base/SOURCES/gcc48-libgomp-20160715.patch

--- libgomp/config/linux/wait.h.jj	2013-01-31 20:29:10.091548989 +0100
+++ libgomp/config/linux/wait.h	2016-07-13 16:57:18.902355979 +0200
@@ -34,13 +34,13 @@

 #define FUTEX_WAIT	0
 #define FUTEX_WAKE	1
-#define FUTEX_PRIVATE_FLAG	128L
+#define FUTEX_PRIVATE_FLAG	128

 #ifdef HAVE_ATTRIBUTE_VISIBILITY
 # pragma GCC visibility push(hidden)
 #endif

-extern long int gomp_futex_wait, gomp_futex_wake;
+extern int gomp_futex_wait, gomp_futex_wake;

 #include <futex.h>

@@ -48,7 +48,9 @@ static inline int do_spin (int *addr, in
 {
   unsigned long long i, count = gomp_spin_count_var;

-  if (__builtin_expect (gomp_managed_threads > gomp_available_cpus, 0))
+  if (__builtin_expect (__atomic_load_n (&gomp_managed_threads,
+                                         MEMMODEL_RELAXED)
+                        > gomp_available_cpus, 0))
     count = gomp_throttled_spin_count_var;
   for (i = 0; i < count; i++)
     if (__builtin_expect (__atomic_load_n (addr, MEMMODEL_RELAXED) != val, 0))
--- libgomp/config/linux/affinity.c.jj	2014-05-15 10:56:37.499502573 +0200
+++ libgomp/config/linux/affinity.c	2016-07-13 16:57:18.902355979 +0200
@@ -352,6 +352,45 @@ gomp_affinity_print_place (void *p)
     fprintf (stderr, ":%lu", len);
 }

+int
+omp_get_place_num_procs (int place_num)
+{
+  if (place_num < 0 || place_num >= gomp_places_list_len)
+    return 0;
+
+  cpu_set_t *cpusetp = (cpu_set_t *) gomp_places_list[place_num];
+  return gomp_cpuset_popcount (gomp_cpuset_size, cpusetp);
+}
+
+void
+omp_get_place_proc_ids (int place_num, int *ids)
+{
+  if (place_num < 0 || place_num >= gomp_places_list_len)
+    return;
+
+  cpu_set_t *cpusetp = (cpu_set_t *) gomp_places_list[place_num];
+  unsigned long i, max = 8 * gomp_cpuset_size;
+  for (i = 0; i < max; i++)
+    if (CPU_ISSET_S (i, gomp_cpuset_size, cpusetp))
+      *ids++ = i;
+}
+
+void
+gomp_get_place_proc_ids_8 (int place_num, int64_t *ids)
+{
+  if (place_num < 0 || place_num >= gomp_places_list_len)
+    return;
+
+  cpu_set_t *cpusetp = (cpu_set_t *) gomp_places_list[place_num];
+  unsigned long i, max = 8 * gomp_cpuset_size;
+  for (i = 0; i < max; i++)
+    if (CPU_ISSET_S (i, gomp_cpuset_size, cpusetp))
+      *ids++ = i;
+}
+
+ialias(omp_get_place_num_procs)
+ialias(omp_get_place_proc_ids)
+
 #else

 #include "../posix/affinity.c"
--- libgomp/config/linux/mutex.c.jj	2013-01-21 16:00:38.220917670 +0100
+++ libgomp/config/linux/mutex.c	2016-07-13 16:57:18.870356375 +0200
@@ -28,8 +28,8 @@

 #include "wait.h"

-long int gomp_futex_wake = FUTEX_WAKE | FUTEX_PRIVATE_FLAG;
-long int gomp_futex_wait = FUTEX_WAIT | FUTEX_PRIVATE_FLAG;
+int gomp_futex_wake = FUTEX_WAKE | FUTEX_PRIVATE_FLAG;
+int gomp_futex_wait = FUTEX_WAIT | FUTEX_PRIVATE_FLAG;

 void
 gomp_mutex_lock_slow (gomp_mutex_t *mutex, int oldval)
--- libgomp/config/posix/affinity.c.jj	2014-05-15 10:56:37.987498844 +0200
+++ libgomp/config/posix/affinity.c	2016-07-15 12:08:28.410015743 +0200
@@ -113,3 +113,27 @@ gomp_affinity_print_place (void *p)
 {
   (void) p;
 }
+
+int
+omp_get_place_num_procs (int place_num)
+{
+  (void) place_num;
+  return 0;
+}
+
+void
+omp_get_place_proc_ids (int place_num, int *ids)
+{
+  (void) place_num;
+  (void) ids;
+}
+
+void
+gomp_get_place_proc_ids_8 (int place_num, int64_t *ids)
+{
+  (void) place_num;
+  (void) ids;
+}
+
+ialias(omp_get_place_num_procs)
+ialias(omp_get_place_proc_ids)
--- libgomp/loop_ull.c.jj	2013-01-21 16:00:46.477871806 +0100
+++ libgomp/loop_ull.c	2016-07-13 16:57:18.918355780 +0200
@@ -174,15 +174,15 @@ GOMP_loop_ull_runtime_start (bool up, go
     {
     case GFS_STATIC:
       return gomp_loop_ull_static_start (up, start, end, incr,
-					 icv->run_sched_modifier,
+					 icv->run_sched_chunk_size,
					 istart, iend);
     case GFS_DYNAMIC:
       return gomp_loop_ull_dynamic_start (up, start, end, incr,
-					  icv->run_sched_modifier,
+					  icv->run_sched_chunk_size,
					  istart, iend);
     case GFS_GUIDED:
       return gomp_loop_ull_guided_start (up, start, end, incr,
-					 icv->run_sched_modifier,
+					 icv->run_sched_chunk_size,
					 istart, iend);
     case GFS_AUTO:
       /* For now map to schedule(static), later on we could play with feedback
@@ -278,15 +278,15 @@ GOMP_loop_ull_ordered_runtime_start (boo
     {
     case GFS_STATIC:
       return gomp_loop_ull_ordered_static_start (up, start, end, incr,
-						 icv->run_sched_modifier,
+						 icv->run_sched_chunk_size,
						 istart, iend);
     case GFS_DYNAMIC:
       return gomp_loop_ull_ordered_dynamic_start (up, start, end, incr,
-						  icv->run_sched_modifier,
+						  icv->run_sched_chunk_size,
						  istart, iend);
     case GFS_GUIDED:
       return gomp_loop_ull_ordered_guided_start (up, start, end, incr,
-						 icv->run_sched_modifier,
+						 icv->run_sched_chunk_size,
						 istart, iend);
     case GFS_AUTO:
       /* For now map to schedule(static), later on we could play with feedback
@@ -298,6 +298,114 @@ GOMP_loop_ull_ordered_runtime_start (boo
     }
 }

