Fix thread_limits to work properly for teams construct

[tlwilmar]

Summary

The thread-limit-var and omp_get_thread_limit API was not perfectly handled for teams construct. Now, when modified by thread_limit clause, omp_get_thread_limit reports the correct value. In addition, the value is restored when leaving the teams construct to what it was in the encountering context.

This is done partly by creating the notion of a Contention Group root (CG root) that keeps track of the thread at the root of each separate CG, the thread-limit-var associated with the CG, and associated counter of active threads within the contention group.

thread-limits are passed from master to worker threads via an entry in the ICV data structure. When a "contention group switch" occurs, a new CG root record is made and passed from master to worker. A thread could potentially have several CG root records if it encounters multiple nested teams constructs (but at the moment the spec doesn't allow for nested teams, so the most one could have currently is 2). The master of the teams masters gets the thread-limit clause value stored to its local ICV structure, and the other teams masters copy it from the master. The thread-limit is set from that ICV copy and restored to the ICV copy when entering and leaving the teams construct.

This change also fixes a bug when the top-level teams construct team gets reused, and OMP_DYNAMIC was true, which can cause the expected size of this team to be smaller than what was actually allocated. The fix updates the size of the team after its threads were reserved.

[protze.joachim]

I'm a bit confused about the explicit copy per task.

runtime/src/kmp.h

1850

Is this entry necessary? From my understanding of the spec, one copy per cg or thread should be sufficient.

runtime/src/kmp_runtime.cpp

7281

If this is the reason for a per task icv, can't we reset the cg value to the initial value after leaving the teams region?

[AndreyChurbanov]

runtime/src/kmp.h

1850

Maybe you are correct, but the specification says it is per-task ICV. So this looks a safer way to go. E.g. in future the setter routine can be added.

runtime/src/kmp_runtime.cpp

5675

This looks incorrect, because worker thread can only belong to one CG at a time, not to chain of CGs.

7281

I'd guess the main reason for a per task ICV is the specification requirement. Technically the implementation can ignore it and make the ICV per contention group, but this sounds a bit risky to me.

[protze.joachim]

runtime/src/kmp.h

1850

The spec does not litteraly say per-task ICV. The scope is "data environment".

data environment: The variables associated with the execution of a given region.

Since the value can only be set/changed on a per-cg base, this should be sufficient.

[AndreyChurbanov]

runtime/src/kmp.h

1850

I lean to agreeing with you that thread-limit-var is rather per-contention-group. Regardless that each task has its own data environment, and spec says "There is one copy of this ICV per data environment", skipping per-task thread-limit-var ICV might be feasible. Though not easy, as the push / pop operations for CG-root structures need to be revised then.

[tlwilmar]

I was going for the per-data-environment specification for the ICV, though I agree it is only needed per contention group. Is there any consensus on this? Leave as is, or remove per-data-environment ICV?

[protze.joachim]

Since there is only one other ICV which might be handled similary (bind-var), I will not further argue for making a special case for thread-limit-var. So leave it as it is.

[AndreyChurbanov]

Per-task ICV works for me.

But please fix the broken CG threads counter in __kmp_free_thread.

[tlwilmar]

Fix update for non-CG_root threads in __kmp_free_thread.

[AndreyChurbanov]

LGTM

[lebedev.ri]

@tlwilmar are you registered on llvm's bugzilla?
This patch introduced a leak, please see https://bugs.llvm.org/show_bug.cgi?id=41494

openmp/trunk/runtime/src/kmp_runtime.cpp

3869 ↗

(On Diff #186316)

This is being leaked in the end

runtime/src/kmp.h

typedef struct kmp_internal_control {
   kmp_int8
       bt_set; /* internal control for whether blocktime is explicitly set */
   int blocktime; /* internal control for blocktime */
 #if KMP_USE_MONITOR
   int bt_intervals; /* internal control for blocktime intervals */
 #endif
   int nproc; /* internal control for #threads for next parallel region (per
                 thread) */
+  int thread_limit; /* internal control for thread-limit-var */
   int max_active_levels; /* internal control for max_active_levels */
   kmp_r_sched_t
       sched; /* internal control for runtime schedule {sched,chunk} pair */
 #if OMP_40_ENABLED
   kmp_proc_bind_t proc_bind; /* internal control for affinity  */
   kmp_int32 default_device; /* internal control for default device */
 #endif // OMP_40_ENABLED
   struct kmp_internal_control *next;

