libcilkrts/runtime/sysdep-unix.c

   1 /*
   2  * sysdep-unix.c
   3  *
   4  *************************************************************************
   5  *
   6  *  @copyright
   7  *  Copyright (C) 2010-2013, Intel Corporation
   8  *  All rights reserved.
   9  *
  10  *  @copyright
  11  *  Redistribution and use in source and binary forms, with or without
  12  *  modification, are permitted provided that the following conditions
  13  *  are met:
  14  *
  15  *    * Redistributions of source code must retain the above copyright
  16  *      notice, this list of conditions and the following disclaimer.
  17  *    * Redistributions in binary form must reproduce the above copyright
  18  *      notice, this list of conditions and the following disclaimer in
  19  *      the documentation and/or other materials provided with the
  20  *      distribution.
  21  *    * Neither the name of Intel Corporation nor the names of its
  22  *      contributors may be used to endorse or promote products derived
  23  *      from this software without specific prior written permission.
  24  *
  25  *  @copyright
  26  *  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  27  *  "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  28  *  LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  29  *  A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  30  *  HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  31  *  INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  32  *  BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
  33  *  OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  34  *  AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  35  *  LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
  36  *  WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  37  *  POSSIBILITY OF SUCH DAMAGE.
  38  *
  39  **************************************************************************
  40  */
  41
  42 #ifdef __linux__
  43     // define _GNU_SOURCE before *any* #include.
  44     // Even <stdint.h> will break later #includes if this macro is not
  45     // already defined when it is #included.
  46 #   define _GNU_SOURCE
  47 #endif
  48
  49 #include "sysdep.h"
  50 #include "os.h"
  51 #include "bug.h"
  52 #include "local_state.h"
  53 #include "signal_node.h"
  54 #include "full_frame.h"
  55 #include "jmpbuf.h"
  56 #include "cilk_malloc.h"
  57 #include "reducer_impl.h"
  58 #include "metacall_impl.h"
  59
  60
  61 // On x86 processors (but not MIC processors), the compiler generated code to
  62 // save the FP state (rounding mode and the like) before calling setjmp.  We
  63 // will need to restore that state when we resume.
  64 #ifndef __MIC__
  65 # if defined(__i386__) || defined(__x86_64)
  66 #   define RESTORE_X86_FP_STATE
  67 # endif // defined(__i386__) || defined(__x86_64)
  68 #endif  // __MIC__
  69
  70 // contains notification macros for VTune.
  71 #include "cilk-ittnotify.h"
  72
  73 #include <stddef.h>
  74
  75 #ifdef __CYGWIN__
  76 // On Cygwin, string.h doesnt declare strcasecmp if __STRICT_ANSI__ is defined
  77 #   undef __STRICT_ANSI__
  78 #endif
  79
  80 #include <string.h>
  81 #include <pthread.h>
  82 #include <unistd.h>
  83
  84 #if defined HAVE_ALLOCA_H
  85 # include <alloca.h>
  86 #elif defined __GNUC__
  87 # define alloca __builtin_alloca
  88 #elif defined _AIX
  89 # define alloca __alloca
  90 #else
  91 # include <stddef.h>
  92 # ifdef  __cplusplus
  93 extern "C"
  94 # endif
  95 void *alloca (size_t);
  96 #endif
  97
  98 #ifdef __APPLE__
  99 //#   include <scheduler.h>  // Angle brackets include Apple's scheduler.h, not ours.
 100 #endif
 101
 102 #ifdef __linux__
 103 #   include <sys/resource.h>
 104 #   include <sys/sysinfo.h>
 105 #endif
 106
 107 #ifdef __FreeBSD__
 108 #   include <sys/resource.h>
 109 // BSD does not define MAP_ANONYMOUS, but *does* define MAP_ANON. Aren't standards great!
