[OpenMP] Have hidden helper team allocate new OS threads only (llvm#8…

…7119)

The hidden helper team pre-allocates the gtid space [1,
num_hidden_helpers] (inclusive). If regular host threads are allocated,
then put back in the thread pool, then the hidden helper team is
initialized, the hidden helper team tries to allocate the threads from
the thread pool with gtids higher than [1, num_hidden_helpers]. Instead,
have the hidden helper team fork OS threads so the correct gtid range
used for hidden helper threads.

Fixes: llvm#87117

openmp/runtime/src/kmp_runtime.cpp

@@ -4431,8 +4431,10 @@ kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
 #endif
   KMP_MB();
 
-  /* first, try to get one from the thread pool */
-  if (__kmp_thread_pool) {
+  /* first, try to get one from the thread pool unless allocating thread is
+   * the main hidden helper thread. The hidden helper team should always
+   * allocate new OS threads. */
+  if (__kmp_thread_pool && !KMP_HIDDEN_HELPER_TEAM(team)) {
     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
     if (new_thr == __kmp_thread_pool_insert_pt) {
@@ -4497,7 +4499,7 @@ kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
   }
 
   /* no, well fork a new one */
-  KMP_ASSERT(__kmp_nth == __kmp_all_nth);
+  KMP_ASSERT(KMP_HIDDEN_HELPER_TEAM(team) || __kmp_nth == __kmp_all_nth);
   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
 
 #if KMP_USE_MONITOR

openmp/runtime/test/tasking/hidden_helper_task/issue-87117.c

@@ -0,0 +1,36 @@
+// RUN: %libomp-compile
+// RUN: env KMP_HOT_TEAMS_MODE=0 KMP_HOT_TEAMS_MAX_LEVEL=1 %libomp-run
+//
+// Force the defaults of:
+// KMP_HOT_TEAMS_MODE=0 means free extra threads after parallel
+//   involving non-hot team
+// KMP_HOT_TEAMS_MAX_LEVEL=1 means only the initial outer team
+//   is a hot team.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <omp.h>
+
+int main() {
+  int a;
+  omp_set_max_active_levels(2);
+// This nested parallel creates extra threads on the thread pool
+#pragma omp parallel num_threads(2)
+  {
+#pragma omp parallel num_threads(2)
+    {
+#pragma omp atomic
+      a++;
+    }
+  }
+
+// Causes assert if hidden helper thread tries to allocate from thread pool
+// instead of creating new OS threads
+#pragma omp parallel num_threads(1)
+  {
+#pragma omp target nowait
+    { a++; }
+  }
+
+  return EXIT_SUCCESS;
+}