Fixed memory leak reported in https://bugs.llvm.org/show_bug.cgi?id=41494

[AndreyChurbanov]

Summary

Might be a bit too verbose (not sure if we need to check that the number of threads reduced to 0), but it fixes the memory leak(s) for me. Multiple leaks were possible in case of multiple foreign threads registered into the library.

[AndreyChurbanov]

formatted

[lebedev.ri]

Thank you for the fix!
I can't really comment on the correctness of the fix, but i can confirm that it does indeed fix the leak i was seeing.

runtime/src/kmp_runtime.cpp

#if OMPT_SUPPORT
   if (ompt_enabled.ompt_callback_thread_end) {
     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
   }
 #endif
 
   TCW_4(__kmp_nth,
         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
-  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
+  i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
                  " to %d\n",
                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
-
+  if (i == 1) {
+    // need to free contention group structure
+    KMP_DEBUG_ASSERT(root->r.r_uber_thread ==
+                     root->r.r_uber_thread->th.th_cg_roots->cg_root);
+    KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL);
+    __kmp_free(root->r.r_uber_thread->th.th_cg_roots);
+    root->r.r_uber_thread->th.th_cg_roots = NULL;
+  }
   __kmp_reap_thread(root->r.r_uber_thread, 1);
 
   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
   // of freeing.
   root->r.r_uber_thread = NULL;
   /* mark root as no longer in use */
   root->r.r_begin = FALSE;