+/* The *_doacross_*_start routines are similar.  The only difference is that
+   this work-share construct is initialized to expect an ORDERED(N) - DOACROSS
+   section, and the worksharing loop iterates always from 0 to COUNTS[0] - 1
+   and other COUNTS array elements tell the library number of iterations
+   in the ordered inner loops.  */
+
+static bool
+gomp_loop_ull_doacross_static_start (unsigned ncounts, gomp_ull *counts,
+				     gomp_ull chunk_size, gomp_ull *istart,
+				     gomp_ull *iend)
+{
+  struct gomp_thread *thr = gomp_thread ();
+
+  thr->ts.static_trip = 0;
+  if (gomp_work_share_start (false))
+    {
+      gomp_loop_ull_init (thr->ts.work_share, true, 0, counts[0], 1,
+			  GFS_STATIC, chunk_size);
+      gomp_doacross_ull_init (ncounts, counts, chunk_size);
+      gomp_work_share_init_done ();
+    }
+
+  return !gomp_iter_ull_static_next (istart, iend);
+}
+
+static bool
+gomp_loop_ull_doacross_dynamic_start (unsigned ncounts, gomp_ull *counts,
+				      gomp_ull chunk_size, gomp_ull *istart,
+				      gomp_ull *iend)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  bool ret;
+
+  if (gomp_work_share_start (false))
+    {
+      gomp_loop_ull_init (thr->ts.work_share, true, 0, counts[0], 1,
+			  GFS_DYNAMIC, chunk_size);
+      gomp_doacross_ull_init (ncounts, counts, chunk_size);
+      gomp_work_share_init_done ();
+    }
+
+#if defined HAVE_SYNC_BUILTINS && defined __LP64__
+  ret = gomp_iter_ull_dynamic_next (istart, iend);
+#else
+  gomp_mutex_lock (&thr->ts.work_share->lock);
+  ret = gomp_iter_ull_dynamic_next_locked (istart, iend);
+  gomp_mutex_unlock (&thr->ts.work_share->lock);
+#endif
+
+  return ret;
+}
+
+static bool
+gomp_loop_ull_doacross_guided_start (unsigned ncounts, gomp_ull *counts,
+				     gomp_ull chunk_size, gomp_ull *istart,
+				     gomp_ull *iend)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  bool ret;
+
+  if (gomp_work_share_start (false))
+    {
+      gomp_loop_ull_init (thr->ts.work_share, true, 0, counts[0], 1,
+			  GFS_GUIDED, chunk_size);
+      gomp_doacross_ull_init (ncounts, counts, chunk_size);
+      gomp_work_share_init_done ();
+    }
+
+#if defined HAVE_SYNC_BUILTINS && defined __LP64__
+  ret = gomp_iter_ull_guided_next (istart, iend);
+#else
+  gomp_mutex_lock (&thr->ts.work_share->lock);
+  ret = gomp_iter_ull_guided_next_locked (istart, iend);
+  gomp_mutex_unlock (&thr->ts.work_share->lock);
+#endif
+
+  return ret;
+}
+
+bool
+GOMP_loop_ull_doacross_runtime_start (unsigned ncounts, gomp_ull *counts,
+				      gomp_ull *istart, gomp_ull *iend)
+{
+  struct gomp_task_icv *icv = gomp_icv (false);
+  switch (icv->run_sched_var)
+    {
+    case GFS_STATIC:
+      return gomp_loop_ull_doacross_static_start (ncounts, counts,
+						  icv->run_sched_chunk_size,
+						  istart, iend);
+    case GFS_DYNAMIC:
+      return gomp_loop_ull_doacross_dynamic_start (ncounts, counts,
+						   icv->run_sched_chunk_size,
+						   istart, iend);
+    case GFS_GUIDED:
+      return gomp_loop_ull_doacross_guided_start (ncounts, counts,
+						  icv->run_sched_chunk_size,
+						  istart, iend);
+    case GFS_AUTO:
+      /* For now map to schedule(static), later on we could play with feedback
+	 driven choice.  */
+      return gomp_loop_ull_doacross_static_start (ncounts, counts,
+						  0, istart, iend);
+    default:
+      abort ();
+    }
+}
+
 /* The *_next routines are called when the thread completes processing of
    the iteration block currently assigned to it.  If the work-share
    construct is bound directly to a parallel construct, then the iteration
@@ -457,6 +565,10 @@ extern __typeof(gomp_loop_ull_dynamic_st
	__attribute__((alias ("gomp_loop_ull_dynamic_start")));
 extern __typeof(gomp_loop_ull_guided_start) GOMP_loop_ull_guided_start
	__attribute__((alias ("gomp_loop_ull_guided_start")));
+extern __typeof(gomp_loop_ull_dynamic_start) GOMP_loop_ull_nonmonotonic_dynamic_start
+	__attribute__((alias ("gomp_loop_ull_dynamic_start")));
+extern __typeof(gomp_loop_ull_guided_start) GOMP_loop_ull_nonmonotonic_guided_start
+	__attribute__((alias ("gomp_loop_ull_guided_start")));

 extern __typeof(gomp_loop_ull_ordered_static_start) GOMP_loop_ull_ordered_static_start
	__attribute__((alias ("gomp_loop_ull_ordered_static_start")));
@@ -465,12 +577,23 @@ extern __typeof(gomp_loop_ull_ordered_dy
 extern __typeof(gomp_loop_ull_ordered_guided_start) GOMP_loop_ull_ordered_guided_start
	__attribute__((alias ("gomp_loop_ull_ordered_guided_start")));

+extern __typeof(gomp_loop_ull_doacross_static_start) GOMP_loop_ull_doacross_static_start
+	__attribute__((alias ("gomp_loop_ull_doacross_static_start")));
+extern __typeof(gomp_loop_ull_doacross_dynamic_start) GOMP_loop_ull_doacross_dynamic_start
+	__attribute__((alias ("gomp_loop_ull_doacross_dynamic_start")));
+extern __typeof(gomp_loop_ull_doacross_guided_start) GOMP_loop_ull_doacross_guided_start
+	__attribute__((alias ("gomp_loop_ull_doacross_guided_start")));
+
 extern __typeof(gomp_loop_ull_static_next) GOMP_loop_ull_static_next
	__attribute__((alias ("gomp_loop_ull_static_next")));
 extern __typeof(gomp_loop_ull_dynamic_next) GOMP_loop_ull_dynamic_next
	__attribute__((alias ("gomp_loop_ull_dynamic_next")));
 extern __typeof(gomp_loop_ull_guided_next) GOMP_loop_ull_guided_next
	__attribute__((alias ("gomp_loop_ull_guided_next")));
+extern __typeof(gomp_loop_ull_dynamic_next) GOMP_loop_ull_nonmonotonic_dynamic_next
+	__attribute__((alias ("gomp_loop_ull_dynamic_next")));
+extern __typeof(gomp_loop_ull_guided_next) GOMP_loop_ull_nonmonotonic_guided_next
+	__attribute__((alias ("gomp_loop_ull_guided_next")));

 extern __typeof(gomp_loop_ull_ordered_static_next) GOMP_loop_ull_ordered_static_next
	__attribute__((alias ("gomp_loop_ull_ordered_static_next")));
@@ -507,6 +630,25 @@ GOMP_loop_ull_guided_start (bool up, gom
 }

