 #if KMP_OS_WINDOWS
1631  kmp_int32 last_upper;
1632 #endif /* KMP_OS_WINDOWS */
1633 } dispatch_private_info32_t;
1634 
1635 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1636  kmp_int64 lb; /* lower-bound */
1637  kmp_int64 ub; /* upper-bound */
1638  kmp_int64 st; /* stride */
1639  kmp_int64 tc; /* trip count (number of iterations) */
1640 
1641  /* parm[1-4] are used in different ways by different scheduling algorithms */
1642  kmp_int64 parm1;
1643  kmp_int64 parm2;
1644  kmp_int64 parm3;
1645  kmp_int64 parm4;
1646 
1647  kmp_int64 count; /* current chunk number for static scheduling */
1648 
1649  kmp_uint64 ordered_lower;
1650  kmp_uint64 ordered_upper;
1651 #if KMP_OS_WINDOWS
1652  kmp_int64 last_upper;
1653 #endif /* KMP_OS_WINDOWS */
1654 } dispatch_private_info64_t;
1655 #endif /* KMP_STATIC_STEAL_ENABLED */
1656 
1657 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1658  union private_info {
1659  dispatch_private_info32_t p32;
1660  dispatch_private_info64_t p64;
1661  } u;
1662  enum sched_type schedule; /* scheduling algorithm */
1663  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1664  kmp_int32 ordered_bumped;
1665  // To retain the structure size after making ordered_iteration scalar
1666  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1667  // Stack of buffers for nest of serial regions
1668  struct dispatch_private_info *next;
1669  kmp_int32 type_size; /* the size of types in private_info */
1670 #if KMP_USE_HIER_SCHED
1671  kmp_int32 hier_id;
1672  void *parent; /* hierarchical scheduling parent pointer */
1673 #endif
1674  enum cons_type pushed_ws;
1675 } dispatch_private_info_t;
1676 
1677 typedef struct dispatch_shared_info32 {
1678  /* chunk index under dynamic, number of idle threads under static-steal;
1679  iteration index otherwise */
1680  volatile kmp_uint32 iteration;
1681  volatile kmp_uint32 num_done;
1682  volatile kmp_uint32 ordered_iteration;
1683  // Dummy to retain the structure size after making ordered_iteration scalar
1684  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1685 } dispatch_shared_info32_t;
1686 
1687 typedef struct dispatch_shared_info64 {
1688  /* chunk index under dynamic, number of idle threads under static-steal;
1689  iteration index otherwise */
1690  volatile kmp_uint64 iteration;
1691  volatile kmp_uint64 num_done;
1692  volatile kmp_uint64 ordered_iteration;
1693  // Dummy to retain the structure size after making ordered_iteration scalar
1694  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1695 } dispatch_shared_info64_t;
1696 
1697 typedef struct dispatch_shared_info {
1698  union shared_info {
1699  dispatch_shared_info32_t s32;
1700  dispatch_shared_info64_t s64;
1701  } u;
1702  volatile kmp_uint32 buffer_index;
1703 #if OMP_45_ENABLED
1704  volatile kmp_int32 doacross_buf_idx; // teamwise index
1705  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1706  kmp_int32 doacross_num_done; // count finished threads
1707 #endif
1708 #if KMP_USE_HIER_SCHED
1709  void *hier;
1710 #endif
1711 #if KMP_USE_HWLOC
1712  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1713  // machines (> 48 cores). Performance analysis showed that a cache thrash
1714  // was occurring and this padding helps alleviate the problem.
1715  char padding[64];
1716 #endif
1717 } dispatch_shared_info_t;
1718 
1719 typedef struct kmp_disp {
1720  /* Vector for ORDERED SECTION */
1721  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1722  /* Vector for END ORDERED SECTION */
1723  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1724 
1725  dispatch_shared_info_t *th_dispatch_sh_current;
1726  dispatch_private_info_t *th_dispatch_pr_current;
1727 
1728  dispatch_private_info_t *th_disp_buffer;
1729  kmp_int32 th_disp_index;
1730 #if OMP_45_ENABLED
1731  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1732  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1733  union { // we can use union here because doacross cannot be used in
1734  // nonmonotonic loops
1735  kmp_int64 *th_doacross_info; // info on loop bounds
1736  kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1737  };
1738 #else
1739 #if KMP_STATIC_STEAL_ENABLED
1740  kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1741  void *dummy_padding[1]; // make it 64 bytes on Intel(R) 64
1742 #else
1743  void *dummy_padding[2]; // make it 64 bytes on Intel(R) 64
1744 #endif
1745 #endif
1746 #if KMP_USE_INTERNODE_ALIGNMENT
1747  char more_padding[INTERNODE_CACHE_LINE];
1748 #endif
1749 } kmp_disp_t;
1750 
1751 /* ------------------------------------------------------------------------ */
1752 /* Barrier stuff */
1753 
1754 /* constants for barrier state update */
1755 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1756 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1757 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1758 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1759 
1760 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1761 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1762 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1763 
1764 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1765 #error "Barrier sleep bit must be smaller than barrier bump bit"
1766 #endif
1767 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1768 #error "Barrier unused bit must be smaller than barrier bump bit"
1769 #endif
1770 
1771 // Constants for release barrier wait state: currently, hierarchical only
1772 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1773 #define KMP_BARRIER_OWN_FLAG \
1774  1 // Normal state; worker waiting on own b_go flag in release
1775 #define KMP_BARRIER_PARENT_FLAG \
1776  2 // Special state; worker waiting on parent's b_go flag in release
1777 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1778  3 // Special state; tells worker to shift from parent to own b_go
1779 #define KMP_BARRIER_SWITCHING \
1780  4 // Special state; worker resets appropriate flag on wake-up
1781 
1782 #define KMP_NOT_SAFE_TO_REAP \
1783  0 // Thread th_reap_state: not safe to reap (tasking)
1784 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1785 
1786 enum barrier_type {
1787  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1788  barriers if enabled) */
1789  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1790 #if KMP_FAST_REDUCTION_BARRIER
1791  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1792 #endif // KMP_FAST_REDUCTION_BARRIER
1793  bs_last_barrier /* Just a placeholder to mark the end */
1794 };
1795 
1796 // to work with reduction barriers just like with plain barriers
1797 #if !KMP_FAST_REDUCTION_BARRIER
1798 #define bs_reduction_barrier bs_plain_barrier
1799 #endif // KMP_FAST_REDUCTION_BARRIER
1800 
1801 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1802  bp_linear_bar =
1803  0, /* Single level (degenerate) tree */
1804  bp_tree_bar =
1805  1, /* Balanced tree with branching factor 2^n */
1806  bp_hyper_bar =
1807  2, /* Hypercube-embedded tree with min branching
1808  factor 2^n */
1809  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1810  bp_last_bar /* Placeholder to mark the end */
1811 } kmp_bar_pat_e;
1812 
1813 #define KMP_BARRIER_ICV_PUSH 1
1814 
1815 /* Record for holding the values of the internal controls stack records */
1816 typedef struct kmp_internal_control {
1817  int serial_nesting_level; /* corresponds to the value of the
1818  th_team_serialized field */
1819  kmp_int8 nested; /* internal control for nested parallelism (per thread) */
1820  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1821  thread) */
1822  kmp_int8
1823  bt_set; /* internal control for whether blocktime is explicitly set */
1824  int blocktime; /* internal control for blocktime */
1825 #if KMP_USE_MONITOR
1826  int bt_intervals; /* internal control for blocktime intervals */
1827 #endif
1828  int nproc; /* internal control for #threads for next parallel region (per
1829  thread) */
1830  int max_active_levels; /* internal control for max_active_levels */
1831  kmp_r_sched_t
1832  sched; /* internal control for runtime schedule {sched,chunk} pair */
1833 #if OMP_40_ENABLED
1834  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1835  kmp_int32 default_device; /* internal control for default device */
1836 #endif // OMP_40_ENABLED
1837  struct kmp_internal_control *next;
1838 } kmp_internal_control_t;
1839 
1840 static inline void copy_icvs(kmp_internal_control_t *dst,
1841  kmp_internal_control_t *src) {
1842  *dst = *src;
1843 }
1844 
1845 /* Thread barrier needs volatile barrier fields */
1846 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1847  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1848  // uses of it). It is not explicitly aligned below, because we *don't* want
1849  // it to be padded -- instead, we fit b_go into the same cache line with
1850  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1851  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1852  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1853  // same NGO store
1854  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1855  KMP_ALIGN_CACHE volatile kmp_uint64
1856  b_arrived; // STATE => task reached synch point.
1857  kmp_uint32 *skip_per_level;
1858  kmp_uint32 my_level;
1859  kmp_int32 parent_tid;
1860  kmp_int32 old_tid;
1861  kmp_uint32 depth;
1862  struct kmp_bstate *parent_bar;
1863  kmp_team_t *team;
1864  kmp_uint64 leaf_state;
1865  kmp_uint32 nproc;
1866  kmp_uint8 base_leaf_kids;
1867  kmp_uint8 leaf_kids;
1868  kmp_uint8 offset;
1869  kmp_uint8 wait_flag;
1870  kmp_uint8 use_oncore_barrier;
1871 #if USE_DEBUGGER
1872  // The following field is intended for the debugger solely. Only the worker
1873  // thread itself accesses this field: the worker increases it by 1 when it
1874  // arrives to a barrier.