 #define set__dynamic(xthread, xval)                                            \
   (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
 #define get__dynamic(xthread)                                                  \
   (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
 
 #define set__nproc(xthread, xval)                                              \
   (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
 
+#define set__thread_limit(xthread, xval)                                       \
+  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
+
 #define set__max_active_levels(xthread, xval)                                  \
   (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
 
 #define set__sched(xthread, xval)                                              \
   (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
 
 #if OMP_40_ENABLED
 

 #endif
 #if OMP_40_ENABLED
 typedef struct kmp_teams_size {
   kmp_int32 nteams; // number of teams in a league
   kmp_int32 nth; // number of threads in each team of the league
 } kmp_teams_size_t;
 #endif
 
+// This struct stores a thread that acts as a "root" for a contention
+// group. Contention groups are rooted at kmp_root threads, but also at
+// each master thread of each team created in the teams construct.
+// This struct therefore also stores a thread_limit associated with
+// that contention group, and a counter to track the number of threads
+// active in that contention group. Each thread has a list of these: CG
+// root threads have an entry in their list in which cg_root refers to
+// the thread itself, whereas other workers in the CG will have a
+// single entry where cg_root is same as the entry containing their CG
+// root. When a thread encounters a teams construct, it will add a new
+// entry to the front of its list, because it now roots a new CG.
+typedef struct kmp_cg_root {
+  kmp_info_p *cg_root; // "root" thread for a contention group
+  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
+  // thread_limit clause for teams masters
+  kmp_int32 cg_thread_limit;
+  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
+  struct kmp_cg_root *up; // pointer to higher level CG root in list
+} kmp_cg_root_t;
+
 // OpenMP thread data structures
 
 typedef struct KMP_ALIGN_CACHE kmp_base_info {
   /* Start with the readonly data which is cache aligned and padded. This is
      written before the thread starts working by the master. Uber masters may
      update themselves later. Usage does not consider serialized regions.  */
   kmp_desc_t th_info;
   kmp_team_p *th_team; /* team we belong to */

 // alignment ???
 #endif /* USE_ITT_BUILD */
 #if KMP_STATS_ENABLED
   kmp_stats_list *th_stats;
 #endif
 #if KMP_OS_UNIX
   std::atomic<bool> th_blocking;
 #endif
+  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
 } kmp_base_info_t;
 
 typedef union KMP_ALIGN_CACHE kmp_info {
   double th_align; /* use worst case alignment */
   char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
   kmp_base_info_t th;
 } kmp_info_t;
 
typedef struct kmp_base_root {
   std::atomic<int> r_in_parallel;
   // GEH: This is misnamed, should be r_active_levels
   kmp_team_t *r_root_team;
   kmp_team_t *r_hot_team;
   kmp_info_t *r_uber_thread;
   kmp_lock_t r_begin_lock;
   volatile int r_begin;
   int r_blocktime; /* blocktime for this root and descendants */
-  int r_cg_nthreads; // count of active threads in a contention group
 } kmp_base_root_t;
 
 typedef union KMP_ALIGN_CACHE kmp_root {
   kmp_base_root_t r;
   double r_align; /* use worst case alignment */
   char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
 } kmp_root_t;
 

runtime/src/kmp_csupport.cpp

#endif
                   );
   __kmp_join_call(loc, gtid
 #if OMPT_SUPPORT
                   ,
                   fork_context_intel
 #endif
                   );
 