 bool
+GOMP_loop_ull_nonmonotonic_dynamic_start (bool up, gomp_ull start,
+					  gomp_ull end, gomp_ull incr,
+					  gomp_ull chunk_size,
+					  gomp_ull *istart, gomp_ull *iend)
+{
+  return gomp_loop_ull_dynamic_start (up, start, end, incr, chunk_size, istart,
+				      iend);
+}
+
+bool
+GOMP_loop_ull_nonmonotonic_guided_start (bool up, gomp_ull start, gomp_ull end,
+					 gomp_ull incr, gomp_ull chunk_size,
+					 gomp_ull *istart, gomp_ull *iend)
+{
+  return gomp_loop_ull_guided_start (up, start, end, incr, chunk_size, istart,
+				     iend);
+}
+
+bool
 GOMP_loop_ull_ordered_static_start (bool up, gomp_ull start, gomp_ull end,
				    gomp_ull incr, gomp_ull chunk_size,
				    gomp_ull *istart, gomp_ull *iend)
@@ -534,6 +676,33 @@ GOMP_loop_ull_ordered_guided_start (bool
 }

 bool
+GOMP_loop_ull_doacross_static_start (unsigned ncounts, gomp_ull *counts,
+				     gomp_ull chunk_size, gomp_ull *istart,
+				     gomp_ull *iend)
+{
+  return gomp_loop_ull_doacross_static_start (ncounts, counts, chunk_size,
+					      istart, iend);
+}
+
+bool
+GOMP_loop_ull_doacross_dynamic_start (unsigned ncounts, gomp_ull *counts,
+				      gomp_ull chunk_size, gomp_ull *istart,
+				      gomp_ull *iend)
+{
+  return gomp_loop_ull_doacross_dynamic_start (ncounts, counts, chunk_size,
+					       istart, iend);
+}
+
+bool
+GOMP_loop_ull_doacross_guided_start (unsigned ncounts, gomp_ull *counts,
+				     gomp_ull chunk_size, gomp_ull *istart,
+				     gomp_ull *iend)
+{
+  return gomp_loop_ull_doacross_guided_start (ncounts, counts, chunk_size,
+					      istart, iend);
+}
+
+bool
 GOMP_loop_ull_static_next (gomp_ull *istart, gomp_ull *iend)
 {
   return gomp_loop_ull_static_next (istart, iend);
@@ -550,6 +719,18 @@ GOMP_loop_ull_guided_next (gomp_ull *ist
 {
   return gomp_loop_ull_guided_next (istart, iend);
 }
+
+bool
+GOMP_loop_ull_nonmonotonic_dynamic_next (gomp_ull *istart, gomp_ull *iend)
+{
+  return gomp_loop_ull_dynamic_next (istart, iend);
+}
+
+bool
+GOMP_loop_ull_nonmonotonic_guided_next (gomp_ull *istart, gomp_ull *iend)
+{
+  return gomp_loop_ull_guided_next (istart, iend);
+}

 bool
 GOMP_loop_ull_ordered_static_next (gomp_ull *istart, gomp_ull *iend)
--- libgomp/team.c.jj	2014-05-15 10:56:32.092524669 +0200
+++ libgomp/team.c	2016-07-13 17:58:01.907291111 +0200
@@ -133,6 +133,25 @@ gomp_thread_start (void *xdata)
   return NULL;
 }

+static inline struct gomp_team *
+get_last_team (unsigned nthreads)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  if (thr->ts.team == NULL)
+    {
+      struct gomp_thread_pool *pool = thr->thread_pool;
+      if (pool != NULL)
+	{
+	  struct gomp_team *last_team = pool->last_team;
+	  if (last_team != NULL && last_team->nthreads == nthreads)
+	    {
+	      pool->last_team = NULL;
+	      return last_team;
+	    }
+	}
+    }
+  return NULL;
+}

 /* Create a new team data structure.  */

@@ -140,18 +159,27 @@ struct gomp_team *
 gomp_new_team (unsigned nthreads)
 {
   struct gomp_team *team;
-  size_t size;
   int i;

-  size = sizeof (*team) + nthreads * (sizeof (team->ordered_release[0])
-				      + sizeof (team->implicit_task[0]));
-  team = gomp_malloc (size);
+  team = get_last_team (nthreads);
+  if (team == NULL)
+    {
+      size_t extra = sizeof (team->ordered_release[0])
+		     + sizeof (team->implicit_task[0]);
+      team = gomp_malloc (sizeof (*team) + nthreads * extra);
+
+#ifndef HAVE_SYNC_BUILTINS
+      gomp_mutex_init (&team->work_share_list_free_lock);
+#endif
+      gomp_barrier_init (&team->barrier, nthreads);
+      gomp_mutex_init (&team->task_lock);
+
+      team->nthreads = nthreads;
+    }

   team->work_share_chunk = 8;
 #ifdef HAVE_SYNC_BUILTINS
   team->single_count = 0;
-#else
-  gomp_mutex_init (&team->work_share_list_free_lock);
 #endif
   team->work_shares_to_free = &team->work_shares[0];
   gomp_init_work_share (&team->work_shares[0], false, nthreads);
@@ -162,15 +190,11 @@ gomp_new_team (unsigned nthreads)
     team->work_shares[i].next_free = &team->work_shares[i + 1];
   team->work_shares[i].next_free = NULL;

-  team->nthreads = nthreads;
-  gomp_barrier_init (&team->barrier, nthreads);
-
   gomp_sem_init (&team->master_release, 0);
   team->ordered_release = (void *) &team->implicit_task[nthreads];
   team->ordered_release[0] = &team->master_release;

-  gomp_mutex_init (&team->task_lock);
-  team->task_queue = NULL;
+  priority_queue_init (&team->task_queue);
   team->task_count = 0;
   team->task_queued_count = 0;
   team->task_running_count = 0;
@@ -186,8 +210,12 @@ gomp_new_team (unsigned nthreads)
 static void
 free_team (struct gomp_team *team)
 {
+#ifndef HAVE_SYNC_BUILTINS
+  gomp_mutex_destroy (&team->work_share_list_free_lock);
+#endif
   gomp_barrier_destroy (&team->barrier);
   gomp_mutex_destroy (&team->task_lock);
+  priority_queue_free (&team->task_queue);
   free (team);
 }

@@ -258,6 +286,8 @@ gomp_free_thread (void *arg __attribute_
       free (pool);
       thr->thread_pool = NULL;
     }
+  if (thr->ts.level == 0 && __builtin_expect (thr->ts.team != NULL, 0))
+    gomp_team_end ();
   if (thr->task != NULL)
     {
       struct gomp_task *task = thr->task;
@@ -287,7 +317,7 @@ gomp_team_start (void (*fn) (void *), vo
   struct gomp_thread **affinity_thr = NULL;

   thr = gomp_thread ();
-  nested = thr->ts.team != NULL;
+  nested = thr->ts.level;
   if (__builtin_expect (thr->thread_pool == NULL, 0))
     {
       thr->thread_pool = gomp_new_thread_pool ();
@@ -894,9 +924,6 @@ gomp_team_end (void)
       while (ws != NULL);
     }
   gomp_sem_destroy (&team->master_release);
-#ifndef HAVE_SYNC_BUILTINS
-  gomp_mutex_destroy (&team->work_share_list_free_lock);
-#endif

   if (__builtin_expect (thr->ts.team != NULL, 0)
       || __builtin_expect (team->nthreads == 1, 0))
--- libgomp/target.c.jj	2014-05-15 10:56:38.313498020 +0200
+++ libgomp/target.c	2016-07-15 16:58:29.249328861 +0200
@@ -22,14 +22,22 @@
    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
    <http://www.gnu.org/licenses/>.  */