1875  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1876 #endif /* USE_DEBUGGER */
1877 } kmp_bstate_t;
1878 
1879 union KMP_ALIGN_CACHE kmp_barrier_union {
1880  double b_align; /* use worst case alignment */
1881  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1882  kmp_bstate_t bb;
1883 };
1884 
1885 typedef union kmp_barrier_union kmp_balign_t;
1886 
1887 /* Team barrier needs only non-volatile arrived counter */
1888 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1889  double b_align; /* use worst case alignment */
1890  char b_pad[CACHE_LINE];
1891  struct {
1892  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1893 #if USE_DEBUGGER
1894  // The following two fields are indended for the debugger solely. Only
1895  // master of the team accesses these fields: the first one is increased by
1896  // 1 when master arrives to a barrier, the second one is increased by one
1897  // when all the threads arrived.
1898  kmp_uint b_master_arrived;
1899  kmp_uint b_team_arrived;
1900 #endif
1901  };
1902 };
1903 
1904 typedef union kmp_barrier_team_union kmp_balign_team_t;
1905 
1906 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
1907  threads when a condition changes. This is to workaround an NPTL bug where
1908  padding was added to pthread_cond_t which caused the initialization routine
1909  to write outside of the structure if compiled on pre-NPTL threads. */
1910 #if KMP_OS_WINDOWS
1911 typedef struct kmp_win32_mutex {
1912  /* The Lock */
1913  CRITICAL_SECTION cs;
1914 } kmp_win32_mutex_t;
1915 
1916 typedef struct kmp_win32_cond {
1917  /* Count of the number of waiters. */
1918  int waiters_count_;
1919 
1920  /* Serialize access to <waiters_count_> */
1921  kmp_win32_mutex_t waiters_count_lock_;
1922 
1923  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
1924  int release_count_;
1925 
1926  /* Keeps track of the current "generation" so that we don't allow */
1927  /* one thread to steal all the "releases" from the broadcast. */
1928  int wait_generation_count_;
1929 
1930  /* A manual-reset event that's used to block and release waiting threads. */
1931  HANDLE event_;
1932 } kmp_win32_cond_t;
1933 #endif
1934 
1935 #if KMP_OS_UNIX
1936 
1937 union KMP_ALIGN_CACHE kmp_cond_union {
1938  double c_align;
1939  char c_pad[CACHE_LINE];
1940  pthread_cond_t c_cond;
1941 };
1942 
1943 typedef union kmp_cond_union kmp_cond_align_t;
1944 
1945 union KMP_ALIGN_CACHE kmp_mutex_union {
1946  double m_align;
1947  char m_pad[CACHE_LINE];
1948  pthread_mutex_t m_mutex;
1949 };
1950 
1951 typedef union kmp_mutex_union kmp_mutex_align_t;
1952 
1953 #endif /* KMP_OS_UNIX */
1954 
1955 typedef struct kmp_desc_base {
1956  void *ds_stackbase;
1957  size_t ds_stacksize;
1958  int ds_stackgrow;
1959  kmp_thread_t ds_thread;
1960  volatile int ds_tid;
1961  int ds_gtid;
1962 #if KMP_OS_WINDOWS
1963  volatile int ds_alive;
1964  DWORD ds_thread_id;
1965 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
1966  However, debugger support (libomp_db) cannot work with handles, because they
1967  uncomparable. For example, debugger requests info about thread with handle h.
1968  h is valid within debugger process, and meaningless within debugee process.
1969  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
1970  within debugee process, but it is a *new* handle which does *not* equal to
1971  any other handle in debugee... The only way to compare handles is convert
1972  them to system-wide ids. GetThreadId() function is available only in
1973  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
1974  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
1975  thread id by call to GetCurrentThreadId() from within the thread and save it
1976  to let libomp_db identify threads. */
1977 #endif /* KMP_OS_WINDOWS */
1978 } kmp_desc_base_t;
1979 
1980 typedef union KMP_ALIGN_CACHE kmp_desc {
1981  double ds_align; /* use worst case alignment */
1982  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1983  kmp_desc_base_t ds;
1984 } kmp_desc_t;
1985 
1986 typedef struct kmp_local {
1987  volatile int this_construct; /* count of single's encountered by thread */
1988  void *reduce_data;
1989 #if KMP_USE_BGET
1990  void *bget_data;
1991  void *bget_list;
1992 #if !USE_CMP_XCHG_FOR_BGET
1993 #ifdef USE_QUEUING_LOCK_FOR_BGET
1994  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
1995 #else
1996  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
1997 // bootstrap lock so we can use it at library
1998 // shutdown.
1999 #endif /* USE_LOCK_FOR_BGET */
2000 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2001 #endif /* KMP_USE_BGET */
2002 
2003  PACKED_REDUCTION_METHOD_T
2004  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2005  __kmpc_end_reduce*() */
2006 
2007 } kmp_local_t;
2008 
2009 #define KMP_CHECK_UPDATE(a, b) \
2010  if ((a) != (b)) \
2011  (a) = (b)
2012 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2013  if ((a) != (b)) \
2014  TCW_SYNC_PTR((a), (b))
2015 
2016 #define get__blocktime(xteam, xtid) \
2017  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2018 #define get__bt_set(xteam, xtid) \
2019  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2020 #if KMP_USE_MONITOR
2021 #define get__bt_intervals(xteam, xtid) \
2022  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2023 #endif
2024 
2025 #define get__nested_2(xteam, xtid) \
2026  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nested)
2027 #define get__dynamic_2(xteam, xtid) \
2028  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2029 #define get__nproc_2(xteam, xtid) \
2030  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2031 #define get__sched_2(xteam, xtid) \
2032  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2033 
2034 #define set__blocktime_team(xteam, xtid, xval) \
2035  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2036  (xval))
2037 
2038 #if KMP_USE_MONITOR
2039 #define set__bt_intervals_team(xteam, xtid, xval) \
2040  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2041  (xval))
2042 #endif
2043 
2044 #define set__bt_set_team(xteam, xtid, xval) \
2045  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2046 
2047 #define set__nested(xthread, xval) \
2048  (((xthread)->th.th_current_task->td_icvs.nested) = (xval))
2049 #define get__nested(xthread) \
2050  (((xthread)->th.th_current_task->td_icvs.nested) ? (FTN_TRUE) : (FTN_FALSE))
2051 
2052 #define set__dynamic(xthread, xval) \
2053  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2054 #define get__dynamic(xthread) \
2055  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2056 
2057 #define set__nproc(xthread, xval) \
2058  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2059 
2060 #define set__max_active_levels(xthread, xval) \
2061  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2062 
2063 #define set__sched(xthread, xval) \
2064  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2065 
2066 #if OMP_40_ENABLED
2067 
2068 #define set__proc_bind(xthread, xval) \
2069  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2070 #define get__proc_bind(xthread) \
2071  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2072 
2073 #endif /* OMP_40_ENABLED */
2074 
2075 // OpenMP tasking data structures
2076 
2077 typedef enum kmp_tasking_mode {
2078  tskm_immediate_exec = 0,
2079  tskm_extra_barrier = 1,
2080  tskm_task_teams = 2,
2081  tskm_max = 2
2082 } kmp_tasking_mode_t;
2083 
2084 extern kmp_tasking_mode_t
2085  __kmp_tasking_mode; /* determines how/when to execute tasks */
2086 extern int __kmp_task_stealing_constraint;
2087 #if OMP_40_ENABLED
2088 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2089 // specified, defaults to 0 otherwise
2090 #endif
2091 #if OMP_45_ENABLED
2092 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2093 extern kmp_int32 __kmp_max_task_priority;
2094 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2095 extern kmp_uint64 __kmp_taskloop_min_tasks;
2096 #endif
2097 
2098 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2099  taskdata first */
2100 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2101 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2102 
2103 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2104 // were spawned and queued since the previous barrier release.