+  // Pop current CG root off list
+  KMP_DEBUG_ASSERT(this_thr->th.th_cg_roots);
+  kmp_cg_root_t *tmp = this_thr->th.th_cg_roots;
+  this_thr->th.th_cg_roots = tmp->up;
+  KA_TRACE(100, ("__kmpc_fork_teams: Thread %p popping node %p and moving up"
+                 " to node %p. cg_nthreads was %d\n",
+                 this_thr, tmp, this_thr->th.th_cg_roots, tmp->cg_nthreads));
+  __kmp_free(tmp);
+  // Restore current task's thread_limit from CG root
+  KMP_DEBUG_ASSERT(this_thr->th.th_cg_roots);
+  this_thr->th.th_current_task->td_icvs.thread_limit =
+      this_thr->th.th_cg_roots->cg_thread_limit;
+
   this_thr->th.th_teams_microtask = NULL;
   this_thr->th.th_teams_level = 0;
   *(kmp_int64 *)(&this_thr->th.th_teams_size) = 0L;
   va_end(ap);
 }
 #endif /* OMP_40_ENABLED */
 
 // I don't think this function should ever have been exported.

runtime/src/kmp_ftn_entry.h

#else
   return __kmp_get_team_size(__kmp_entry_gtid(), KMP_DEREF level);
 #endif
 }
 
 int FTN_STDCALL KMP_EXPAND_NAME(FTN_GET_THREAD_LIMIT)(void) {
 #ifdef KMP_STUB
   return 1; // TO DO: clarify whether it returns 1 or 0?
 #else
+  int gtid;
+  kmp_info_t *thread;
   if (!__kmp_init_serial) {
     __kmp_serial_initialize();
   }
-  /* global ICV */
-  return __kmp_cg_max_nth;
+
+  gtid = __kmp_entry_gtid();
+  thread = __kmp_threads[gtid];
+  return thread->th.th_current_task->td_icvs.thread_limit;
 #endif
 }
 
 int FTN_STDCALL KMP_EXPAND_NAME(FTN_IN_FINAL)(void) {
 #ifdef KMP_STUB
   return 0; // TO DO: clarify whether it returns 1 or 0?
 #else
   if (!TCR_4(__kmp_init_parallel)) {

runtime/src/kmp_runtime.cpp

#if OMP_40_ENABLED
                                  ,
                                  int enter_teams
 #endif /* OMP_40_ENABLED */
                                  ) {
   int capacity;
   int new_nthreads;
   KMP_DEBUG_ASSERT(__kmp_init_serial);
   KMP_DEBUG_ASSERT(root && parent_team);
+  kmp_info_t *this_thr = parent_team->t.t_threads[master_tid];
 
   // If dyn-var is set, dynamically adjust the number of desired threads,
   // according to the method specified by dynamic_mode.
   new_nthreads = set_nthreads;
   if (!get__dynamic_2(parent_team, master_tid)) {
     ;
   }
 #ifdef USE_LOAD_BALANCE
if (__kmp_nth + new_nthreads -
     }
     KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
                   "reservation to %d threads\n",
                   master_tid, tl_nthreads));
     new_nthreads = tl_nthreads;
   }
 
   // Respect OMP_THREAD_LIMIT
-  if (root->r.r_cg_nthreads + new_nthreads -
+  int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads;
+  int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit;
+  if (cg_nthreads + new_nthreads -
           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
-      __kmp_cg_max_nth) {
-    int tl_nthreads = __kmp_cg_max_nth - root->r.r_cg_nthreads +
+      max_cg_threads) {
+    int tl_nthreads = max_cg_threads - cg_nthreads +
                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
     if (tl_nthreads <= 0) {
       tl_nthreads = 1;
     }
 
     // If dyn-var is false, emit a 1-time warning.
     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
       __kmp_reserve_warn = 1;

 #endif /* OMP_40_ENABLED */
                      ) {
   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
   kmp_team_t *team;
   kmp_team_t *parent_team;
   kmp_info_t *master_th;
   kmp_root_t *root;
   int master_active;
-  int i;
 
   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
 
   /* setup current data */
   master_th = __kmp_threads[gtid];
   root = master_th->th.th_root;
   team = master_th->th.th_team;
   parent_team = team->t.t_parent;
if (master_th->th.th_teams_microtask && !exit_teams &&
     // inside the teams construct, so that at the next parallel same (hot) team
     // works, only adjust nesting levels
 