-/* This file handles the maintainence of threads in response to team
-   creation and termination.  */
+/* This file contains the support of offloading.  */

+#include "config.h"
 #include "libgomp.h"
+#include "oacc-plugin.h"
+#include "oacc-int.h"
+#include "gomp-constants.h"
 #include <limits.h>
 #include <stdbool.h>
 #include <stdlib.h>
+#ifdef HAVE_INTTYPES_H
+# include <inttypes.h>  /* For PRIu64.  */
+#endif
 #include <string.h>
+#include <assert.h>
+#include <errno.h>

 attribute_hidden int
 gomp_get_num_devices (void)
@@ -37,22 +45,87 @@ gomp_get_num_devices (void)
   return 0;
 }

-/* Called when encountering a target directive.  If DEVICE
-   is -1, it means use device-var ICV.  If it is -2 (or any other value
-   larger than last available hw device, use host fallback.
-   FN is address of host code, OPENMP_TARGET contains value of the
-   __OPENMP_TARGET__ symbol in the shared library or binary that invokes
-   GOMP_target.  HOSTADDRS, SIZES and KINDS are arrays
-   with MAPNUM entries, with addresses of the host objects,
-   sizes of the host objects (resp. for pointer kind pointer bias
-   and assumed sizeof (void *) size) and kinds.  */
+/* This function should be called from every offload image while loading.
+   It gets the descriptor of the host func and var tables HOST_TABLE, TYPE of
+   the target, and TARGET_DATA needed by target plugin.  */

 void
-GOMP_target (int device, void (*fn) (void *), const void *openmp_target,
-	     size_t mapnum, void **hostaddrs, size_t *sizes,
-	     unsigned char *kinds)
+GOMP_offload_register_ver (unsigned version, const void *host_table,
+			   int target_type, const void *target_data)
+{
+  (void) version;
+  (void) host_table;
+  (void) target_type;
+  (void) target_data;
+}
+
+void
+GOMP_offload_register (const void *host_table, int target_type,
+		       const void *target_data)
+{
+  (void) host_table;
+  (void) target_type;
+  (void) target_data;
+}
+
+/* This function should be called from every offload image while unloading.
+   It gets the descriptor of the host func and var tables HOST_TABLE, TYPE of
+   the target, and TARGET_DATA needed by target plugin.  */
+
+void
+GOMP_offload_unregister_ver (unsigned version, const void *host_table,
+			     int target_type, const void *target_data)
+{
+  (void) version;
+  (void) host_table;
+  (void) target_type;
+  (void) target_data;
+}
+
+void
+GOMP_offload_unregister (const void *host_table, int target_type,
+			 const void *target_data)
+{
+  (void) host_table;
+  (void) target_type;
+  (void) target_data;
+}
+
+/* This function initializes the target device, specified by DEVICEP.  DEVICEP
+   must be locked on entry, and remains locked on return.  */
+
+attribute_hidden void
+gomp_init_device (struct gomp_device_descr *devicep)
+{
+  devicep->state = GOMP_DEVICE_INITIALIZED;
+}
+
+attribute_hidden void
+gomp_unload_device (struct gomp_device_descr *devicep)
+{
+}
+
+/* Free address mapping tables.  MM must be locked on entry, and remains locked
+   on return.  */
+
+attribute_hidden void
+gomp_free_memmap (struct splay_tree_s *mem_map)
+{
+  while (mem_map->root)
+    {
+      struct target_mem_desc *tgt = mem_map->root->key.tgt;
+
+      splay_tree_remove (mem_map, &mem_map->root->key);
+      free (tgt->array);
+      free (tgt);
+    }
+}
+
+/* Host fallback for GOMP_target{,_ext} routines.  */
+
+static void
+gomp_target_fallback (void (*fn) (void *), void **hostaddrs)
 {
-  /* Host fallback.  */
   struct gomp_thread old_thr, *thr = gomp_thread ();
   old_thr = *thr;
   memset (thr, '\0', sizeof (*thr));
@@ -66,10 +139,167 @@ GOMP_target (int device, void (*fn) (voi
   *thr = old_thr;
 }