2105 #define KMP_TASKING_ENABLED(task_team) \
2106  (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2107 
2114 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2115 
2116 #if OMP_40_ENABLED || OMP_45_ENABLED
2117 typedef union kmp_cmplrdata {
2118 #if OMP_45_ENABLED
2119  kmp_int32 priority;
2120 #endif // OMP_45_ENABLED
2121 #if OMP_40_ENABLED
2122  kmp_routine_entry_t
2123  destructors; /* pointer to function to invoke deconstructors of
2124  firstprivate C++ objects */
2125 #endif // OMP_40_ENABLED
2126  /* future data */
2127 } kmp_cmplrdata_t;
2128 #endif
2129 
2130 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2133 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2134  void *shareds;
2135  kmp_routine_entry_t
2136  routine;
2137  kmp_int32 part_id;
2138 #if OMP_40_ENABLED || OMP_45_ENABLED
2139  kmp_cmplrdata_t
2140  data1; /* Two known optional additions: destructors and priority */
2141  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2142 /* future data */
2143 #endif
2144  /* private vars */
2145 } kmp_task_t;
2146 
2151 #if OMP_40_ENABLED
2152 typedef struct kmp_taskgroup {
2153  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2154  std::atomic<kmp_int32>
2155  cancel_request; // request for cancellation of this taskgroup
2156  struct kmp_taskgroup *parent; // parent taskgroup
2157 // TODO: change to OMP_50_ENABLED, need to change build tools for this to work
2158 #if OMP_45_ENABLED
2159  // Block of data to perform task reduction
2160  void *reduce_data; // reduction related info
2161  kmp_int32 reduce_num_data; // number of data items to reduce
2162 #endif
2163 } kmp_taskgroup_t;
2164 
2165 // forward declarations
2166 typedef union kmp_depnode kmp_depnode_t;
2167 typedef struct kmp_depnode_list kmp_depnode_list_t;
2168 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2169 
2170 typedef struct kmp_depend_info {
2171  kmp_intptr_t base_addr;
2172  size_t len;
2173  struct {
2174  bool in : 1;
2175  bool out : 1;
2176  } flags;
2177 } kmp_depend_info_t;
2178 
2179 struct kmp_depnode_list {
2180  kmp_depnode_t *node;
2181  kmp_depnode_list_t *next;
2182 };
2183 
2184 typedef struct kmp_base_depnode {
2185  kmp_depnode_list_t *successors;
2186  kmp_task_t *task;
2187 
2188  kmp_lock_t lock;
2189 
2190 #if KMP_SUPPORT_GRAPH_OUTPUT
2191  kmp_uint32 id;
2192 #endif
2193 
2194  std::atomic<kmp_int32> npredecessors;
2195  std::atomic<kmp_int32> nrefs;
2196 } kmp_base_depnode_t;
2197 
2198 union KMP_ALIGN_CACHE kmp_depnode {
2199  double dn_align; /* use worst case alignment */
2200  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2201  kmp_base_depnode_t dn;
2202 };
2203 
2204 struct kmp_dephash_entry {
2205  kmp_intptr_t addr;
2206  kmp_depnode_t *last_out;
2207  kmp_depnode_list_t *last_ins;
2208  kmp_dephash_entry_t *next_in_bucket;
2209 };
2210 
2211 typedef struct kmp_dephash {
2212  kmp_dephash_entry_t **buckets;
2213  size_t size;
2214 #ifdef KMP_DEBUG
2215  kmp_uint32 nelements;
2216  kmp_uint32 nconflicts;
2217 #endif
2218 } kmp_dephash_t;
2219 
2220 #endif
2221 
2222 #ifdef BUILD_TIED_TASK_STACK
2223 
2224 /* Tied Task stack definitions */
2225 typedef struct kmp_stack_block {
2226  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2227  struct kmp_stack_block *sb_next;
2228  struct kmp_stack_block *sb_prev;
2229 } kmp_stack_block_t;
2230 
2231 typedef struct kmp_task_stack {
2232  kmp_stack_block_t ts_first_block; // first block of stack entries
2233  kmp_taskdata_t **ts_top; // pointer to the top of stack
2234  kmp_int32 ts_entries; // number of entries on the stack
2235 } kmp_task_stack_t;
2236 
2237 #endif // BUILD_TIED_TASK_STACK
2238 
2239 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2240  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2241  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2242  unsigned final : 1; /* task is final(1) so execute immediately */
2243  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2244  code path */
2245 #if OMP_40_ENABLED
2246  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2247  invoke destructors from the runtime */
2248 #if OMP_45_ENABLED
2249  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2250  context of the RTL) */
2251  unsigned priority_specified : 1; /* set if the compiler provides priority
2252  setting for the task */
2253  unsigned reserved : 10; /* reserved for compiler use */
2254 #else
2255  unsigned reserved : 12; /* reserved for compiler use */
2256 #endif
2257 #else // OMP_40_ENABLED
2258  unsigned reserved : 13; /* reserved for compiler use */
2259 #endif // OMP_40_ENABLED
2260 
2261  /* Library flags */ /* Total library flags must be 16 bits */
2262  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2263  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2264  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2265  // (1) or may be deferred (0)
2266  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2267  // (0) [>= 2 threads]
2268  /* If either team_serial or tasking_ser is set, task team may be NULL */
2269  /* Task State Flags: */
2270  unsigned started : 1; /* 1==started, 0==not started */
2271  unsigned executing : 1; /* 1==executing, 0==not executing */
2272  unsigned complete : 1; /* 1==complete, 0==not complete */
2273  unsigned freed : 1; /* 1==freed, 0==allocateed */
2274  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2275  unsigned reserved31 : 7; /* reserved for library use */
2276 
2277 } kmp_tasking_flags_t;
2278 
2279 struct kmp_taskdata { /* aligned during dynamic allocation */
2280  kmp_int32 td_task_id; /* id, assigned by debugger */
2281  kmp_tasking_flags_t td_flags; /* task flags */
2282  kmp_team_t *td_team; /* team for this task */
2283  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2284  /* Currently not used except for perhaps IDB */
2285  kmp_taskdata_t *td_parent; /* parent task */
2286  kmp_int32 td_level; /* task nesting level */
2287  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2288  ident_t *td_ident; /* task identifier */
2289  // Taskwait data.
2290  ident_t *td_taskwait_ident;
2291  kmp_uint32 td_taskwait_counter;
2292  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2293  KMP_ALIGN_CACHE kmp_internal_control_t
2294  td_icvs; /* Internal control variables for the task */
2295  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2296  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2297  deallocated */
2298  std::atomic<kmp_int32>
2299  td_incomplete_child_tasks; /* Child tasks not yet complete */
2300 #if OMP_40_ENABLED
2301  kmp_taskgroup_t
2302  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2303  kmp_dephash_t
2304  *td_dephash; // Dependencies for children tasks are tracked from here
2305  kmp_depnode_t
2306  *td_depnode; // Pointer to graph node if this task has dependencies
2307 #endif // OMP_40_ENABLED
2308 #if OMP_45_ENABLED
2309  kmp_task_team_t *td_task_team;
2310  kmp_int32 td_size_alloc; // The size of task structure, including shareds etc.
2311 #if defined(KMP_GOMP_COMPAT)
2312  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2313  kmp_int32 td_size_loop_bounds;
2314 #endif
2315 #endif // OMP_45_ENABLED
2316  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2317 #if defined(KMP_GOMP_COMPAT) && OMP_45_ENABLED
2318  // GOMP sends in a copy function for copy constructors
2319  void (*td_copy_func)(void *, void *);
2320 #endif
2321 #if OMPT_SUPPORT
2322  ompt_task_info_t ompt_task_info;
2323 #endif
2324 }; // struct kmp_taskdata
2325 
2326 // Make sure padding above worked
2327 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2328 
2329 // Data for task team but per thread
2330 typedef struct kmp_base_thread_data {
2331  kmp_info_p *td_thr; // Pointer back to thread info
2332  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2333  // queued?
2334  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2335  kmp_taskdata_t *
2336  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2337  kmp_int32 td_deque_size; // Size of deck
2338  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2339  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2340  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2341  // GEH: shouldn't this be volatile since used in while-spin?
2342  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2343 #ifdef BUILD_TIED_TASK_STACK
2344  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2345 // scheduling constraint
2346 #endif // BUILD_TIED_TASK_STACK
2347 } kmp_base_thread_data_t;
2348 
2349 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2350 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2351 
2352 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2353 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2354 
2355 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2356  kmp_base_thread_data_t td;
2357  double td_align; /* use worst case alignment */
2358  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2359 } kmp_thread_data_t;
2360 
2361 // Data for task teams which are used when tasking is enabled for the team
2362 typedef struct kmp_base_task_team {
2363  kmp_bootstrap_lock_t
2364  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2365  /* must be bootstrap lock since used at library shutdown*/
2366  kmp_task_team_t *tt_next; /* For linking the task team free list */
2367  kmp_thread_data_t
2368  *tt_threads_data; /* Array of per-thread structures for task team */
2369  /* Data survives task team deallocation */
2370  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2371  executing this team? */
2372  /* TRUE means tt_threads_data is set up and initialized */
2373  kmp_int32 tt_nproc; /* #threads in team */
2374  kmp_int32
2375  tt_max_threads; /* number of entries allocated for threads_data array */
2376 #if OMP_45_ENABLED
2377  kmp_int32
2378  tt_found_proxy_tasks; /* Have we found proxy tasks since last barrier */
2379 #endif
2380  kmp_int32 tt_untied_task_encountered;
2381 
2382  KMP_ALIGN_CACHE
2383  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2384 
2385  KMP_ALIGN_CACHE
2386  volatile kmp_uint32
2387  tt_active; /* is the team still actively executing tasks */
2388 } kmp_base_task_team_t;
2389 
2390 union KMP_ALIGN_CACHE kmp_task_team {
2391  kmp_base_task_team_t tt;
2392  double tt_align; /* use worst case alignment */
2393  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2394 };
2395 
2396 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2397 // Free lists keep same-size free memory slots for fast memory allocation
2398 // routines
2399 typedef struct kmp_free_list {
2400  void *th_free_list_self; // Self-allocated tasks free list
2401  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2402  // threads
2403  void *th_free_list_other; // Non-self free list (to be returned to owner's
2404  // sync list)
2405 } kmp_free_list_t;
2406 #endif
2407 #if KMP_NESTED_HOT_TEAMS
2408 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2409 // are not put in teams pool, and they don't put threads in threads pool.
