[OpenMP] Initialize runtime in the forked child process

[hbae]

Summary

This patch enables serial initialization in the forked child process
to fix unstable runtime behavior when used with Python-based AI tools.

[AndreyChurbanov]

LGTM

openmp/runtime/src/z_Linux_util.cpp

 }
 
 static void __kmp_atfork_prepare(void) {
   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
 }
 
 static void __kmp_atfork_parent(void) {
-  __kmp_release_bootstrap_lock(&__kmp_initz_lock);
   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
+  __kmp_release_bootstrap_lock(&__kmp_initz_lock);
 }
 
 /* Reset the library so execution in the child starts "all over again" with
    clean data structures in initial states.  Don't worry about freeing memory
    allocated by parent, just abandon it to be safe. */
 static void __kmp_atfork_child(void) {
   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
+  __kmp_release_bootstrap_lock(&__kmp_initz_lock);
   /* TODO make sure this is done right for nested/sibling */
   // ATT:  Memory leaks are here? TODO: Check it and fix.
   /* KMP_ASSERT( 0 ); */
 
   ++__kmp_fork_count;
 
 #if KMP_AFFINITY_SUPPORTED
 #if KMP_OS_LINUX || KMP_OS_FREEBSD
   // reset the affinity in the child to the initial thread
   // affinity in the parent
   kmp_set_thread_affinity_mask_initial();
 #endif
   // Set default not to bind threads tightly in the child (we’re expecting
   // over-subscription after the fork and this can improve things for
   // scripting languages that use OpenMP inside process-parallel code).
   __kmp_affinity_type = affinity_none;
   if (__kmp_nested_proc_bind.bind_types != NULL) {
     __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
   }
 #endif // KMP_AFFINITY_SUPPORTED
 
-  __kmp_init_runtime = FALSE;
 #if KMP_USE_MONITOR
   __kmp_init_monitor = 0;
 #endif
   __kmp_init_parallel = FALSE;
   __kmp_init_middle = FALSE;
   __kmp_init_serial = FALSE;
   TCW_4(__kmp_init_gtid, FALSE);
   __kmp_init_common = FALSE;
#endif
   __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
   __kmp_init_bootstrap_lock(&__kmp_console_lock);
   __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
 
 #if USE_ITT_BUILD
   __kmp_itt_reset(); // reset ITT's global state
 #endif /* USE_ITT_BUILD */
 
+  __kmp_serial_initialize();
+
   /* This is necessary to make sure no stale data is left around */
   /* AC: customers complain that we use unsafe routines in the atfork
      handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
      in dynamic_link when check the presence of shared tbbmalloc library.
      Suggestion is to make the library initialization lazier, similar
      to what done for __kmpc_begin(). */
   // TODO: synchronize all static initializations with regular library
   //       startup; look at kmp_global.cpp and etc.