+/* Calculate alignment and size requirements of a private copy of data shared
+   as GOMP_MAP_FIRSTPRIVATE and store them to TGT_ALIGN and TGT_SIZE.  */
+
+static inline void
+calculate_firstprivate_requirements (size_t mapnum, size_t *sizes,
+				     unsigned short *kinds, size_t *tgt_align,
+				     size_t *tgt_size)
+{
+  size_t i;
+  for (i = 0; i < mapnum; i++)
+    if ((kinds[i] & 0xff) == GOMP_MAP_FIRSTPRIVATE)
+      {
+	size_t align = (size_t) 1 << (kinds[i] >> 8);
+	if (*tgt_align < align)
+	  *tgt_align = align;
+	*tgt_size = (*tgt_size + align - 1) & ~(align - 1);
+	*tgt_size += sizes[i];
+      }
+}
+
+/* Copy data shared as GOMP_MAP_FIRSTPRIVATE to DST.  */
+
+static inline void
+copy_firstprivate_data (char *tgt, size_t mapnum, void **hostaddrs,
+			size_t *sizes, unsigned short *kinds, size_t tgt_align,
+			size_t tgt_size)
+{
+  uintptr_t al = (uintptr_t) tgt & (tgt_align - 1);
+  if (al)
+    tgt += tgt_align - al;
+  tgt_size = 0;
+  size_t i;
+  for (i = 0; i < mapnum; i++)
+    if ((kinds[i] & 0xff) == GOMP_MAP_FIRSTPRIVATE)
+      {
+	size_t align = (size_t) 1 << (kinds[i] >> 8);
+	tgt_size = (tgt_size + align - 1) & ~(align - 1);
+	memcpy (tgt + tgt_size, hostaddrs[i], sizes[i]);
+	hostaddrs[i] = tgt + tgt_size;
+	tgt_size = tgt_size + sizes[i];
+      }
+}
+
+/* Called when encountering a target directive.  If DEVICE
+   is GOMP_DEVICE_ICV, it means use device-var ICV.  If it is
+   GOMP_DEVICE_HOST_FALLBACK (or any value
+   larger than last available hw device), use host fallback.
+   FN is address of host code, UNUSED is part of the current ABI, but
+   we're not actually using it.  HOSTADDRS, SIZES and KINDS are arrays
+   with MAPNUM entries, with addresses of the host objects,
+   sizes of the host objects (resp. for pointer kind pointer bias
+   and assumed sizeof (void *) size) and kinds.  */
+
+void
+GOMP_target (int device, void (*fn) (void *), const void *unused,
+	     size_t mapnum, void **hostaddrs, size_t *sizes,
+	     unsigned char *kinds)
+{
+  return gomp_target_fallback (fn, hostaddrs);
+}
+
+/* Like GOMP_target, but KINDS is 16-bit, UNUSED is no longer present,
+   and several arguments have been added:
+   FLAGS is a bitmask, see GOMP_TARGET_FLAG_* in gomp-constants.h.
+   DEPEND is array of dependencies, see GOMP_task for details.
+
+   ARGS is a pointer to an array consisting of a variable number of both
+   device-independent and device-specific arguments, which can take one two
+   elements where the first specifies for which device it is intended, the type
+   and optionally also the value.  If the value is not present in the first
+   one, the whole second element the actual value.  The last element of the
+   array is a single NULL.  Among the device independent can be for example
+   NUM_TEAMS and THREAD_LIMIT.
+
+   NUM_TEAMS is positive if GOMP_teams will be called in the body with
+   that value, or 1 if teams construct is not present, or 0, if
+   teams construct does not have num_teams clause and so the choice is
+   implementation defined, and -1 if it can't be determined on the host
+   what value will GOMP_teams have on the device.
+   THREAD_LIMIT similarly is positive if GOMP_teams will be called in the
+   body with that value, or 0, if teams construct does not have thread_limit
+   clause or the teams construct is not present, or -1 if it can't be
+   determined on the host what value will GOMP_teams have on the device.  */
+
+void
+GOMP_target_ext (int device, void (*fn) (void *), size_t mapnum,
+		 void **hostaddrs, size_t *sizes, unsigned short *kinds,
+		 unsigned int flags, void **depend, void **args)
+{
+  size_t tgt_align = 0, tgt_size = 0;
+  bool fpc_done = false;
+
+  if (flags & GOMP_TARGET_FLAG_NOWAIT)
+    {
+      struct gomp_thread *thr = gomp_thread ();
+      if (thr->ts.team
+	  && !thr->task->final_task)
+	{
+	  gomp_create_target_task (NULL, fn, mapnum, hostaddrs,
+				   sizes, kinds, flags, depend, args,
+				   GOMP_TARGET_TASK_BEFORE_MAP);
+	  return;
+	}
+    }
+
+  /* If there are depend clauses, but nowait is not present
+     (or we are in a final task), block the parent task until the
+     dependencies are resolved and then just continue with the rest
+     of the function as if it is a merged task.  */
+  if (depend != NULL)
+    {
+      struct gomp_thread *thr = gomp_thread ();
+      if (thr->task && thr->task->depend_hash)
+	{
+	  /* If we might need to wait, copy firstprivate now.  */
+	  calculate_firstprivate_requirements (mapnum, sizes, kinds,
+					       &tgt_align, &tgt_size);
+	  if (tgt_align)
+	    {
+	      char *tgt = gomp_alloca (tgt_size + tgt_align - 1);
+	      copy_firstprivate_data (tgt, mapnum, hostaddrs, sizes, kinds,
+				      tgt_align, tgt_size);
+	    }
+	  fpc_done = true;
+	  gomp_task_maybe_wait_for_dependencies (depend);
+	}
+    }
+
+  if (!fpc_done)
+    {
+      calculate_firstprivate_requirements (mapnum, sizes, kinds,
+					   &tgt_align, &tgt_size);
+      if (tgt_align)
+	{
+	  char *tgt = gomp_alloca (tgt_size + tgt_align - 1);
+	  copy_firstprivate_data (tgt, mapnum, hostaddrs, sizes, kinds,
+				  tgt_align, tgt_size);
+	}
+    }
+  gomp_target_fallback (fn, hostaddrs);
+}
+
+/* Host fallback for GOMP_target_data{,_ext} routines.  */
+
+static void
+gomp_target_data_fallback (void)
+{
+}
+
 void
-GOMP_target_data (int device, const void *openmp_target, size_t mapnum,
+GOMP_target_data (int device, const void *unused, size_t mapnum,
		  void **hostaddrs, size_t *sizes, unsigned char *kinds)
 {
+  return gomp_target_data_fallback ();
+}
+
+void
+GOMP_target_data_ext (int device, size_t mapnum, void **hostaddrs,
+		      size_t *sizes, unsigned short *kinds)
+{
+  return gomp_target_data_fallback ();
 }

 void
@@ -78,12 +308,112 @@ GOMP_target_end_data (void)
 }

 void
-GOMP_target_update (int device, const void *openmp_target, size_t mapnum,
+GOMP_target_update (int device, const void *unused, size_t mapnum,
		    void **hostaddrs, size_t *sizes, unsigned char *kinds)
 {
 }