2410 typedef struct kmp_hot_team_ptr {
2411  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2412  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2413 } kmp_hot_team_ptr_t;
2414 #endif
2415 #if OMP_40_ENABLED
2416 typedef struct kmp_teams_size {
2417  kmp_int32 nteams; // number of teams in a league
2418  kmp_int32 nth; // number of threads in each team of the league
2419 } kmp_teams_size_t;
2420 #endif
2421 
2422 // OpenMP thread data structures
2423 
2424 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2425  /* Start with the readonly data which is cache aligned and padded. This is
2426  written before the thread starts working by the master. Uber masters may
2427  update themselves later. Usage does not consider serialized regions. */
2428  kmp_desc_t th_info;
2429  kmp_team_p *th_team; /* team we belong to */
2430  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2431  kmp_info_p *th_next_pool; /* next available thread in the pool */
2432  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2433  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2434 
2435  /* The following are cached from the team info structure */
2436  /* TODO use these in more places as determined to be needed via profiling */
2437  int th_team_nproc; /* number of threads in a team */
2438  kmp_info_p *th_team_master; /* the team's master thread */
2439  int th_team_serialized; /* team is serialized */
2440 #if OMP_40_ENABLED
2441  microtask_t th_teams_microtask; /* save entry address for teams construct */
2442  int th_teams_level; /* save initial level of teams construct */
2443 /* it is 0 on device but may be any on host */
2444 #endif
2445 
2446 /* The blocktime info is copied from the team struct to the thread sruct */
2447 /* at the start of a barrier, and the values stored in the team are used */
2448 /* at points in the code where the team struct is no longer guaranteed */
2449 /* to exist (from the POV of worker threads). */
2450 #if KMP_USE_MONITOR
2451  int th_team_bt_intervals;
2452  int th_team_bt_set;
2453 #else
2454  kmp_uint64 th_team_bt_intervals;
2455 #endif
2456 
2457 #if KMP_AFFINITY_SUPPORTED
2458  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2459 #endif
2460 
2461  /* The data set by the master at reinit, then R/W by the worker */
2462  KMP_ALIGN_CACHE int
2463  th_set_nproc; /* if > 0, then only use this request for the next fork */
2464 #if KMP_NESTED_HOT_TEAMS
2465  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2466 #endif
2467 #if OMP_40_ENABLED
2468  kmp_proc_bind_t
2469  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2470  kmp_teams_size_t
2471  th_teams_size; /* number of teams/threads in teams construct */
2472 #if KMP_AFFINITY_SUPPORTED
2473  int th_current_place; /* place currently bound to */
2474  int th_new_place; /* place to bind to in par reg */
2475  int th_first_place; /* first place in partition */
2476  int th_last_place; /* last place in partition */
2477 #endif
2478 #endif
2479 #if USE_ITT_BUILD
2480  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2481  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2482  kmp_uint64 th_frame_time; /* frame timestamp */
2483 #endif /* USE_ITT_BUILD */
2484  kmp_local_t th_local;
2485  struct private_common *th_pri_head;
2486 
2487  /* Now the data only used by the worker (after initial allocation) */
2488  /* TODO the first serial team should actually be stored in the info_t
2489  structure. this will help reduce initial allocation overhead */
2490  KMP_ALIGN_CACHE kmp_team_p
2491  *th_serial_team; /*serialized team held in reserve*/
2492 
2493 #if OMPT_SUPPORT
2494  ompt_thread_info_t ompt_thread_info;
2495 #endif
2496 
2497  /* The following are also read by the master during reinit */
2498  struct common_table *th_pri_common;
2499 
2500  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2501  /* while awaiting queuing lock acquire */
2502 
2503  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2504 
2505  ident_t *th_ident;
2506  unsigned th_x; // Random number generator data
2507  unsigned th_a; // Random number generator data
2508 
2509  /* Tasking-related data for the thread */
2510  kmp_task_team_t *th_task_team; // Task team struct
2511  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2512  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2513  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2514  // at nested levels
2515  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2516  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2517  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2518  // tasking, thus safe to reap
2519 
2520  /* More stuff for keeping track of active/sleeping threads (this part is
2521  written by the worker thread) */
2522  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2523  int th_active; // ! sleeping; 32 bits for TCR/TCW
2524  struct cons_header *th_cons; // used for consistency check
2525 #if KMP_USE_HIER_SCHED
2526  // used for hierarchical scheduling
2527  kmp_hier_private_bdata_t *th_hier_bar_data;
2528 #endif
2529 
2530  /* Add the syncronizing data which is cache aligned and padded. */
2531  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2532 
2533  KMP_ALIGN_CACHE volatile kmp_int32
2534  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2535 
2536 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2537 #define NUM_LISTS 4
2538  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2539 // allocation routines
2540 #endif
2541 
2542 #if KMP_OS_WINDOWS
2543  kmp_win32_cond_t th_suspend_cv;
2544  kmp_win32_mutex_t th_suspend_mx;
2545  int th_suspend_init;
2546 #endif
2547 #if KMP_OS_UNIX
2548  kmp_cond_align_t th_suspend_cv;
2549  kmp_mutex_align_t th_suspend_mx;
2550  int th_suspend_init_count;
2551 #endif
2552 
2553 #if USE_ITT_BUILD
2554  kmp_itt_mark_t th_itt_mark_single;
2555 // alignment ???
2556 #endif /* USE_ITT_BUILD */
2557 #if KMP_STATS_ENABLED
2558  kmp_stats_list *th_stats;
2559 #endif
2560 #if KMP_OS_UNIX
2561  std::atomic<bool> th_blocking;
2562 #endif
2563 } kmp_base_info_t;
2564 
2565 typedef union KMP_ALIGN_CACHE kmp_info {
2566  double th_align; /* use worst case alignment */
2567  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2568  kmp_base_info_t th;
2569 } kmp_info_t;
2570 
2571 // OpenMP thread team data structures
2572 
2573 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2574 
2575 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2576  double dt_align; /* use worst case alignment */
2577  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2578  kmp_base_data_t dt;
2579 } kmp_sleep_team_t;
2580 
2581 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2582  double dt_align; /* use worst case alignment */
2583  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2584  kmp_base_data_t dt;
2585 } kmp_ordered_team_t;
2586 
2587 typedef int (*launch_t)(int gtid);
2588 
2589 /* Minimum number of ARGV entries to malloc if necessary */
2590 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2591 
2592 // Set up how many argv pointers will fit in cache lines containing
2593 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2594 // larger value for more space between the master write/worker read section and
2595 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2596 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2597 #define KMP_INLINE_ARGV_BYTES \
2598  (4 * CACHE_LINE - \
2599  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2600  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2601  CACHE_LINE))
2602 #else
2603 #define KMP_INLINE_ARGV_BYTES \
2604  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2605 #endif
2606 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2607 
2608 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2609  // Synchronization Data
2610  // ---------------------------------------------------------------------------
2611  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2612  kmp_balign_team_t t_bar[bs_last_barrier];
2613  std::atomic<int> t_construct; // count of single directive encountered by team
2614  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2615 
2616  // Master only
2617  // ---------------------------------------------------------------------------
2618  KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2619  int t_master_this_cons; // "this_construct" single counter of master in parent
2620  // team
2621  ident_t *t_ident; // if volatile, have to change too much other crud to
2622  // volatile too
2623  kmp_team_p *t_parent; // parent team
2624  kmp_team_p *t_next_pool; // next free team in the team pool
2625  kmp_disp_t *t_dispatch; // thread's dispatch data
2626  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2627 #if OMP_40_ENABLED
2628  kmp_proc_bind_t t_proc_bind; // bind type for par region
2629 #endif // OMP_40_ENABLED
2630 #if USE_ITT_BUILD
2631  kmp_uint64 t_region_time; // region begin timestamp
2632 #endif /* USE_ITT_BUILD */
2633 
2634  // Master write, workers read
2635  // --------------------------------------------------------------------------
2636  KMP_ALIGN_CACHE void **t_argv;
2637  int t_argc;
2638  int t_nproc; // number of threads in team
2639  microtask_t t_pkfn;
2640  launch_t t_invoke; // procedure to launch the microtask
2641 
2642 #if OMPT_SUPPORT
2643  ompt_team_info_t ompt_team_info;
2644  ompt_lw_taskteam_t *ompt_serialized_team_info;
2645 #endif
2646 
2647 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2648  kmp_int8 t_fp_control_saved;
2649  kmp_int8 t_pad2b;
2650  kmp_int16 t_x87_fpu_control_word; // FP control regs
2651  kmp_uint32 t_mxcsr;
2652 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2653 
2654  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2655 
2656  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2657  kmp_taskdata_t
2658  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2659  int t_level; // nested parallel level
2660 
2661  KMP_ALIGN_CACHE int t_max_argc;
2662  int t_max_nproc; // max threads this team can handle (dynamicly expandable)
2663  int t_serialized; // levels deep of serialized teams
2664  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2665  int t_id; // team's id, assigned by debugger.
2666  int t_active_level; // nested active parallel level
2667  kmp_r_sched_t t_sched; // run-time schedule for the team
2668 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2669  int t_first_place; // first & last place in parent thread's partition.
2670  int t_last_place; // Restore these values to master after par region.
2671 #endif // OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2672  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2673 // omp_set_num_threads() call
2674 
2675 // Read/write by workers as well
2676 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2677  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2678  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2679  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2680  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2681  char dummy_padding[1024];
2682 #endif
2683  // Internal control stack for additional nested teams.