 110 #   define MAP_ANONYMOUS MAP_ANON
 111 #endif
 112
 113 #ifdef  __VXWORKS__
 114 #   include <vxWorks.h>
 115 #   include <vxCpuLib.h>
 116 #endif
 117
 118 struct global_sysdep_state
 119 {
 120     pthread_t *threads;    ///< Array of pthreads for system workers
 121     size_t pthread_t_size; ///< for cilk_db
 122 };
 123
 124 static void internal_enforce_global_visibility();
 125
 126
 127 COMMON_SYSDEP
 128 void __cilkrts_init_worker_sysdep(struct __cilkrts_worker *w)
 129 {
 130     ITT_SYNC_CREATE(w, "Scheduler");
 131 }
 132
 133 COMMON_SYSDEP
 134 void __cilkrts_destroy_worker_sysdep(struct __cilkrts_worker *w)
 135 {
 136 }
 137
 138 COMMON_SYSDEP
 139 void __cilkrts_init_global_sysdep(global_state_t *g)
 140 {
 141     internal_enforce_global_visibility();
 142
 143     __cilkrts_init_tls_variables();
 144
 145     CILK_ASSERT(g->total_workers >= g->P - 1);
 146     g->sysdep = __cilkrts_malloc(sizeof (struct global_sysdep_state));
 147     CILK_ASSERT(g->sysdep);
 148     g->sysdep->pthread_t_size = sizeof (pthread_t);
 149
 150     // TBD: Should this value be g->total_workers, or g->P?
 151     //      Need to check what we are using this field for.
 152     g->sysdep->threads = __cilkrts_malloc(sizeof(pthread_t) * g->total_workers);
 153     CILK_ASSERT(g->sysdep->threads);
 154
 155     return;
 156 }
 157
 158 COMMON_SYSDEP
 159 void __cilkrts_destroy_global_sysdep(global_state_t *g)
 160 {
 161     if (g->sysdep->threads)
 162         __cilkrts_free(g->sysdep->threads);
 163     __cilkrts_free(g->sysdep);
 164 }
 165
 166 /*************************************************************
 167   Creation of worker threads:
 168 *************************************************************/
 169
 170 static void internal_run_scheduler_with_exceptions(__cilkrts_worker *w)
 171 {
 172     /* We assume the stack grows down. */
 173     char var;
 174     __cilkrts_cilkscreen_establish_c_stack(&var - 1000000, &var);
 175
 176     __cilkrts_run_scheduler_with_exceptions(w);
 177 }
 178
 179
 180
 181 /*
 182  * scheduler_thread_proc_for_system_worker
 183  *
 184  * Thread start function called when we start a new worker.
 185  *
 186  */
 187 NON_COMMON void* scheduler_thread_proc_for_system_worker(void *arg)
 188 {
 189     /*int status;*/
 190     __cilkrts_worker *w = (__cilkrts_worker *)arg;
 191
 192 #ifdef __INTEL_COMPILER
 193 #ifdef USE_ITTNOTIFY
 194     // Name the threads for Advisor.  They don't want a worker number.
 195     __itt_thread_set_name("Cilk Worker");
 196 #endif // defined USE_ITTNOTIFY
 197 #endif // defined __INTEL_COMPILER
 198
 199     /* Worker startup is serialized
 200     status = pthread_mutex_lock(&__cilkrts_global_mutex);
 201     CILK_ASSERT(status == 0);*/
 202     CILK_ASSERT(w->l->type == WORKER_SYSTEM);
 203     /*status = pthread_mutex_unlock(&__cilkrts_global_mutex);
 204     CILK_ASSERT(status == 0);*/
 205
 206     __cilkrts_set_tls_worker(w);
 207
 208     // Create a cilk fiber for this worker on this thread.