 void
+GOMP_target_update_ext (int device, size_t mapnum, void **hostaddrs,
+			size_t *sizes, unsigned short *kinds,
+			unsigned int flags, void **depend)
+{
+  /* If there are depend clauses, but nowait is not present,
+     block the parent task until the dependencies are resolved
+     and then just continue with the rest of the function as if it
+     is a merged task.  Until we are able to schedule task during
+     variable mapping or unmapping, ignore nowait if depend clauses
+     are not present.  */
+  if (depend != NULL)
+    {
+      struct gomp_thread *thr = gomp_thread ();
+      if (thr->task && thr->task->depend_hash)
+	{
+	  if ((flags & GOMP_TARGET_FLAG_NOWAIT)
+	      && thr->ts.team
+	      && !thr->task->final_task)
+	    {
+	      if (gomp_create_target_task (NULL, (void (*) (void *)) NULL,
+					   mapnum, hostaddrs, sizes, kinds,
+					   flags | GOMP_TARGET_FLAG_UPDATE,
+					   depend, NULL, GOMP_TARGET_TASK_DATA))
+		return;
+	    }
+	  else
+	    {
+	      struct gomp_team *team = thr->ts.team;
+	      /* If parallel or taskgroup has been cancelled, don't start new
+		 tasks.  */
+	      if (team
+		  && (gomp_team_barrier_cancelled (&team->barrier)
+		      || (thr->task->taskgroup
+			  && thr->task->taskgroup->cancelled)))
+		return;
+
+	      gomp_task_maybe_wait_for_dependencies (depend);
+	    }
+	}
+    }
+}
+
+void
+GOMP_target_enter_exit_data (int device, size_t mapnum, void **hostaddrs,
+			     size_t *sizes, unsigned short *kinds,
+			     unsigned int flags, void **depend)
+{
+  /* If there are depend clauses, but nowait is not present,
+     block the parent task until the dependencies are resolved
+     and then just continue with the rest of the function as if it
+     is a merged task.  Until we are able to schedule task during
+     variable mapping or unmapping, ignore nowait if depend clauses
+     are not present.  */
+  if (depend != NULL)
+    {
+      struct gomp_thread *thr = gomp_thread ();
+      if (thr->task && thr->task->depend_hash)
+	{
+	  if ((flags & GOMP_TARGET_FLAG_NOWAIT)
+	      && thr->ts.team
+	      && !thr->task->final_task)
+	    {
+	      if (gomp_create_target_task (NULL, (void (*) (void *)) NULL,
+					   mapnum, hostaddrs, sizes, kinds,
+					   flags, depend, NULL,
+					   GOMP_TARGET_TASK_DATA))
+		return;
+	    }
+	  else
+	    {
+	      struct gomp_team *team = thr->ts.team;
+	      /* If parallel or taskgroup has been cancelled, don't start new
+		 tasks.  */
+	      if (team
+		  && (gomp_team_barrier_cancelled (&team->barrier)
+		      || (thr->task->taskgroup
+			  && thr->task->taskgroup->cancelled)))
+		return;
+
+	      gomp_task_maybe_wait_for_dependencies (depend);
+	    }
+	}
+    }
+}
+
+bool
+gomp_target_task_fn (void *data)
+{
+  struct gomp_target_task *ttask = (struct gomp_target_task *) data;
+
+  if (ttask->fn != NULL)
+    {
+      ttask->state = GOMP_TARGET_TASK_FALLBACK;
+      gomp_target_fallback (ttask->fn, ttask->hostaddrs);
+      return false;
+    }
+  return false;
+}
+
+void
 GOMP_teams (unsigned int num_teams, unsigned int thread_limit)
 {
   if (thread_limit)
@@ -94,3 +424,153 @@ GOMP_teams (unsigned int num_teams, unsi
     }
   (void) num_teams;
 }
+
+void *
+omp_target_alloc (size_t size, int device_num)
+{
+  if (device_num == GOMP_DEVICE_HOST_FALLBACK)
+    return malloc (size);
+
+  return NULL;
+}
+
+void
+omp_target_free (void *device_ptr, int device_num)
+{
+  if (device_ptr == NULL)
+    return;
+
+  if (device_num == GOMP_DEVICE_HOST_FALLBACK)
+    {
+      free (device_ptr);
+      return;
+    }
+}
+
+int
+omp_target_is_present (void *ptr, int device_num)
+{
+  if (ptr == NULL)
+    return 1;
+
+  if (device_num == GOMP_DEVICE_HOST_FALLBACK)
+    return 1;
+
+  return 0;
+}
+
+int
+omp_target_memcpy (void *dst, void *src, size_t length, size_t dst_offset,
+		   size_t src_offset, int dst_device_num, int src_device_num)
+{
+  if (dst_device_num != GOMP_DEVICE_HOST_FALLBACK)
+    return EINVAL;
+  if (src_device_num != GOMP_DEVICE_HOST_FALLBACK)
+    return EINVAL;
+  memcpy ((char *) dst + dst_offset, (char *) src + src_offset, length);
+  return 0;
+}
+
+#define HALF_SIZE_T (((size_t) 1) << (8 * sizeof (size_t) / 2))
+
+#define __builtin_mul_overflow(x, y, z) \
+  ({ bool retval = false;					\
+     size_t xval = (x);						\
+     size_t yval = (y);						\
+     size_t zval = xval * yval;					\
+     if (__builtin_expect ((xval | yval) >= HALF_SIZE_T, 0))	\
+       {							\
+         if (xval && zval / xval != yval)			\
+	   retval = true;					\
+       }							\
+     *(z) = zval;						\
+     retval; })
+
+static int
+omp_target_memcpy_rect_worker (void *dst, void *src, size_t element_size,
+			       int num_dims, const size_t *volume,
+			       const size_t *dst_offsets,
+			       const size_t *src_offsets,
+			       const size_t *dst_dimensions,
+			       const size_t *src_dimensions)
+{
+  size_t dst_slice = element_size;
+  size_t src_slice = element_size;
+  size_t j, dst_off, src_off, length;
+  int i, ret;
+
+
+  if (num_dims == 1)
+    {
+      if (__builtin_mul_overflow (element_size, volume[0], &length)
+	  || __builtin_mul_overflow (element_size, dst_offsets[0], &dst_off)
+	  || __builtin_mul_overflow (element_size, src_offsets[0], &src_off))
+	return EINVAL;
+      memcpy ((char *) dst + dst_off, (char *) src + src_off, length);
+      ret = 1;
+      return ret ? 0 : EINVAL;
+    }
+
+  /* FIXME: it would be nice to have some plugin function to handle
+     num_dims == 2 and num_dims == 3 more efficiently.  Larger ones can
+     be handled in the generic recursion below, and for host-host it
+     should be used even for any num_dims >= 2.  */
+
+  for (i = 1; i < num_dims; i++)
+    if (__builtin_mul_overflow (dst_slice, dst_dimensions[i], &dst_slice)
+	|| __builtin_mul_overflow (src_slice, src_dimensions[i], &src_slice))
+      return EINVAL;
+  if (__builtin_mul_overflow (dst_slice, dst_offsets[0], &dst_off)
+      || __builtin_mul_overflow (src_slice, src_offsets[0], &src_off))
+    return EINVAL;
+  for (j = 0; j < volume[0]; j++)
+    {
+      ret = omp_target_memcpy_rect_worker ((char *) dst + dst_off,
+					   (char *) src + src_off,
+					   element_size, num_dims - 1,
+					   volume + 1, dst_offsets + 1,
+					   src_offsets + 1, dst_dimensions + 1,
+					   src_dimensions + 1);
+      if (ret)
+	return ret;
+      dst_off += dst_slice;
+      src_off += src_slice;
+    }
+  return 0;
+}
+
+int
+omp_target_memcpy_rect (void *dst, void *src, size_t element_size,
+			int num_dims, const size_t *volume,
+			const size_t *dst_offsets,
+			const size_t *src_offsets,
+			const size_t *dst_dimensions,
+			const size_t *src_dimensions,
+			int dst_device_num, int src_device_num)
+{
+  if (!dst && !src)
+    return INT_MAX;
+
+  if (dst_device_num != GOMP_DEVICE_HOST_FALLBACK)
+    return EINVAL;
+  if (src_device_num != GOMP_DEVICE_HOST_FALLBACK)
+    return EINVAL;
+
+  int ret = omp_target_memcpy_rect_worker (dst, src, element_size, num_dims,
+					   volume, dst_offsets, src_offsets,
+					   dst_dimensions, src_dimensions);
+  return ret;
+}
+
+int
+omp_target_associate_ptr (void *host_ptr, void *device_ptr, size_t size,
+			  size_t device_offset, int device_num)
+{
+  return EINVAL;
+}
+
+int
+omp_target_disassociate_ptr (void *ptr, int device_num)
+{
+  return EINVAL;
+}
--- libgomp/fortran.c.jj	2014-05-15 10:56:31.593531223 +0200
+++ libgomp/fortran.c	2016-07-13 16:57:04.432535397 +0200
@@ -67,12 +67,20 @@ ialias_redirect (omp_get_active_level)
 ialias_redirect (omp_in_final)
 ialias_redirect (omp_get_cancellation)
 ialias_redirect (omp_get_proc_bind)
+ialias_redirect (omp_get_num_places)
+ialias_redirect (omp_get_place_num_procs)
+ialias_redirect (omp_get_place_proc_ids)
+ialias_redirect (omp_get_place_num)
+ialias_redirect (omp_get_partition_num_places)
+ialias_redirect (omp_get_partition_place_nums)
 ialias_redirect (omp_set_default_device)
 ialias_redirect (omp_get_default_device)
 ialias_redirect (omp_get_num_devices)
 ialias_redirect (omp_get_num_teams)
 ialias_redirect (omp_get_team_num)
 ialias_redirect (omp_is_initial_device)
+ialias_redirect (omp_get_initial_device)
+ialias_redirect (omp_get_max_task_priority)
 #endif