2684  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2685 // for SERIALIZED teams nested 2 or more levels deep
2686 #if OMP_40_ENABLED
2687  // typed flag to store request state of cancellation
2688  std::atomic<kmp_int32> t_cancel_request;
2689 #endif
2690  int t_master_active; // save on fork, restore on join
2691  kmp_taskq_t t_taskq; // this team's task queue
2692  void *t_copypriv_data; // team specific pointer to copyprivate data array
2693 #if KMP_OS_WINDOWS
2694  std::atomic<kmp_uint32> t_copyin_counter;
2695 #endif
2696 #if USE_ITT_BUILD
2697  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2698 #endif /* USE_ITT_BUILD */
2699 } kmp_base_team_t;
2700 
2701 union KMP_ALIGN_CACHE kmp_team {
2702  kmp_base_team_t t;
2703  double t_align; /* use worst case alignment */
2704  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2705 };
2706 
2707 typedef union KMP_ALIGN_CACHE kmp_time_global {
2708  double dt_align; /* use worst case alignment */
2709  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2710  kmp_base_data_t dt;
2711 } kmp_time_global_t;
2712 
2713 typedef struct kmp_base_global {
2714  /* cache-aligned */
2715  kmp_time_global_t g_time;
2716 
2717  /* non cache-aligned */
2718  volatile int g_abort;
2719  volatile int g_done;
2720 
2721  int g_dynamic;
2722  enum dynamic_mode g_dynamic_mode;
2723 } kmp_base_global_t;
2724 
2725 typedef union KMP_ALIGN_CACHE kmp_global {
2726  kmp_base_global_t g;
2727  double g_align; /* use worst case alignment */
2728  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2729 } kmp_global_t;
2730 
2731 typedef struct kmp_base_root {
2732  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2733  // (r_in_parallel>= 0)
2734  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2735  // the synch overhead or keeping r_active
2736  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2737  // GEH: This is misnamed, should be r_in_parallel
2738  volatile int r_nested; // TODO: GEH - This is unused, just remove it entirely.
2739  // keeps a count of active parallel regions per root
2740  std::atomic<int> r_in_parallel;
2741  // GEH: This is misnamed, should be r_active_levels
2742  kmp_team_t *r_root_team;
2743  kmp_team_t *r_hot_team;
2744  kmp_info_t *r_uber_thread;
2745  kmp_lock_t r_begin_lock;
2746  volatile int r_begin;
2747  int r_blocktime; /* blocktime for this root and descendants */
2748  int r_cg_nthreads; // count of active threads in a contention group
2749 } kmp_base_root_t;
2750 
2751 typedef union KMP_ALIGN_CACHE kmp_root {
2752  kmp_base_root_t r;
2753  double r_align; /* use worst case alignment */
2754  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2755 } kmp_root_t;
2756 
2757 struct fortran_inx_info {
2758  kmp_int32 data;
2759 };
2760 
2761 /* ------------------------------------------------------------------------ */
2762 
2763 extern int __kmp_settings;
2764 extern int __kmp_duplicate_library_ok;
2765 #if USE_ITT_BUILD
2766 extern int __kmp_forkjoin_frames;
2767 extern int __kmp_forkjoin_frames_mode;
2768 #endif
2769 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2770 extern int __kmp_determ_red;
2771 
2772 #ifdef KMP_DEBUG
2773 extern int kmp_a_debug;
2774 extern int kmp_b_debug;
2775 extern int kmp_c_debug;
2776 extern int kmp_d_debug;
2777 extern int kmp_e_debug;
2778 extern int kmp_f_debug;
2779 #endif /* KMP_DEBUG */
2780 
2781 /* For debug information logging using rotating buffer */
2782 #define KMP_DEBUG_BUF_LINES_INIT 512
2783 #define KMP_DEBUG_BUF_LINES_MIN 1
2784 
2785 #define KMP_DEBUG_BUF_CHARS_INIT 128
2786 #define KMP_DEBUG_BUF_CHARS_MIN 2
2787 
2788 extern int
2789  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2790 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2791 extern int
2792  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2793 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2794  entry pointer */
2795 
2796 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2797 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2798  printed in buffer so far */
2799 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2800  recommended in warnings */
2801 /* end rotating debug buffer */
2802 
2803 #ifdef KMP_DEBUG
2804 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2805 
2806 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2807 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2808 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2809 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2810 extern int __kmp_par_range_lb;
2811 extern int __kmp_par_range_ub;
2812 #endif
2813 
2814 /* For printing out dynamic storage map for threads and teams */
2815 extern int
2816  __kmp_storage_map; /* True means print storage map for threads and teams */
2817 extern int __kmp_storage_map_verbose; /* True means storage map includes
2818  placement info */
2819 extern int __kmp_storage_map_verbose_specified;
2820 
2821 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2822 extern kmp_cpuinfo_t __kmp_cpuinfo;
2823 #endif
2824 
2825 extern volatile int __kmp_init_serial;
2826 extern volatile int __kmp_init_gtid;
2827 extern volatile int __kmp_init_common;
2828 extern volatile int __kmp_init_middle;
2829 extern volatile int __kmp_init_parallel;
2830 #if KMP_USE_MONITOR
2831 extern volatile int __kmp_init_monitor;
2832 #endif
2833 extern volatile int __kmp_init_user_locks;
2834 extern int __kmp_init_counter;
2835 extern int __kmp_root_counter;
2836 extern int __kmp_version;
2837 
2838 /* list of address of allocated caches for commons */
2839 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2840 
2841 /* Barrier algorithm types and options */
2842 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2843 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2844 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2845 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2846 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2847 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2848 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2849 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2850 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2851 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2852 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2853 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2854 
2855 /* Global Locks */
2856 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2857 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2858 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2859 extern kmp_bootstrap_lock_t
2860  __kmp_exit_lock; /* exit() is not always thread-safe */
2861 #if KMP_USE_MONITOR
2862 extern kmp_bootstrap_lock_t
2863  __kmp_monitor_lock; /* control monitor thread creation */
2864 #endif
2865 extern kmp_bootstrap_lock_t
2866  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2867  __kmp_threads expansion to co-exist */
2868 
2869 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
2870 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
2871 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2872 
2873 /* used for yielding spin-waits */
2874 extern unsigned int __kmp_init_wait; /* initial number of spin-tests */
2875 extern unsigned int __kmp_next_wait; /* susequent number of spin-tests */
2876 
2877 extern enum library_type __kmp_library;
2878 
2879 extern enum sched_type __kmp_sched; /* default runtime scheduling */
2880 extern enum sched_type __kmp_static; /* default static scheduling method */
2881 extern enum sched_type __kmp_guided; /* default guided scheduling method */
2882 extern enum sched_type __kmp_auto; /* default auto scheduling method */
2883 extern int __kmp_chunk; /* default runtime chunk size */
2884 
2885 extern size_t __kmp_stksize; /* stack size per thread */
2886 #if KMP_USE_MONITOR
2887 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
2888 #endif
2889 extern size_t __kmp_stkoffset; /* stack offset per thread */
2890 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
2891 
2892 extern size_t
2893  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
2894 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
2895 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
2896 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
2897 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
2898 extern int __kmp_generate_warnings; /* should we issue warnings? */
2899 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
2900 
2901 #ifdef DEBUG_SUSPEND
2902 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
2903 #endif
2904 
2905 extern kmp_uint32 __kmp_yield_init;
2906 extern kmp_uint32 __kmp_yield_next;
2907 
2908 #if KMP_USE_MONITOR
2909 extern kmp_uint32 __kmp_yielding_on;
2910 #endif
2911 extern kmp_uint32 __kmp_yield_cycle;
2912 extern kmp_int32 __kmp_yield_on_count;
2913 extern kmp_int32 __kmp_yield_off_count;
2914 
2915 /* ------------------------------------------------------------------------- */
2916 extern int __kmp_allThreadsSpecified;
2917 
2918 extern size_t __kmp_align_alloc;
2919 /* following data protected by initialization routines */
2920 extern int __kmp_xproc; /* number of processors in the system */
2921 extern int __kmp_avail_proc; /* number of processors available to the process */
2922 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
2923 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
2924 // maximum total number of concurrently-existing threads on device
2925 extern int __kmp_max_nth;
2926 // maximum total number of concurrently-existing threads in a contention group
2927 extern int __kmp_cg_max_nth;
2928 extern int __kmp_teams_max_nth; // max threads used in a teams construct
2929 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
2930  __kmp_root */
2931 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
2932  region a la OMP_NUM_THREADS */
2933 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
2934  initialization */
2935 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
2936  used (fixed) */
2937 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
2938  (__kmpc_threadprivate_cached()) */
2939 extern int __kmp_dflt_nested; /* nested parallelism enabled by default a la
2940  OMP_NESTED */
2941 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
2942  blocking (env setting) */
2943 #if KMP_USE_MONITOR
2944 extern int
2945  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
2946 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
2947  blocking */
2948 #endif
2949 #ifdef KMP_ADJUST_BLOCKTIME
2950 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
2951 #endif /* KMP_ADJUST_BLOCKTIME */
2952 #ifdef KMP_DFLT_NTH_CORES
2953 extern int __kmp_ncores; /* Total number of cores for threads placement */
2954 #endif
2955 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
2956 extern int __kmp_abort_delay;
2957 
2958 extern int __kmp_need_register_atfork_specified;
2959 extern int
2960  __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
2961  install fork handler */
2962 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
2963  0 - not set, will be set at runtime
2964  1 - using stack search
2965  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
2966  X*) or TlsGetValue(Windows* OS))
2967  3 - static TLS (__declspec(thread) __kmp_gtid),
2968  Linux* OS .so only. */
2969 extern int
2970  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
2971 #ifdef KMP_TDATA_GTID
2972 extern KMP_THREAD_LOCAL int __kmp_gtid;
2973 #endif
2974 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
2975 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
2976 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2977 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
2978 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
2979 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
2980 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2981 
2982 extern int __kmp_dflt_max_active_levels; /* max_active_levels for nested
2983  parallelism enabled by default via
2984  OMP_MAX_ACTIVE_LEVELS */
2985 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
2986  concurrent execution per team */
2987 #if KMP_NESTED_HOT_TEAMS
2988 extern int __kmp_hot_teams_mode;
2989 extern int __kmp_hot_teams_max_level;
2990 #endif
2991 
2992 #if KMP_OS_LINUX
2993 extern enum clock_function_type __kmp_clock_function;
2994 extern int __kmp_clock_function_param;
2995 #endif /* KMP_OS_LINUX */
2996 
2997 #if KMP_MIC_SUPPORTED
2998 extern enum mic_type __kmp_mic_type;
2999 #endif
3000 
3001 #ifdef USE_LOAD_BALANCE
3002 extern double __kmp_load_balance_interval; // load balance algorithm interval
3003 #endif /* USE_LOAD_BALANCE */
3004 
3005 // OpenMP 3.1 - Nested num threads array
3006 typedef struct kmp_nested_nthreads_t {
3007  int *nth;
3008  int size;
3009  int used;
3010 } kmp_nested_nthreads_t;
3011 
3012 extern kmp_nested_nthreads_t __kmp_nested_nth;
3013 
3014 #if KMP_USE_ADAPTIVE_LOCKS
3015 
3016 // Parameters for the speculative lock backoff system.