 209     START_INTERVAL(w, INTERVAL_FIBER_ALLOCATE_FROM_THREAD) {
 210         w->l->scheduling_fiber = cilk_fiber_allocate_from_thread();
 211         cilk_fiber_set_owner(w->l->scheduling_fiber, w);
 212     } STOP_INTERVAL(w, INTERVAL_FIBER_ALLOCATE_FROM_THREAD);
 213
 214     internal_run_scheduler_with_exceptions(w);
 215
 216     START_INTERVAL(w, INTERVAL_FIBER_DEALLOCATE_FROM_THREAD) {
 217         // Deallocate the scheduling fiber.  This operation reverses the
 218         // effect cilk_fiber_allocate_from_thread() and must be done in this
 219         // thread before it exits.
 220         int ref_count = cilk_fiber_deallocate_from_thread(w->l->scheduling_fiber);
 221         // Scheduling fibers should never have extra references to them.
 222         // We only get extra references into fibers because of Windows
 223         // exceptions.
 224         CILK_ASSERT(0 == ref_count);
 225         w->l->scheduling_fiber = NULL;
 226     } STOP_INTERVAL(w, INTERVAL_FIBER_DEALLOCATE_FROM_THREAD);
 227
 228     return 0;
 229 }
 230
 231
 232 /*
 233  * __cilkrts_user_worker_scheduling_stub
 234  *
 235  * Routine for the scheduling fiber created for an imported user
 236  * worker thread.  This method is analogous to
 237  * scheduler_thread_proc_for_system_worker.
 238  *
 239  */
 240 void __cilkrts_user_worker_scheduling_stub(cilk_fiber* fiber, void* null_arg)
 241 {
 242     __cilkrts_worker *w = __cilkrts_get_tls_worker();
 243
 244     // Sanity check.
 245     CILK_ASSERT(WORKER_USER == w->l->type);
 246
 247     // Enter the scheduling loop on the user worker.
 248     // This function will never return.
 249     __cilkrts_run_scheduler_with_exceptions(w);
 250
 251     // A WORKER_USER, at some point, will resume on the original stack and leave
 252     // Cilk.  Under no circumstances do we ever exit off of the bottom of this
 253     // stack.
 254     CILK_ASSERT(0);
 255 }
 256
 257 /**
 258  * We are exporting a function with this name to Inspector?
 259  * What a confusing name...
 260  *
 261  * This function is exported so Piersol's stack trace displays
 262  * reasonable information.
 263  */
 264 void* __cilkrts_worker_stub(void* arg)
 265 {
 266     return scheduler_thread_proc_for_system_worker(arg);
 267 }
 268
 269
 270
 271 // /* Return the lesser of the argument and the operating system
 272 //    limit on the number of workers (threads) that may or ought
 273 //    to be created. */
 274 // int sysdep_thread_limit(int n, int physical_cpus)
 275 // {
 276 //     /* On Linux thread creation fails somewhere short of the
 277 //        number of available processes. */
 278 //     struct rlimit lim;
 279
 280 //     if (n > 256 + 2 * physical_cpus)
 281 //         n = 256 + 2 * physical_cpus;
 282
 283 //     if (getrlimit(RLIMIT_NPROC, &lim) == 0 && lim.rlim_cur != RLIM_INFINITY)
 284 //     {
 285 //         /* If the limit reads 0 or absurdly small, ignore it. */
 286 //         unsigned int maxproc = (lim.rlim_cur * 3 + 3) / 4;
 287 //         if (maxproc > 8 + 2 * physical_cpus && maxproc < n)
 288 //             n = maxproc;
 289 //     }
 290 //     return n;
 291 // }
 292
 293
 294
 295 static void write_version_file (global_state_t *, int);
 296
 297 /* Create n worker threads from base..top-1
 298  */
 299 static void create_threads(global_state_t *g, int base, int top)
 300 {
 301     // TBD(11/30/12): We want to insert code providing the option of
 302     // pinning system workers to cores.