 #ifndef LIBGOMP_GNU_SYMBOL_VERSIONING
@@ -342,35 +350,35 @@ omp_get_wtime_ (void)
 }

 void
-omp_set_schedule_ (const int32_t *kind, const int32_t *modifier)
+omp_set_schedule_ (const int32_t *kind, const int32_t *chunk_size)
 {
-  omp_set_schedule (*kind, *modifier);
+  omp_set_schedule (*kind, *chunk_size);
 }

 void
-omp_set_schedule_8_ (const int32_t *kind, const int64_t *modifier)
+omp_set_schedule_8_ (const int32_t *kind, const int64_t *chunk_size)
 {
-  omp_set_schedule (*kind, TO_INT (*modifier));
+  omp_set_schedule (*kind, TO_INT (*chunk_size));
 }

 void
-omp_get_schedule_ (int32_t *kind, int32_t *modifier)
+omp_get_schedule_ (int32_t *kind, int32_t *chunk_size)
 {
   omp_sched_t k;
-  int m;
-  omp_get_schedule (&k, &m);
+  int cs;
+  omp_get_schedule (&k, &cs);
   *kind = k;
-  *modifier = m;
+  *chunk_size = cs;
 }

 void
-omp_get_schedule_8_ (int32_t *kind, int64_t *modifier)
+omp_get_schedule_8_ (int32_t *kind, int64_t *chunk_size)
 {
   omp_sched_t k;
-  int m;
-  omp_get_schedule (&k, &m);
+  int cs;
+  omp_get_schedule (&k, &cs);
   *kind = k;
-  *modifier = m;
+  *chunk_size = cs;
 }

 int32_t
@@ -451,6 +459,69 @@ omp_get_proc_bind_ (void)
   return omp_get_proc_bind ();
 }

+int32_t
+omp_get_num_places_ (void)
+{
+  return omp_get_num_places ();
+}
+
+int32_t
+omp_get_place_num_procs_ (const int32_t *place_num)
+{
+  return omp_get_place_num_procs (*place_num);
+}
+
+int32_t
+omp_get_place_num_procs_8_ (const int64_t *place_num)
+{
+  return omp_get_place_num_procs (TO_INT (*place_num));
+}
+
+void
+omp_get_place_proc_ids_ (const int32_t *place_num, int32_t *ids)
+{
+  omp_get_place_proc_ids (*place_num, (int *) ids);
+}
+
+void
+omp_get_place_proc_ids_8_ (const int64_t *place_num, int64_t *ids)
+{
+  gomp_get_place_proc_ids_8 (TO_INT (*place_num), ids);
+}
+
+int32_t
+omp_get_place_num_ (void)
+{
+  return omp_get_place_num ();
+}
+
+int32_t
+omp_get_partition_num_places_ (void)
+{
+  return omp_get_partition_num_places ();
+}
+
+void
+omp_get_partition_place_nums_ (int32_t *place_nums)
+{
+  omp_get_partition_place_nums ((int *) place_nums);
+}
+
+void
+omp_get_partition_place_nums_8_ (int64_t *place_nums)
+{
+  if (gomp_places_list == NULL)
+    return;
+
+  struct gomp_thread *thr = gomp_thread ();
+  if (thr->place == 0)
+    gomp_init_affinity ();
+
+  unsigned int i;
+  for (i = 0; i < thr->ts.place_partition_len; i++)
+    *place_nums++ = (int64_t) thr->ts.place_partition_off + i;
+}
+
 void
 omp_set_default_device_ (const int32_t *device_num)
 {
@@ -492,3 +563,15 @@ omp_is_initial_device_ (void)
 {
   return omp_is_initial_device ();
 }
+
+int32_t
+omp_get_initial_device_ (void)
+{
+  return omp_get_initial_device ();
+}
+
+int32_t
+omp_get_max_task_priority_ (void)
+{
+  return omp_get_max_task_priority ();
+}
--- libgomp/libgomp.map.jj	2014-05-15 10:56:31.927533549 +0200
+++ libgomp/libgomp.map	2016-07-13 16:57:04.434535373 +0200
@@ -134,6 +134,36 @@ OMP_4.0 {
	omp_is_initial_device_;
 } OMP_3.1;