3017 struct kmp_adaptive_backoff_params_t {
3018  // Number of soft retries before it counts as a hard retry.
3019  kmp_uint32 max_soft_retries;
3020  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3021  // the right
3022  kmp_uint32 max_badness;
3023 };
3024 
3025 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3026 
3027 #if KMP_DEBUG_ADAPTIVE_LOCKS
3028 extern const char *__kmp_speculative_statsfile;
3029 #endif
3030 
3031 #endif // KMP_USE_ADAPTIVE_LOCKS
3032 
3033 #if OMP_40_ENABLED
3034 extern int __kmp_display_env; /* TRUE or FALSE */
3035 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3036 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3037 #endif
3038 
3039 /* ------------------------------------------------------------------------- */
3040 
3041 /* the following are protected by the fork/join lock */
3042 /* write: lock read: anytime */
3043 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3044 /* read/write: lock */
3045 extern volatile kmp_team_t *__kmp_team_pool;
3046 extern volatile kmp_info_t *__kmp_thread_pool;
3047 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3048 
3049 // total num threads reachable from some root thread including all root threads
3050 extern volatile int __kmp_nth;
3051 /* total number of threads reachable from some root thread including all root
3052  threads, and those in the thread pool */
3053 extern volatile int __kmp_all_nth;
3054 extern int __kmp_thread_pool_nth;
3055 extern std::atomic<int> __kmp_thread_pool_active_nth;
3056 
3057 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3058 /* end data protected by fork/join lock */
3059 /* ------------------------------------------------------------------------- */
3060 
3061 extern kmp_global_t __kmp_global; /* global status */
3062 
3063 extern kmp_info_t __kmp_monitor;
3064 // For Debugging Support Library
3065 extern std::atomic<kmp_uint32> __kmp_team_counter;
3066 // For Debugging Support Library
3067 extern std::atomic<kmp_uint32> __kmp_task_counter;
3068 
3069 #if USE_DEBUGGER
3070 #define _KMP_GEN_ID(counter) \
3071  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3072 #else
3073 #define _KMP_GEN_ID(counter) (~0)
3074 #endif /* USE_DEBUGGER */
3075 
3076 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3077 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3078 
3079 /* ------------------------------------------------------------------------ */
3080 
3081 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3082  size_t size, char const *format, ...);
3083 
3084 extern void __kmp_serial_initialize(void);
3085 extern void __kmp_middle_initialize(void);
3086 extern void __kmp_parallel_initialize(void);
3087 
3088 extern void __kmp_internal_begin(void);
3089 extern void __kmp_internal_end_library(int gtid);
3090 extern void __kmp_internal_end_thread(int gtid);
3091 extern void __kmp_internal_end_atexit(void);
3092 extern void __kmp_internal_end_fini(void);
3093 extern void __kmp_internal_end_dtor(void);
3094 extern void __kmp_internal_end_dest(void *);
3095 
3096 extern int __kmp_register_root(int initial_thread);
3097 extern void __kmp_unregister_root(int gtid);
3098 
3099 extern int __kmp_ignore_mppbeg(void);
3100 extern int __kmp_ignore_mppend(void);
3101 
3102 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3103 extern void __kmp_exit_single(int gtid);
3104 
3105 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3106 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3107 
3108 #ifdef USE_LOAD_BALANCE
3109 extern int __kmp_get_load_balance(int);
3110 #endif
3111 
3112 extern int __kmp_get_global_thread_id(void);
3113 extern int __kmp_get_global_thread_id_reg(void);
3114 extern void __kmp_exit_thread(int exit_status);
3115 extern void __kmp_abort(char const *format, ...);
3116 extern void __kmp_abort_thread(void);
3117 KMP_NORETURN extern void __kmp_abort_process(void);
3118 extern void __kmp_warn(char const *format, ...);
3119 
3120 extern void __kmp_set_num_threads(int new_nth, int gtid);
3121 
3122 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3123 // registered.
3124 static inline kmp_info_t *__kmp_entry_thread() {
3125  int gtid = __kmp_entry_gtid();
3126 
3127  return __kmp_threads[gtid];
3128 }
3129 
3130 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3131 extern int __kmp_get_max_active_levels(int gtid);
3132 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3133 extern int __kmp_get_team_size(int gtid, int level);
3134 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3135 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3136 
3137 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3138 extern void __kmp_init_random(kmp_info_t *thread);
3139 
3140 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3141 extern void __kmp_adjust_num_threads(int new_nproc);
3142 
3143 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3144 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3145 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3146 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3147 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3148 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3149 
3150 #if USE_FAST_MEMORY
3151 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3152  size_t size KMP_SRC_LOC_DECL);
3153 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3154 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3155 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3156 #define __kmp_fast_allocate(this_thr, size) \
3157  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3158 #define __kmp_fast_free(this_thr, ptr) \
3159  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3160 #endif
3161 
3162 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3163 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3164  size_t elsize KMP_SRC_LOC_DECL);
3165 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3166  size_t size KMP_SRC_LOC_DECL);
3167 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3168 #define __kmp_thread_malloc(th, size) \
3169  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3170 #define __kmp_thread_calloc(th, nelem, elsize) \
3171  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3172 #define __kmp_thread_realloc(th, ptr, size) \
3173  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3174 #define __kmp_thread_free(th, ptr) \
3175  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3176 
3177 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3178 #define KMP_INTERNAL_FREE(p) free(p)
3179 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3180 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3181 
3182 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3183 
3184 #if OMP_40_ENABLED
3185 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3186  kmp_proc_bind_t proc_bind);
3187 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3188  int num_threads);
3189 #endif
3190 
3191 extern void __kmp_yield(int cond);
3192 
3193 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3194  enum sched_type schedule, kmp_int32 lb,
3195  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3196 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3197  enum sched_type schedule, kmp_uint32 lb,
3198  kmp_uint32 ub, kmp_int32 st,
3199  kmp_int32 chunk);
3200 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3201  enum sched_type schedule, kmp_int64 lb,
3202  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3203 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3204  enum sched_type schedule, kmp_uint64 lb,
3205  kmp_uint64 ub, kmp_int64 st,
3206  kmp_int64 chunk);
3207 
3208 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3209  kmp_int32 *p_last, kmp_int32 *p_lb,
3210  kmp_int32 *p_ub, kmp_int32 *p_st);
3211 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3212  kmp_int32 *p_last, kmp_uint32 *p_lb,
3213  kmp_uint32 *p_ub, kmp_int32 *p_st);
3214 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3215  kmp_int32 *p_last, kmp_int64 *p_lb,
3216  kmp_int64 *p_ub, kmp_int64 *p_st);
3217 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3218  kmp_int32 *p_last, kmp_uint64 *p_lb,
3219  kmp_uint64 *p_ub, kmp_int64 *p_st);
3220 
3221 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3222 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3223 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3224 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3225 
3226 #ifdef KMP_GOMP_COMPAT
3227 
3228 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3229  enum sched_type schedule, kmp_int32 lb,
3230  kmp_int32 ub, kmp_int32 st,
3231  kmp_int32 chunk, int push_ws);
3232 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3233  enum sched_type schedule, kmp_uint32 lb,
3234  kmp_uint32 ub, kmp_int32 st,
3235  kmp_int32 chunk, int push_ws);
3236 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3237  enum sched_type schedule, kmp_int64 lb,
3238  kmp_int64 ub, kmp_int64 st,
3239  kmp_int64 chunk, int push_ws);
3240 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3241  enum sched_type schedule, kmp_uint64 lb,
3242  kmp_uint64 ub, kmp_int64 st,
3243  kmp_int64 chunk, int push_ws);
3244 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3245 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3246 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3247 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3248 
3249 #endif /* KMP_GOMP_COMPAT */
3250 
3251 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3252 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3253 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3254 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3255 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3256 extern kmp_uint32 __kmp_wait_yield_4(kmp_uint32 volatile *spinner,
3257  kmp_uint32 checker,
3258  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3259  void *obj);
3260 extern void __kmp_wait_yield_4_ptr(void *spinner, kmp_uint32 checker,
3261  kmp_uint32 (*pred)(void *, kmp_uint32),
3262  void *obj);
3263 
3264 class kmp_flag_32;
3265 class kmp_flag_64;
3266 class kmp_flag_oncore;
3267 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3268  int final_spin
3269 #if USE_ITT_BUILD
3270  ,
3271  void *itt_sync_obj
3272 #endif
3273  );
3274 extern void __kmp_release_64(kmp_flag_64 *flag);
3275 
3276 extern void __kmp_infinite_loop(void);
3277 
3278 extern void __kmp_cleanup(void);
3279 
3280 #if KMP_HANDLE_SIGNALS
3281 extern int __kmp_handle_signals;
3282 extern void __kmp_install_signals(int parallel_init);
3283 extern void __kmp_remove_signals(void);
3284 #endif
3285 
3286 extern void __kmp_clear_system_time(void);
3287 extern void __kmp_read_system_time(double *delta);
3288 
3289 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3290 
3291 extern void __kmp_expand_host_name(char *buffer, size_t size);
3292 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3293 
3294 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3295 extern void
3296 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3297 #endif
3298 
3299 extern void
3300 __kmp_runtime_initialize(void); /* machine specific initialization */