 303     for (int i = base; i < top; i++) {
 304         int status = pthread_create(&g->sysdep->threads[i],
 305                                     NULL,
 306                                     scheduler_thread_proc_for_system_worker,
 307                                     g->workers[i]);
 308         if (status != 0)
 309             __cilkrts_bug("Cilk runtime error: thread creation (%d) failed: %d\n", i, status);
 310     }
 311 }
 312
 313 #if PARALLEL_THREAD_CREATE
 314 static int volatile threads_created = 0;
 315
 316 // Create approximately half of the worker threads, and then become a worker
 317 // ourselves.
 318 static void * create_threads_and_work (void * arg)
 319 {
 320     global_state_t *g = ((__cilkrts_worker *)arg)->g;
 321
 322     create_threads(g, g->P/2, g->P-1);
 323     // Let the initial thread know that we're done.
 324     threads_created = 1;
 325
 326     // Ideally this turns into a tail call that wipes out this stack frame.
 327     return scheduler_thread_proc_for_system_worker(arg);
 328 }
 329 #endif
 330 void __cilkrts_start_workers(global_state_t *g, int n)
 331 {
 332     g->workers_running = 1;
 333     g->work_done = 0;
 334
 335     if (!g->sysdep->threads)
 336         return;
 337
 338     // Do we actually have any threads to create?
 339     if (n > 0)
 340     {
 341 #if PARALLEL_THREAD_CREATE
 342             int status;
 343             // We create (a rounded up) half of the threads, thread one creates the rest
 344             int half_threads = (n+1)/2;
 345
 346             // Create the first thread passing a different thread function, so that it creates threads itself
 347             status = pthread_create(&g->sysdep->threads[0], NULL, create_threads_and_work, g->workers[0]);
 348
 349             if (status != 0)
 350                 __cilkrts_bug("Cilk runtime error: thread creation (0) failed: %d\n", status);
 351
 352             // Then the rest of the ones we have to create
 353             create_threads(g, 1, half_threads);
 354
 355             // Now wait for the first created thread to tell us it's created all of its threads.
 356             // We could maybe drop this a bit lower and overlap with write_version_file.
 357             while (!threads_created)
 358                 __cilkrts_yield();
 359 #else
 360             // Simply create all the threads linearly here.
 361             create_threads(g, 0, n);
 362 #endif
 363     }
 364     // write the version information to a file if the environment is configured
 365     // for it (the function makes the check).
 366     write_version_file(g, n);
 367
 368
 369     return;
 370 }
 371
 372 void __cilkrts_stop_workers(global_state_t *g)
 373 {
 374     int i;
 375
 376     // Tell the workers to give up
 377
 378     g->work_done = 1;
 379
 380     if (g->workers_running == 0)
 381         return;
 382
 383     if (!g->sysdep->threads)
 384         return;
 385
 386     /* Make them all runnable. */
 387     if (g->P > 1) {
 388         CILK_ASSERT(g->workers[0]->l->signal_node);
 389         signal_node_msg(g->workers[0]->l->signal_node, 1);
 390     }
 391
 392         for (i = 0; i < g->P - 1; ++i) {
 393             int sc_status;
 394             void *th_status;
 395
 396             sc_status = pthread_join(g->sysdep->threads[i], &th_status);
 397             if (sc_status != 0)
 398                 __cilkrts_bug("Cilk runtime error: thread join (%d) failed: %d\n", i, sc_status);
 399         }
 400
 401     g->workers_running = 0;
 402
 403
 404     return;
 405 }
 406
 407
 408 /*
 409  * @brief Returns the stack address for resuming execution of sf.
 410  *
 411  * This method takes in the top of the stack to use, and then returns
 412  * a properly aligned address for resuming execution of sf.
 413  *
 414  *  @param sf           -   The stack frame we want to resume executing.
 415  *  @param stack_base   -   The top of the stack we want to execute sf on.
 416  *
 417  */
 418 static char* get_sp_for_executing_sf(char* stack_base,
 419                                      full_frame *ff,
 420                                      __cilkrts_stack_frame *sf)
 421 {
 422 // The original calculation that had been done to correct the stack
 423 // pointer when resuming execution.