     /* Decrement our nested depth level */
     team->t.t_level--;
     team->t.t_active_level--;
     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
 
-    /* Restore number of threads in the team if needed */
+    // Restore number of threads in the team if needed. This code relies on
+    // the proper adjustment of th_teams_size.nth after the fork in
+    // __kmp_teams_master on each teams master in the case that
+    // __kmp_reserve_threads reduced it.
     if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
       int old_num = master_th->th.th_team_nproc;
       int new_num = master_th->th.th_teams_size.nth;
       kmp_info_t **other_threads = team->t.t_threads;
       team->t.t_nproc = new_num;
-      for (i = 0; i < old_num; ++i) {
+      for (int i = 0; i < old_num; ++i) {
         other_threads[i]->th.th_team_nproc = new_num;
       }
       // Adjust states of non-used threads of the team
-      for (i = old_num; i < new_num; ++i) {
+      for (int i = old_num; i < new_num; ++i) {
         // Re-initialize thread's barrier data.
-        int b;
+        KMP_DEBUG_ASSERT(other_threads[i]);
         kmp_balign_t *balign = other_threads[i]->th.th_bar;
-        for (b = 0; b < bs_last_barrier; ++b) {
+        for (int b = 0; b < bs_last_barrier; ++b) {
           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
 #if USE_DEBUGGER
           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
 #endif
         }
         if (__kmp_tasking_mode != tskm_immediate_exec) {
           // Synchronize thread's task state
kmp_internal_control_t g_icvs = {
     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
 #if KMP_USE_MONITOR
     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
 // intervals
 #endif
     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
     // next parallel region (per thread)
     // (use a max ub on value if __kmp_parallel_initialize not called yet)
+    __kmp_cg_max_nth, // int thread_limit;
     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
     // for max_active_levels
     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
 // {sched,chunk} pair
 #if OMP_40_ENABLED
     __kmp_nested_proc_bind.bind_types[0],
     __kmp_default_device,
 #endif /* OMP_40_ENABLED */
static void __kmp_initialize_root(kmp_root_t *root) {
 
   /* setup the root state structure */
   __kmp_init_lock(&root->r.r_begin_lock);
   root->r.r_begin = FALSE;
   root->r.r_active = FALSE;
   root->r.r_in_parallel = 0;
   root->r.r_blocktime = __kmp_dflt_blocktime;
   root->r.r_nested = __kmp_dflt_nested;
-  root->r.r_cg_nthreads = 1;
 
   /* setup the root team for this task */
   /* allocate the root team structure */
   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
 
   root_team =
       __kmp_allocate_team(root,
                           1, // new_nproc
#endif
   }
 #endif /* KMP_AFFINITY_SUPPORTED */
 #if OMP_50_ENABLED
   root_thread->th.th_def_allocator = __kmp_def_allocator;
   root_thread->th.th_prev_level = 0;
   root_thread->th.th_prev_num_threads = 1;
 #endif
 
+  kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
+  tmp->cg_root = root_thread;
+  tmp->cg_thread_limit = __kmp_cg_max_nth;
+  tmp->cg_nthreads = 1;
+  KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"
+                 " cg_nthreads init to 1\n",
+                 root_thread, tmp));
+  tmp->up = NULL;
+  root_thread->th.th_cg_roots = tmp;
+
   __kmp_root_counter++;
 
 #if OMPT_SUPPORT
   if (!initial_thread && ompt_enabled.enabled) {
 
     kmp_info_t *root_thread = ompt_get_thread();
 
     ompt_set_thread_state(root_thread, ompt_state_overhead);
#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_cg_nthreads--;
+  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));
 
   __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;
if (!this_thr->th.th_pri_common) {
     if (__kmp_storage_map) {
       __kmp_print_storage_map_gtid(
           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
     }
     this_thr->th.th_pri_head = NULL;
   }
 