3301 extern void __kmp_runtime_destroy(void);
3302 
3303 #if KMP_AFFINITY_SUPPORTED
3304 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3305  kmp_affin_mask_t *mask);
3306 extern void __kmp_affinity_initialize(void);
3307 extern void __kmp_affinity_uninitialize(void);
3308 extern void __kmp_affinity_set_init_mask(
3309  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3310 #if OMP_40_ENABLED
3311 extern void __kmp_affinity_set_place(int gtid);
3312 #endif
3313 extern void __kmp_affinity_determine_capable(const char *env_var);
3314 extern int __kmp_aux_set_affinity(void **mask);
3315 extern int __kmp_aux_get_affinity(void **mask);
3316 extern int __kmp_aux_get_affinity_max_proc();
3317 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3318 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3319 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3320 extern void __kmp_balanced_affinity(int tid, int team_size);
3321 #if KMP_OS_LINUX
3322 extern int kmp_set_thread_affinity_mask_initial(void);
3323 #endif
3324 #endif /* KMP_AFFINITY_SUPPORTED */
3325 
3326 extern void __kmp_cleanup_hierarchy();
3327 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3328 
3329 #if KMP_USE_FUTEX
3330 
3331 extern int __kmp_futex_determine_capable(void);
3332 
3333 #endif // KMP_USE_FUTEX
3334 
3335 extern void __kmp_gtid_set_specific(int gtid);
3336 extern int __kmp_gtid_get_specific(void);
3337 
3338 extern double __kmp_read_cpu_time(void);
3339 
3340 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3341 
3342 #if KMP_USE_MONITOR
3343 extern void __kmp_create_monitor(kmp_info_t *th);
3344 #endif
3345 
3346 extern void *__kmp_launch_thread(kmp_info_t *thr);
3347 
3348 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3349 
3350 #if KMP_OS_WINDOWS
3351 extern int __kmp_still_running(kmp_info_t *th);
3352 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3353 extern void __kmp_free_handle(kmp_thread_t tHandle);
3354 #endif
3355 
3356 #if KMP_USE_MONITOR
3357 extern void __kmp_reap_monitor(kmp_info_t *th);
3358 #endif
3359 extern void __kmp_reap_worker(kmp_info_t *th);
3360 extern void __kmp_terminate_thread(int gtid);
3361 
3362 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag);
3363 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag);
3364 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
3365 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag);
3366 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag);
3367 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
3368 
3369 extern void __kmp_elapsed(double *);
3370 extern void __kmp_elapsed_tick(double *);
3371 
3372 extern void __kmp_enable(int old_state);
3373 extern void __kmp_disable(int *old_state);
3374 
3375 extern void __kmp_thread_sleep(int millis);
3376 
3377 extern void __kmp_common_initialize(void);
3378 extern void __kmp_common_destroy(void);
3379 extern void __kmp_common_destroy_gtid(int gtid);
3380 
3381 #if KMP_OS_UNIX
3382 extern void __kmp_register_atfork(void);
3383 #endif
3384 extern void __kmp_suspend_initialize(void);
3385 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3386 
3387 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3388  int tid);
3389 #if OMP_40_ENABLED
3390 extern kmp_team_t *
3391 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3392 #if OMPT_SUPPORT
3393  ompt_data_t ompt_parallel_data,
3394 #endif
3395  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3396  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3397 #else
3398 extern kmp_team_t *
3399 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3400 #if OMPT_SUPPORT
3401  ompt_id_t ompt_parallel_id,
3402 #endif
3403  kmp_internal_control_t *new_icvs,
3404  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3405 #endif // OMP_40_ENABLED
3406 extern void __kmp_free_thread(kmp_info_t *);
3407 extern void __kmp_free_team(kmp_root_t *,
3408  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3409 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3410 
3411 /* ------------------------------------------------------------------------ */
3412 
3413 extern void __kmp_initialize_bget(kmp_info_t *th);
3414 extern void __kmp_finalize_bget(kmp_info_t *th);
3415 
3416 KMP_EXPORT void *kmpc_malloc(size_t size);
3417 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3418 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3419 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3420 KMP_EXPORT void kmpc_free(void *ptr);
3421 
3422 /* declarations for internal use */
3423 
3424 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3425  size_t reduce_size, void *reduce_data,
3426  void (*reduce)(void *, void *));
3427 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3428 
3433 enum fork_context_e {
3434  fork_context_gnu,
3436  fork_context_intel,
3437  fork_context_last
3438 };
3439 extern int __kmp_fork_call(ident_t *loc, int gtid,
3440  enum fork_context_e fork_context, kmp_int32 argc,
3441  microtask_t microtask, launch_t invoker,
3442 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
3443 #if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX
3444  va_list *ap
3445 #else
3446  va_list ap
3447 #endif
3448  );
3449 
3450 extern void __kmp_join_call(ident_t *loc, int gtid
3451 #if OMPT_SUPPORT
3452  ,
3453  enum fork_context_e fork_context
3454 #endif
3455 #if OMP_40_ENABLED
3456  ,
3457  int exit_teams = 0
3458 #endif
3459  );
3460 
3461 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3462 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3463 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3464 extern int __kmp_invoke_task_func(int gtid);
3465 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3466  kmp_info_t *this_thr,
3467  kmp_team_t *team);
3468 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3469  kmp_info_t *this_thr,
3470  kmp_team_t *team);
3471 
3472 // should never have been exported
3473 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3474 #if OMP_40_ENABLED
3475 extern int __kmp_invoke_teams_master(int gtid);
3476 extern void __kmp_teams_master(int gtid);
3477 #endif
3478 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3479 extern void __kmp_user_set_library(enum library_type arg);
3480 extern void __kmp_aux_set_library(enum library_type arg);
3481 extern void __kmp_aux_set_stacksize(size_t arg);
3482 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3483 extern void __kmp_aux_set_defaults(char const *str, int len);
3484 
3485 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3486 void kmpc_set_blocktime(int arg);
3487 void ompc_set_nested(int flag);
3488 void ompc_set_dynamic(int flag);
3489 void ompc_set_num_threads(int arg);
3490 
3491 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3492  kmp_team_t *team, int tid);
3493 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3494 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3495  kmp_tasking_flags_t *flags,
3496  size_t sizeof_kmp_task_t,
3497  size_t sizeof_shareds,
3498  kmp_routine_entry_t task_entry);
3499 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3500  kmp_team_t *team, int tid,
3501  int set_curr_task);
3502 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3503 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3504 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3505  kmp_flag_32 *flag, int final_spin,
3506  int *thread_finished,
3507 #if USE_ITT_BUILD
3508  void *itt_sync_obj,
3509 #endif /* USE_ITT_BUILD */
3510  kmp_int32 is_constrained);
3511 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3512  kmp_flag_64 *flag, int final_spin,
3513  int *thread_finished,
3514 #if USE_ITT_BUILD
3515  void *itt_sync_obj,
3516 #endif /* USE_ITT_BUILD */
3517  kmp_int32 is_constrained);
3518 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3519  kmp_flag_oncore *flag, int final_spin,
3520  int *thread_finished,
3521 #if USE_ITT_BUILD
3522  void *itt_sync_obj,
3523 #endif /* USE_ITT_BUILD */
3524  kmp_int32 is_constrained);
3525 
3526 extern void __kmp_free_task_team(kmp_info_t *thread,
3527  kmp_task_team_t *task_team);
3528 extern void __kmp_reap_task_teams(void);
3529 extern void __kmp_wait_to_unref_task_teams(void);
3530 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3531  int always);
3532 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3533 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3534 #if USE_ITT_BUILD
3535  ,
3536  void *itt_sync_obj
3537 #endif /* USE_ITT_BUILD */
3538  ,
3539  int wait = 1);
3540 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3541  int gtid);
3542 
3543 extern int __kmp_is_address_mapped(void *addr);
3544 extern kmp_uint64 __kmp_hardware_timestamp(void);
3545 
3546 #if KMP_OS_UNIX
3547 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3548 #endif
3549 
3550 /* ------------------------------------------------------------------------ */
3551 //
3552 // Assembly routines that have no compiler intrinsic replacement
3553 //
3554 
3555 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3556 
3557 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
3558 
3559 #define __kmp_load_mxcsr(p) _mm_setcsr(*(p))
3560 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
3561 
3562 extern void __kmp_load_x87_fpu_control_word(kmp_int16 *p);
3563 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
3564 extern void __kmp_clear_x87_fpu_status_word();
3565 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
3566 
3567 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3568 
3569 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3570  void *argv[]
3571 #if OMPT_SUPPORT
3572  ,
3573  void **exit_frame_ptr
3574 #endif
3575  );
3576 
3577 /* ------------------------------------------------------------------------ */
3578 
3579 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3580 KMP_EXPORT void __kmpc_end(ident_t *);
3581 
3582 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3583  kmpc_ctor_vec ctor,
3584  kmpc_cctor_vec cctor,
3585  kmpc_dtor_vec dtor,
3586  size_t vector_length);
3587 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3588  kmpc_ctor ctor, kmpc_cctor cctor,
3589  kmpc_dtor dtor);
3590 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3591  void *data, size_t size);
3592 
3593 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3594 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3595 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3596 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3597 
3598 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3599 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3600  kmpc_micro microtask, ...);
3601 
3602 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3603 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3604 
3605 KMP_EXPORT void __kmpc_flush(ident_t *);
3606 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3607 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3608 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3609 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3610 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3611 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3612  kmp_critical_name *);