 424 //
 425 // But this code was never getting called in the eng branch anyway...
 426 //
 427 // TBD(11/30/12): This logic needs to be revisited to make sure that
 428 // we are doing the proper calculation in reserving space for outgoing
 429 // arguments on all platforms and architectures.
 430 #if 0
 431     /* Preserve outgoing argument space and stack alignment on steal.
 432        Outgoing argument space is bounded by the difference between
 433        stack and frame pointers.  Some user code is known to rely on
 434        16 byte alignment.  Maintain 32 byte alignment for future
 435        compatibility. */
 436 #define SMASK 31 /* 32 byte alignment */
 437     if (sf) {
 438         char *fp = FP(sf), *sp = SP(sf);
 439         int fp_align = (int)(size_t)fp & SMASK;
 440         ptrdiff_t space = fp - sp;
 441
 442         fprintf(stderr, "Here: fp = %p, sp = %p\n", fp, sp);
 443         char *top_aligned = (char *)((((size_t)stack_base - SMASK) & ~(size_t)SMASK) | fp_align);
 444         /* Don't allocate an unreasonable amount of stack space. */
 445
 446         fprintf(stderr, "Here: stack_base = %p, top_aligned=%p, space=%ld\n",
 447                 stack_base, top_aligned, space);
 448         if (space < 32)
 449             space = 32 + (space & SMASK);
 450         else if (space > 40 * 1024)
 451             space = 40 * 1024 + (space & SMASK);
 452
 453         return top_aligned - space;
 454     }
 455 #endif
 456
 457 #define PERFORM_FRAME_SIZE_CALCULATION 0
 458
 459     char* new_stack_base = stack_base - 256;
 460
 461 #if PERFORM_FRAME_SIZE_CALCULATION
 462     // If there is a frame size saved, then use that as the
 463     // correction instead of 256.
 464     if (ff->frame_size > 0) {
 465         if (ff->frame_size < 40*1024) {
 466             new_stack_base = stack_base - ff->frame_size;
 467         }
 468         else {
 469             // If for some reason, our frame size calculation is giving us
 470             // a number which is bigger than about 10 pages, then
 471             // there is likely something wrong here?  Don't allocate
 472             // an unreasonable amount of space.
 473             new_stack_base = stack_base - 40*1024;
 474         }
 475     }
 476 #endif
 477
 478     // Whatever correction we choose, align the final stack top.
 479     // This alignment seems to be necessary in particular on 32-bit
 480     // Linux, and possibly Mac. (Is 32-byte alignment is sufficient?)
 481     /* 256-byte alignment. Why not? */
 482     const uintptr_t align_mask = ~(256 -1);
 483     new_stack_base = (char*)((size_t)new_stack_base & align_mask);
 484     return new_stack_base;
 485 }
 486
 487 char* sysdep_reset_jump_buffers_for_resume(cilk_fiber* fiber,
 488                                            full_frame *ff,
 489                                            __cilkrts_stack_frame *sf)
 490 {
 491 #if FIBER_DEBUG >= 4
 492     fprintf(stderr, "ThreadId=%p (fiber_proc_to_resume), Fiber %p.  sf = %p. ff=%p, ff->sync_sp=%p\n",
 493             cilkos_get_current_thread_id(),
 494             fiber,
 495             sf,
 496             ff, ff->sync_sp);
 497 #endif
 498
 499     CILK_ASSERT(fiber);
 500     void* sp = (void*)get_sp_for_executing_sf(cilk_fiber_get_stack_base(fiber), ff, sf);
 501     SP(sf) = sp;
 502
 503     /* Debugging: make sure stack is accessible. */
 504     ((volatile char *)sp)[-1];
 505
 506     // Adjust the saved_sp to account for the SP we're about to run.  This will
 507     // allow us to track fluctations in the stack
 508 #if FIBER_DEBUG >= 4
 509     fprintf(stderr, "ThreadId=%p, about to take stack ff=%p, sp=%p, sync_sp=%p\n",
 510             cilkos_get_current_thread_id(),
 511             ff,
 512             sp,
 513             ff->sync_sp);
 514 #endif
 515     __cilkrts_take_stack(ff, sp);
 516     return sp;
 517 }
 518
 519
 520 NORETURN sysdep_longjmp_to_sf(char* new_sp,
 521                               __cilkrts_stack_frame *sf,
 522                               full_frame *ff_for_exceptions /* UNUSED on Unix */)
 523 {
 524 #if FIBER_DEBUG >= 3
 525     fprintf(stderr,
 526             "ThreadId=%p. resume user code, sf=%p, new_sp = %p, original SP(sf) = %p, FP(sf) = %p\n",
 527             cilkos_get_current_thread_id(), sf, new_sp, SP(sf), FP(sf));
 528 #endif
 529
 530     // Set the stack pointer.