+  if (this_thr != master && // Master's CG root is initialized elsewhere
+      this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
+    // Make new thread's CG root same as master's
+    KMP_DEBUG_ASSERT(master->th.th_cg_roots);
+    this_thr->th.th_cg_roots = master->th.th_cg_roots;
+    // Increment new thread's CG root's counter to add the new thread
+    this_thr->th.th_cg_roots->cg_nthreads++;
+    KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
+                   " node %p of thread %p to %d\n",
+                   this_thr, this_thr->th.th_cg_roots,
+                   this_thr->th.th_cg_roots->cg_root,
+                   this_thr->th.th_cg_roots->cg_nthreads));
+    this_thr->th.th_current_task->td_icvs.thread_limit =
+        this_thr->th.th_cg_roots->cg_thread_limit;
+  }
+
   /* Initialize dynamic dispatch */
   {
     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
     // Use team max_nproc since this will never change for the team.
     size_t disp_size =
         sizeof(dispatch_private_info_t) *
         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
   KMP_DEBUG_ASSERT(root && team);
 #if !KMP_NESTED_HOT_TEAMS
   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
 #endif
   KMP_MB();
 
   /* first, try to get one from the thread pool */
   if (__kmp_thread_pool) {
-
     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) {
       __kmp_thread_pool_insert_pt = NULL;
     }
     TCW_4(new_thr->th.th_in_pool, FALSE);
     // Don't touch th_active_in_pool or th_active.
     // The worker thread adjusts those flags as it sleeps/awakens.
     __kmp_thread_pool_nth--;
 
     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
     KMP_ASSERT(!new_thr->th.th_team);
     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0);
 
     /* setup the thread structure */
     __kmp_initialize_info(new_thr, team, new_tid,
                           new_thr->th.th_info.ds.ds_gtid);
     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
 
     TCW_4(__kmp_nth, __kmp_nth + 1);
-    root->r.r_cg_nthreads++;
 
     new_thr->th.th_task_state = 0;
     new_thr->th.th_task_state_top = 0;
     new_thr->th.th_task_state_stack_sz = 4;
 
 #ifdef KMP_ADJUST_BLOCKTIME
     /* Adjust blocktime back to zero if necessary */
     /* Middle initialization might not have occurred yet */
#endif
   TCW_4(new_thr->th.th_in_pool, FALSE);
   new_thr->th.th_active_in_pool = FALSE;
   TCW_4(new_thr->th.th_active, TRUE);
 
   /* adjust the global counters */
   __kmp_all_nth++;
   __kmp_nth++;
 
-  root->r.r_cg_nthreads++;
-
   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
   // numbers of procs, and method #2 (keyed API call) for higher numbers.
   if (__kmp_adjust_gtid_mode) {
     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
       if (TCR_4(__kmp_gtid_mode) != 2) {
         TCW_4(__kmp_gtid_mode, 2);
       }
     } else {
if (level < __kmp_hot_teams_max_level && hot_teams &&
       use_hot_team = 1;
     } else {
       use_hot_team = 0;
     }
   }
 #endif
   // Optimization to use a "hot" team
   if (use_hot_team && new_nproc > 1) {
-    KMP_DEBUG_ASSERT(new_nproc == max_nproc);
+    KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
 #if KMP_NESTED_HOT_TEAMS
     team = hot_teams[level].hot_team;
 #else
     team = root->r.r_hot_team;
 #endif
 #if KMP_DEBUG
     if (__kmp_tasking_mode != tskm_immediate_exec) {
       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
#if KMP_NESTED_HOT_TEAMS
       }
 #endif // KMP_NESTED_HOT_TEAMS
       team->t.t_nproc = new_nproc;
       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
       __kmp_reinitialize_team(team, new_icvs,
                               root->r.r_uber_thread->th.th_ident);
 
-      /* update the remaining threads */
+      // Update remaining threads
       for (f = 0; f < new_nproc; ++f) {
         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
       }
+
       // restore the current task state of the master thread: should be the
       // implicit task
       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
                     team->t.t_threads[0], team));
 
       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
 
 #ifdef KMP_DEBUG
#endif // KMP_NESTED_HOT_TEAMS
       __kmp_initialize_team(team, new_nproc, new_icvs,
                             root->r.r_uber_thread->th.th_ident);
 