3613 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3614  kmp_critical_name *);
3615 
3616 #if OMP_45_ENABLED
3617 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3618  kmp_critical_name *, uintptr_t hint);
3619 #endif
3620 
3621 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3622 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3623 
3624 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3625  kmp_int32 global_tid);
3626 
3627 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3628 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3629 
3630 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3631  kmp_int32 schedtype, kmp_int32 *plastiter,
3632  kmp_int *plower, kmp_int *pupper,
3633  kmp_int *pstride, kmp_int incr,
3634  kmp_int chunk);
3635 
3636 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3637 
3638 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3639  size_t cpy_size, void *cpy_data,
3640  void (*cpy_func)(void *, void *),
3641  kmp_int32 didit);
3642 
3643 extern void KMPC_SET_NUM_THREADS(int arg);
3644 extern void KMPC_SET_DYNAMIC(int flag);
3645 extern void KMPC_SET_NESTED(int flag);
3646 
3647 /* Taskq interface routines */
3648 KMP_EXPORT kmpc_thunk_t *__kmpc_taskq(ident_t *loc, kmp_int32 global_tid,
3649  kmpc_task_t taskq_task,
3650  size_t sizeof_thunk,
3651  size_t sizeof_shareds, kmp_int32 flags,
3652  kmpc_shared_vars_t **shareds);
3653 KMP_EXPORT void __kmpc_end_taskq(ident_t *loc, kmp_int32 global_tid,
3654  kmpc_thunk_t *thunk);
3655 KMP_EXPORT kmp_int32 __kmpc_task(ident_t *loc, kmp_int32 global_tid,
3656  kmpc_thunk_t *thunk);
3657 KMP_EXPORT void __kmpc_taskq_task(ident_t *loc, kmp_int32 global_tid,
3658  kmpc_thunk_t *thunk, kmp_int32 status);
3659 KMP_EXPORT void __kmpc_end_taskq_task(ident_t *loc, kmp_int32 global_tid,
3660  kmpc_thunk_t *thunk);
3661 KMP_EXPORT kmpc_thunk_t *__kmpc_task_buffer(ident_t *loc, kmp_int32 global_tid,
3662  kmpc_thunk_t *taskq_thunk,
3663  kmpc_task_t task);
3664 
3665 /* OMP 3.0 tasking interface routines */
3666 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3667  kmp_task_t *new_task);
3668 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3669  kmp_int32 flags,
3670  size_t sizeof_kmp_task_t,
3671  size_t sizeof_shareds,
3672  kmp_routine_entry_t task_entry);
3673 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3674  kmp_task_t *task);
3675 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3676  kmp_task_t *task);
3677 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3678  kmp_task_t *new_task);
3679 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3680 
3681 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3682  int end_part);
3683 
3684 #if TASK_UNUSED
3685 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3686 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3687  kmp_task_t *task);
3688 #endif // TASK_UNUSED
3689 
3690 /* ------------------------------------------------------------------------ */
3691 
3692 #if OMP_40_ENABLED
3693 
3694 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3695 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3696 
3697 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3698  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3699  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3700  kmp_depend_info_t *noalias_dep_list);
3701 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3702  kmp_int32 ndeps,
3703  kmp_depend_info_t *dep_list,
3704  kmp_int32 ndeps_noalias,
3705  kmp_depend_info_t *noalias_dep_list);
3706 extern void __kmp_release_deps(kmp_int32 gtid, kmp_taskdata_t *task);
3707 extern void __kmp_dephash_free_entries(kmp_info_t *thread, kmp_dephash_t *h);
3708 extern void __kmp_dephash_free(kmp_info_t *thread, kmp_dephash_t *h);
3709 
3710 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3711  bool serialize_immediate);
3712 
3713 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3714  kmp_int32 cncl_kind);
3715 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3716  kmp_int32 cncl_kind);
3717 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3718 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3719 
3720 #if OMP_45_ENABLED
3721 
3722 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3723 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3724 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3725  kmp_int32 if_val, kmp_uint64 *lb,
3726  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3727  kmp_int32 sched, kmp_uint64 grainsize,
3728  void *task_dup);
3729 #endif
3730 // TODO: change to OMP_50_ENABLED, need to change build tools for this to work
3731 #if OMP_45_ENABLED
3732 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3733 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3734 #endif
3735 
3736 #endif
3737 
3738 /* Lock interface routines (fast versions with gtid passed in) */
3739 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3740  void **user_lock);
3741 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3742  void **user_lock);
3743 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3744  void **user_lock);
3745 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3746  void **user_lock);
3747 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3748 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3749  void **user_lock);
3750 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3751  void **user_lock);
3752 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3753  void **user_lock);
3754 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3755 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3756  void **user_lock);
3757 
3758 #if OMP_45_ENABLED
3759 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3760  void **user_lock, uintptr_t hint);
3761 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3762  void **user_lock,
3763  uintptr_t hint);
3764 #endif
3765 
3766 /* Interface to fast scalable reduce methods routines */
3767 
3768 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3769  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3770  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3771  kmp_critical_name *lck);
3772 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3773  kmp_critical_name *lck);
3774 KMP_EXPORT kmp_int32 __kmpc_reduce(
3775  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3776  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3777  kmp_critical_name *lck);
3778 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3779  kmp_critical_name *lck);
3780 
3781 /* Internal fast reduction routines */
3782 
3783 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3784  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3785  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3786  kmp_critical_name *lck);
3787 
3788 // this function is for testing set/get/determine reduce method
3789 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3790 
3791 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3792 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3793 
3794 // C++ port
3795 // missing 'extern "C"' declarations
3796 
3797 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3798 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3799 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3800  kmp_int32 num_threads);
3801 
3802 #if OMP_40_ENABLED
3803 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3804  int proc_bind);
3805 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3806  kmp_int32 num_teams,
3807  kmp_int32 num_threads);
3808 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3809  kmpc_micro microtask, ...);
3810 #endif
3811 #if OMP_45_ENABLED
3812 struct kmp_dim { // loop bounds info casted to kmp_int64
3813  kmp_int64 lo; // lower
3814  kmp_int64 up; // upper
3815  kmp_int64 st; // stride
3816 };
3817 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3818  kmp_int32 num_dims,
3819  const struct kmp_dim *dims);
3820 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3821  const kmp_int64 *vec);
3822 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3823  const kmp_int64 *vec);
3824 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3825 #endif
3826 
3827 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3828  void *data, size_t size,
3829  void ***cache);
3830 
3831 // Symbols for MS mutual detection.
3832 extern int _You_must_link_with_exactly_one_OpenMP_library;
3833 extern int _You_must_link_with_Intel_OpenMP_library;
3834 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3835 extern int _You_must_link_with_Microsoft_OpenMP_library;
3836 #endif
3837 
3838 // The routines below are not exported.
3839 // Consider making them 'static' in corresponding source files.
3840 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3841  void *data_addr, size_t pc_size);
3842 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3843  void *data_addr,
3844  size_t pc_size);
3845 void __kmp_threadprivate_resize_cache(int newCapacity);
3846 void __kmp_cleanup_threadprivate_caches();
3847 
3848 // ompc_, kmpc_ entries moved from omp.h.
3849 #if KMP_OS_WINDOWS
3850 #define KMPC_CONVENTION __cdecl
3851 #else
3852 #define KMPC_CONVENTION
3853 #endif
3854 
3855 #ifndef __OMP_H
3856 typedef enum omp_sched_t {
3857  omp_sched_static = 1,
3858  omp_sched_dynamic = 2,
3859  omp_sched_guided = 3,
3860  omp_sched_auto = 4
3861 } omp_sched_t;
3862 typedef void *kmp_affinity_mask_t;
3863 #endif
3864 
3865 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3866 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3867 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3868 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3869 KMP_EXPORT int KMPC_CONVENTION
3870 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3871 KMP_EXPORT int KMPC_CONVENTION
3872 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3873 KMP_EXPORT int KMPC_CONVENTION
3874 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3875 
3876 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3877 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3878 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3879 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
3880 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
3881 
3882 #if OMP_50_ENABLED
3883 enum kmp_target_offload_kind {
3884  tgt_disabled = 0,
3885  tgt_default = 1,
3886  tgt_mandatory = 2
3887 };
3888 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3889 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
3890 extern kmp_target_offload_kind_t __kmp_target_offload;
3891 extern int __kmpc_get_target_offload();
3892 #endif
3893 
3894 #ifdef __cplusplus
3895 }
3896 #endif
3897 
3898 #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:1424
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:1447
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:1441
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:1453
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:1431
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:1418
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:1400
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