 531     SP(sf) = new_sp;
 532
 533 #ifdef RESTORE_X86_FP_STATE
 534     if (CILK_FRAME_VERSION_VALUE(sf->flags) >= 1) {
 535         // Restore the floating point state that was set in this frame at the
 536         // last spawn.
 537         //
 538         // This feature is only available in ABI 1 or later frames, and only
 539         // needed on IA64 or Intel64 processors.
 540         restore_x86_fp_state(sf);
 541     }
 542 #endif
 543
 544     CILK_LONGJMP(sf->ctx);
 545 }
 546
 547
 548 #include <stddef.h>
 549 #include <stdlib.h>
 550 #include <string.h>
 551 #include <sys/mman.h>
 552 #include <errno.h>
 553
 554
 555 void __cilkrts_make_unrunnable_sysdep(__cilkrts_worker *w,
 556                                       full_frame *ff,
 557                                       __cilkrts_stack_frame *sf,
 558                                       int is_loot,
 559                                       const char *why)
 560 {
 561     (void)w; /* unused */
 562     sf->except_data = 0;
 563
 564     if (is_loot)
 565     {
 566         if (ff->frame_size == 0)
 567         ff->frame_size = __cilkrts_get_frame_size(sf);
 568
 569         // Null loot's sp for debugging purposes (so we'll know it's not valid)
 570     SP(sf) = 0;
 571     }
 572 }
 573
 574 /*
 575  * __cilkrts_sysdep_is_worker_thread_id
 576  *
 577  * Returns true if the thread ID specified matches the thread ID we saved
 578  * for a worker.
 579  */
 580
 581 int __cilkrts_sysdep_is_worker_thread_id(global_state_t *g,
 582                                          int i,
 583                                          void *thread_id)
 584 {
 585 #if defined( __linux__) || defined(__VXWORKS__)
 586     pthread_t tid = *(pthread_t *)thread_id;
 587     if (i < 0 || i > g->total_workers)
 588         return 0;
 589     return g->sysdep->threads[i] == tid;
 590 #else
 591     // Needs to be implemented
 592     return 0;
 593 #endif
 594 }
 595
 596
 597
 598
 599 /*************************************************************
 600   Version information:
 601 *************************************************************/
 602
 603 #include <dlfcn.h>
 604 #include "internal/cilk_version.h"
 605 #include <stdio.h>
 606 #include <sys/utsname.h>
 607
 608 #ifdef __VXWORKS__
 609 #include <version.h>
 610 # endif
 611
 612 /* (Non-static) dummy function is used by get_runtime_path() to find the path
 613  * to the .so containing the Cilk runtime.
 614  */
 615 void dummy_function() { }
 616
 617 /* return a string with the path to the Cilk runtime, or "unknown" if the path
 618  * cannot be determined.
