[OpenMP] Initialize runtime in the forked child process

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

Differential Revision: https://reviews.llvm.org/D93230

Author: hansangbae

openmp/runtime/src/z_Linux_util.cpp

@@ -1272,15 +1272,16 @@ static void __kmp_atfork_prepare(void) {
 }
 
 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 ); */
@@ -1302,7 +1303,6 @@ static void __kmp_atfork_child(void) {
   }
 #endif // KMP_AFFINITY_SUPPORTED
 
-  __kmp_init_runtime = FALSE;
 #if KMP_USE_MONITOR
   __kmp_init_monitor = 0;
 #endif
@@ -1355,6 +1355,8 @@ static void __kmp_atfork_child(void) {
   __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