       /* reinitialize the threads */
       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
       for (f = 0; f < team->t.t_nproc; ++f)
         __kmp_initialize_info(team->t.t_threads[f], team, f,
                               __kmp_gtid_from_tid(f, team));
+
       if (level) { // set th_task_state for new threads in nested hot team
         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
         // only need to set the th_task_state for the new threads. th_task_state
         // for master thread will not be accurate until after this in
         // __kmp_fork_call(), so we look to the master's memo_stack to get the
         // correct value.
         for (f = old_nproc; f < team->t.t_nproc; ++f)
           team->t.t_threads[f]->th.th_task_state =
#endif
         }
       }
 
       // Delete task teams
       int tt_idx;
       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
         if (task_team != NULL) {
-          for (f = 0; f < team->t.t_nproc;
-               ++f) { // Have all threads unref task teams
+          for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
+            KMP_DEBUG_ASSERT(team->t.t_threads[f]);
             team->t.t_threads[f]->th.th_task_team = NULL;
           }
           KA_TRACE(
               20,
               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
                __kmp_get_gtid(), task_team, team->t.t_id));
 #if KMP_NESTED_HOT_TEAMS
           __kmp_free_task_team(master, task_team);
for (f = 1; f < team->t.t_nproc; ++f) {
       __kmp_free_thread(team->t.t_threads[f]);
       team->t.t_threads[f] = NULL;
     }
 
     /* put the team back in the team pool */
     /* TODO limit size of team pool, call reap_team if pool too large */
     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
     __kmp_team_pool = (volatile kmp_team_t *)team;
+  } else { // Check if team was created for the masters in a teams construct
+    // See if first worker is a CG root
+    KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
+                     team->t.t_threads[1]->th.th_cg_roots);
+    if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
+      // Clean up the CG root nodes on workers so that this team can be re-used
+      for (f = 1; f < team->t.t_nproc; ++f) {
+        kmp_info_t *thr = team->t.t_threads[f];
+        KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
+                         thr->th.th_cg_roots->cg_root == thr);
+        // Pop current CG root off list
+        kmp_cg_root_t *tmp = thr->th.th_cg_roots;
+        thr->th.th_cg_roots = tmp->up;
+        KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
+                       " up to node %p. cg_nthreads was %d\n",
+                       thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
+        __kmp_free(tmp);
+        // Restore current task's thread_limit from CG root
+        if (thr->th.th_cg_roots)
+          thr->th.th_current_task->td_icvs.thread_limit =
+              thr->th.th_cg_roots->cg_thread_limit;
+      }
+    }
   }
 
   KMP_MB();
 }
 
 /* reap the team.  destroy it, reclaim all its resources and free its memory */
 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
   kmp_team_t *next_pool = team->t.t_next_pool;

 // back on the hot team in reverse order.  This could cause bad cache
 // locality problems on programs where the size of the hot team regularly
 // grew and shrunk.
 //
 // Now, for single-level parallelism, the OMP tid is alway == gtid.
 void __kmp_free_thread(kmp_info_t *this_th) {
   int gtid;
   kmp_info_t **scan;
-  kmp_root_t *root = this_th->th.th_root;
 
   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
 
   KMP_DEBUG_ASSERT(this_th);
 
   // When moving thread to pool, switch thread to wait on own b_go flag, and
   // uninitialized (NULL team).
   int b;
   kmp_balign_t *balign = this_th->th.th_bar;
   for (b = 0; b < bs_last_barrier; ++b) {
     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
     balign[b].bb.team = NULL;
     balign[b].bb.leaf_kids = 0;
   }
   this_th->th.th_task_state = 0;
   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
 
   /* put thread back on the free pool */
   TCW_PTR(this_th->th.th_team, NULL);
   TCW_PTR(this_th->th.th_root, NULL);
   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
 