 619  */
 620 static const char *get_runtime_path ()
 621 {
 622 #ifdef __CYGWIN__
 623     // Cygwin doesn't support dladdr, which sucks
 624     return "unknown";
 625 #else
 626     Dl_info info;
 627     if (0 == dladdr(dummy_function, &info)) return "unknown";
 628     return info.dli_fname;
 629 #endif
 630 }
 631
 632 /* if the environment variable, CILK_VERSION, is defined, writes the version
 633  * information to the specified file.
 634  * g is the global state that was just created, and n is the number of workers
 635  * that were made (or requested from RML) for it.
 636  */
 637 static void write_version_file (global_state_t *g, int n)
 638 {
 639     const char *env;      // environment variable.
 640     char buf[256];        // print buffer.
 641     time_t t;
 642     FILE *fp;
 643     struct utsname sys_info;
 644     int err;              // error code from system calls.
 645
 646     // if CILK_VERSION is not set, or if the file cannot be opened, fail
 647     // silently.  Otherwise open the file for writing (or use stderr or stdout
 648     // if the user specifies).
 649     if (NULL == (env = getenv("CILK_VERSION"))) return;
 650     if (0 == strcasecmp(env, "stderr"))         fp = stderr;
 651     else if (0 == strcasecmp(env, "stdout"))    fp = stdout;
 652     else if (NULL == (fp = fopen(env, "w")))    return;
 653
 654     // get a string for the current time.  E.g.,
 655     // Cilk runtime initialized: Thu Jun 10 13:28:00 2010
 656     t = time(NULL);
 657     strftime(buf, 256, "%a %b %d %H:%M:%S %Y", localtime(&t));
 658     fprintf(fp, "Cilk runtime initialized: %s\n", buf);
 659
 660     // Print runtime info.  E.g.,
 661     // Cilk runtime information
 662     // ========================
 663     // Cilk version: 2.0.0 Build 9184
 664     // Built by willtor on host willtor-desktop
 665     // Compilation date: Thu Jun 10 13:27:42 2010
 666     // Compiled with ICC V99.9.9, ICC build date: 20100610
 667
 668     fprintf(fp, "\nCilk runtime information\n");
 669     fprintf(fp, "========================\n");
 670     fprintf(fp, "Cilk version: %d.%d.%d Build %d\n",
 671             VERSION_MAJOR,
 672             VERSION_MINOR,
 673             VERSION_REV,
 674             VERSION_BUILD);
 675 #ifdef __VXWORKS__
 676     char * vxWorksVer = VXWORKS_VERSION;
 677     fprintf(fp, "Cross compiled for %s\n",vxWorksVer);
 678     // user and host not avalible if VxWorks cross compiled on windows build host
 679 #else
 680
 681     // User and host are not available for GCC builds
 682 #ifdef BUILD_USER
 683     fprintf(fp, "Built by "BUILD_USER" on host "BUILD_HOST"\n");
 684 #endif // BUILD_USER
 685 #endif // __VXWORKS__
 686
 687     // GCC has requested that this be removed for GCC builds
 688 #ifdef BUILD_USER
 689     fprintf(fp, "Compilation date: "__DATE__" "__TIME__"\n");
 690 #endif // BUILD_USER
 691
 692 #ifdef __INTEL_COMPILER
 693     // Compiled by the Intel C/C++ compiler.
 694     fprintf(fp, "Compiled with ICC V%d.%d.%d, ICC build date: %d\n",
 695             __INTEL_COMPILER / 100,
 696             (__INTEL_COMPILER / 10) % 10,
 697             __INTEL_COMPILER % 10,
 698             __INTEL_COMPILER_BUILD_DATE);
 699 #else
 700     // Compiled by GCC.
 701     fprintf(fp, "Compiled with GCC V%d.%d.%d\n",
 702             __GNUC__,
 703             __GNUC_MINOR__,
 704             __GNUC_PATCHLEVEL__);
 705 #endif // defined __INTEL_COMPILER
 706
 707     // Print system info.  E.g.,
 708     // System information
 709     // ==================
 710     // Cilk runtime path: /opt/icc/64/lib/libcilkrts.so.5
 711     // System OS: Linux, release 2.6.28-19-generic
 712     // System architecture: x86_64
 713
 714     err = uname(&sys_info);
 715     fprintf(fp, "\nSystem information\n");
 716     fprintf(fp, "==================\n");
 717     fprintf(fp, "Cilk runtime path: %s\n", get_runtime_path());
 718     fprintf(fp, "System OS: %s, release %s\n",
 719             err < 0 ? "unknown" : sys_info.sysname,
 720             err < 0 ? "?" : sys_info.release);
 721     fprintf(fp, "System architecture: %s\n",
 722             err < 0 ? "unknown" : sys_info.machine);
 723
 724     // Print thread info.  E.g.,
 725     // Thread information
 726     // ==================
 727     // System cores: 8
 728     // Cilk workers requested: 8
 729     // Thread creator: Private
 730
 731     fprintf(fp, "\nThread information\n");
 732     fprintf(fp, "==================\n");
 733 #ifdef __VXWORKS__
 734     fprintf(fp, "System cores: %d\n", (int)__builtin_popcount(vxCpuEnabledGet()));
 735 #else
 736     fprintf(fp, "System cores: %d\n", (int)sysconf(_SC_NPROCESSORS_ONLN));
 737 #endif
 738     fprintf(fp, "Cilk workers requested: %d\n", n);
 739 #if (PARALLEL_THREAD_CREATE)
 740         fprintf(fp, "Thread creator: Private (parallel)\n");
 741 #else
 742         fprintf(fp, "Thread creator: Private\n");
 743 #endif
 744
 745     if (fp != stderr && fp != stdout) fclose(fp);
 746     else fflush(fp); // flush the handle buffer if it is stdout or stderr.
 747 }
 748
 749
 750 /*
 751  * __cilkrts_establish_c_stack
 752  *
 753  * Tell Cilkscreen about the user stack bounds.
 754  *
 755  * Note that the Cilk V1 runtime only included the portion of the stack from
 756  * the entry into Cilk, down.  We don't appear to be able to find that, but
 757  * I think this will be sufficient.
 758  */
 759
 760 void __cilkrts_establish_c_stack(void)
 761 {
 762     /* FIXME: Not implemented. */
 763
 764     /* TBD: Do we need this */
 765     /*
 766     void __cilkrts_cilkscreen_establish_c_stack(char *begin, char *end);
 767
 768     size_t r;
 769     MEMORY_BASIC_INFORMATION mbi;
 770
 771     r = VirtualQuery (&mbi,
 772                       &mbi,
 773                       sizeof(mbi));
 774
 775     __cilkrts_cilkscreen_establish_c_stack((char *)mbi.BaseAddress,
 776                                    (char *)mbi.BaseAddress + mbi.RegionSize);
 777     */
 778 }
 779
 780
 781 /*
 782  * internal_enforce_global_visibility
 783  *
 784  * Ensure global visibility of public symbols, for proper Cilk-TBB interop.
 785  *
 786  * If Cilk runtime is loaded dynamically, its symbols might remain unavailable
 787  * for global search with dladdr; that might prevent TBB from finding Cilk
 788  * in the process address space and initiating the interop protocol.
 789  * The workaround is for the library to open itself with RTLD_GLOBAL flag.
 790  */
 791
 792 static __attribute__((noinline))
 793 void internal_enforce_global_visibility()
 794 {
 795     void* handle = dlopen( get_runtime_path(), RTLD_GLOBAL|RTLD_LAZY );
 796
 797     /* For proper reference counting, close the handle immediately. */
 798     if( handle) dlclose(handle);
 799 }
 800
 801 /*
 802   Local Variables: **
 803   c-file-style:"bsd" **
 804   c-basic-offset:4 **
 805   indent-tabs-mode:nil **
 806   End: **
 807 */