+  while (this_th->th.th_cg_roots) {
+    this_th->th.th_cg_roots->cg_nthreads--;
+    KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
+                   " %p of thread  %p to %d\n",
+                   this_th, this_th->th.th_cg_roots,
+                   this_th->th.th_cg_roots->cg_root,
+                   this_th->th.th_cg_roots->cg_nthreads));
+    kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
+    if (tmp->cg_root == this_th) { // Thread is a cg_root
+      KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
+      KA_TRACE(
+          5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
+      this_th->th.th_cg_roots = tmp->up;
+      __kmp_free(tmp);
+    } else { // Worker thread
+      this_th->th.th_cg_roots = NULL;
+      break;
+    }
+  }
+
   /* If the implicit task assigned to this thread can be used by other threads
    * -> multiple threads can share the data and try to free the task at
    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
    * with higher probability when hot team is disabled but can occurs even when
    * the hot team is enabled */
   __kmp_free_implicit_task(this_th);
   this_th->th.th_current_task = NULL;
 
void __kmp_free_thread(kmp_info_t *this_th) {
   __kmp_thread_pool_insert_pt = *scan = this_th;
   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
                    (this_th->th.th_info.ds.ds_gtid <
                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
   TCW_4(this_th->th.th_in_pool, TRUE);
   __kmp_thread_pool_nth++;
 
   TCW_4(__kmp_nth, __kmp_nth - 1);
-  root->r.r_cg_nthreads--;
 
 #ifdef KMP_ADJUST_BLOCKTIME
   /* Adjust blocktime back to user setting or default if necessary */
   /* Middle initialization might never have occurred                */
   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
     if (__kmp_nth <= __kmp_avail_proc) {
       __kmp_zero_bt = FALSE;
void __kmp_teams_master(int gtid) {
   kmp_info_t *thr = __kmp_threads[gtid];
   kmp_team_t *team = thr->th.th_team;
   ident_t *loc = team->t.t_ident;
   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
+
+  // This thread is a new CG root.  Set up the proper variables.
+  kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
+  tmp->cg_root = thr; // Make thr the CG root
+  // Init to thread limit that was stored when league masters were forked
+  tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
+  tmp->cg_nthreads = 1; // Init counter to one active thread, this one
+  KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
+                 " cg_threads to 1\n",
+                 thr, tmp));
+  tmp->up = thr->th.th_cg_roots;
+  thr->th.th_cg_roots = tmp;
+
 // Launch league of teams now, but not let workers execute
 // (they hang on fork barrier until next parallel)
 #if INCLUDE_SSC_MARKS
   SSC_MARK_FORKING();
 #endif
   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
 #if INCLUDE_SSC_MARKS
   SSC_MARK_JOINING();
 #endif
-
+  // If the team size was reduced from the limit, set it to the new size
+  if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
+    thr->th.th_teams_size.nth = thr->th.th_team_nproc;
   // AC: last parameter "1" eliminates join barrier which won't work because
   // worker threads are in a fork barrier waiting for more parallel regions
   __kmp_join_call(loc, gtid
 #if OMPT_SUPPORT
                   ,
                   fork_context_intel
 #endif
                   ,
if (num_threads == 0) {
     if (!TCR_4(__kmp_init_middle))
       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
     num_threads = __kmp_avail_proc / num_teams;
     if (num_teams * num_threads > __kmp_teams_max_nth) {
       // adjust num_threads w/o warning as it is not user setting
       num_threads = __kmp_teams_max_nth / num_teams;
     }
   } else {
+    // This thread will be the master of the league masters
+    // Store new thread limit; old limit is saved in th_cg_roots list
+    thr->th.th_current_task->td_icvs.thread_limit = num_threads;
+
     if (num_teams * num_threads > __kmp_teams_max_nth) {
       int new_threads = __kmp_teams_max_nth / num_teams;
       if (!__kmp_reserve_warn) { // user asked for too many threads
-        __kmp_reserve_warn = 1; // that conflicts with KMP_TEAMS_THREAD_LIMIT
+        __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
         __kmp_msg(kmp_ms_warning,
                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
       }
       num_threads = new_threads;
     }
   }
   thr->th.th_teams_size.nth = num_threads;