+OMP_4.5 {
+  global:
+	omp_get_max_task_priority;
+	omp_get_max_task_priority_;
+	omp_get_num_places;
+	omp_get_num_places_;
+	omp_get_place_num_procs;
+	omp_get_place_num_procs_;
+	omp_get_place_num_procs_8_;
+	omp_get_place_proc_ids;
+	omp_get_place_proc_ids_;
+	omp_get_place_proc_ids_8_;
+	omp_get_place_num;
+	omp_get_place_num_;
+	omp_get_partition_num_places;
+	omp_get_partition_num_places_;
+	omp_get_partition_place_nums;
+	omp_get_partition_place_nums_;
+	omp_get_partition_place_nums_8_;
+	omp_get_initial_device;
+	omp_get_initial_device_;
+	omp_target_alloc;
+	omp_target_free;
+	omp_target_is_present;
+	omp_target_memcpy;
+	omp_target_memcpy_rect;
+	omp_target_associate_ptr;
+	omp_target_disassociate_ptr;
+} OMP_4.0;
+
 GOMP_1.0 {
   global:
	GOMP_atomic_end;
@@ -227,3 +257,158 @@ GOMP_4.0 {
	GOMP_target_update;
	GOMP_teams;
 } GOMP_3.0;
+
+GOMP_4.0.1 {
+  global:
+	GOMP_offload_register;
+	GOMP_offload_unregister;
+} GOMP_4.0;
+
+GOMP_4.5 {
+  global:
+	GOMP_target_ext;
+	GOMP_target_data_ext;
+	GOMP_target_update_ext;
+	GOMP_target_enter_exit_data;
+	GOMP_taskloop;
+	GOMP_taskloop_ull;
+	GOMP_offload_register_ver;
+	GOMP_offload_unregister_ver;
+	GOMP_loop_doacross_dynamic_start;
+	GOMP_loop_doacross_guided_start;
+	GOMP_loop_doacross_runtime_start;
+	GOMP_loop_doacross_static_start;
+	GOMP_doacross_post;
+	GOMP_doacross_wait;
+	GOMP_loop_ull_doacross_dynamic_start;
+	GOMP_loop_ull_doacross_guided_start;
+	GOMP_loop_ull_doacross_runtime_start;
+	GOMP_loop_ull_doacross_static_start;
+	GOMP_doacross_ull_post;
+	GOMP_doacross_ull_wait;
+	GOMP_loop_nonmonotonic_dynamic_next;
+	GOMP_loop_nonmonotonic_dynamic_start;
+	GOMP_loop_nonmonotonic_guided_next;
+	GOMP_loop_nonmonotonic_guided_start;
+	GOMP_loop_ull_nonmonotonic_dynamic_next;
+	GOMP_loop_ull_nonmonotonic_dynamic_start;
+	GOMP_loop_ull_nonmonotonic_guided_next;
+	GOMP_loop_ull_nonmonotonic_guided_start;
+	GOMP_parallel_loop_nonmonotonic_dynamic;
+	GOMP_parallel_loop_nonmonotonic_guided;
+} GOMP_4.0.1;
+
+OACC_2.0 {
+  global:
+	acc_get_num_devices;
+	acc_get_num_devices_h_;
+	acc_set_device_type;
+	acc_set_device_type_h_;
+	acc_get_device_type;
+	acc_get_device_type_h_;
+	acc_set_device_num;
+	acc_set_device_num_h_;
+	acc_get_device_num;
+	acc_get_device_num_h_;
+	acc_async_test;
+	acc_async_test_h_;
+	acc_async_test_all;
+	acc_async_test_all_h_;
+	acc_wait;
+	acc_wait_h_;
+	acc_wait_async;
+	acc_wait_async_h_;
+	acc_wait_all;
+	acc_wait_all_h_;
+	acc_wait_all_async;
+	acc_wait_all_async_h_;
+	acc_init;
+	acc_init_h_;
+	acc_shutdown;
+	acc_shutdown_h_;
+	acc_on_device;
+	acc_on_device_h_;
+	acc_malloc;
+	acc_free;
+	acc_copyin;
+	acc_copyin_32_h_;
+	acc_copyin_64_h_;
+	acc_copyin_array_h_;
+	acc_present_or_copyin;
+	acc_present_or_copyin_32_h_;
+	acc_present_or_copyin_64_h_;
+	acc_present_or_copyin_array_h_;
+	acc_create;
+	acc_create_32_h_;
+	acc_create_64_h_;
+	acc_create_array_h_;
+	acc_present_or_create;
+	acc_present_or_create_32_h_;
+	acc_present_or_create_64_h_;
+	acc_present_or_create_array_h_;
+	acc_copyout;
+	acc_copyout_32_h_;
+	acc_copyout_64_h_;
+	acc_copyout_array_h_;
+	acc_delete;
+	acc_delete_32_h_;
+	acc_delete_64_h_;
+	acc_delete_array_h_;
+	acc_update_device;
+	acc_update_device_32_h_;
+	acc_update_device_64_h_;
+	acc_update_device_array_h_;
+	acc_update_self;
+	acc_update_self_32_h_;
+	acc_update_self_64_h_;
+	acc_update_self_array_h_;
+	acc_map_data;
+	acc_unmap_data;
+	acc_deviceptr;
+	acc_hostptr;
+	acc_is_present;
+	acc_is_present_32_h_;
+	acc_is_present_64_h_;
+	acc_is_present_array_h_;
+	acc_memcpy_to_device;
+	acc_memcpy_from_device;
+	acc_get_current_cuda_device;
+	acc_get_current_cuda_context;
+	acc_get_cuda_stream;
+	acc_set_cuda_stream;
+};
+
+GOACC_2.0 {
+  global:
+	GOACC_data_end;
+	GOACC_data_start;
+	GOACC_enter_exit_data;
+	GOACC_parallel;
+	GOACC_update;
+	GOACC_wait;
+	GOACC_get_thread_num;
+	GOACC_get_num_threads;
+};
+
+GOACC_2.0.1 {
+  global:
+	GOACC_declare;
+	GOACC_parallel_keyed;
+} GOACC_2.0;
+
+GOMP_PLUGIN_1.0 {
+  global:
+	GOMP_PLUGIN_malloc;
+	GOMP_PLUGIN_malloc_cleared;
+	GOMP_PLUGIN_realloc;
+	GOMP_PLUGIN_debug;
+	GOMP_PLUGIN_error;
+	GOMP_PLUGIN_fatal;
+	GOMP_PLUGIN_async_unmap_vars;
+	GOMP_PLUGIN_acc_thread;
+};
+
+GOMP_PLUGIN_1.1 {
+  global:
+	GOMP_PLUGIN_target_task_completion;
+} GOMP_PLUGIN_1.0;
--- libgomp/ordered.c.jj	2013-01-21 16:00:46.137873657 +0100
+++ libgomp/ordered.c	2016-07-13 16:57:18.918355780 +0200
@@ -25,6 +25,9 @@
 /* This file handles the ORDERED construct.  */

 #include "libgomp.h"
+#include <stdarg.h>
+#include <string.h>
+#include "doacross.h"


 /* This function is called when first allocating an iteration block.  That
@@ -249,3 +252,533 @@ void
 GOMP_ordered_end (void)
 {
 }
+
+/* DOACROSS initialization.  */
+
+#define MAX_COLLAPSED_BITS (__SIZEOF_LONG__ * __CHAR_BIT__)
+
+void
+gomp_doacross_init (unsigned ncounts, long *counts, long chunk_size)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_team *team = thr->ts.team;
+  struct gomp_work_share *ws = thr->ts.work_share;
+  unsigned int i, bits[MAX_COLLAPSED_BITS], num_bits = 0;
+  unsigned long ent, num_ents, elt_sz, shift_sz;
+  struct gomp_doacross_work_share *doacross;
+
+  if (team == NULL || team->nthreads == 1)
+    return;
+
+  for (i = 0; i < ncounts; i++)
+    {
+      /* If any count is 0, GOMP_doacross_{post,wait} can't be called.  */
+      if (counts[i] == 0)
+	return;
+
+      if (num_bits <= MAX_COLLAPSED_BITS)
+	{
+	  unsigned int this_bits;
+	  if (counts[i] == 1)
+	    this_bits = 1;
+	  else
+	    this_bits = __SIZEOF_LONG__ * __CHAR_BIT__
+			- __builtin_clzl (counts[i] - 1);
+	  if (num_bits + this_bits <= MAX_COLLAPSED_BITS)
+	    {
+	      bits[i] = this_bits;
+	      num_bits += this_bits;
+	    }
+	  else
+	    num_bits = MAX_COLLAPSED_BITS + 1;
+	}
+    }
+
+  if (ws->sched == GFS_STATIC)
+    num_ents = team->nthreads;
+  else if (ws->sched == GFS_GUIDED)
+    num_ents = counts[0];
+  else
+    num_ents = (counts[0] - 1) / chunk_size + 1;
+  if (num_bits <= MAX_COLLAPSED_BITS)
+    {
+      elt_sz = sizeof (unsigned long);
+      shift_sz = ncounts * sizeof (unsigned int);
+    }
+  else
+    {
+      elt_sz = sizeof (unsigned long) * ncounts;
+      shift_sz = 0;
+    }
+  elt_sz = (elt_sz + 63) & ~63UL;
+
+  doacross = gomp_malloc (sizeof (*doacross) + 63 + num_ents * elt_sz
+			  + shift_sz);
+  doacross->chunk_size = chunk_size;
+  doacross->elt_sz = elt_sz;
+  doacross->ncounts = ncounts;
+  doacross->flattened = false;
+  doacross->array = (unsigned char *)
+		    ((((uintptr_t) (doacross + 1)) + 63 + shift_sz)
+		     & ~(uintptr_t) 63);
+  if (num_bits <= MAX_COLLAPSED_BITS)
+    {
+      unsigned int shift_count = 0;
+      doacross->flattened = true;
+      for (i = ncounts; i > 0; i--)
+	{
+	  doacross->shift_counts[i - 1] = shift_count;
+	  shift_count += bits[i - 1];
+	}
+      for (ent = 0; ent < num_ents; ent++)
+	*(unsigned long *) (doacross->array + ent * elt_sz) = 0;
+    }
+  else
+    for (ent = 0; ent < num_ents; ent++)
+      memset (doacross->array + ent * elt_sz, '\0',
+	      sizeof (unsigned long) * ncounts);
+  if (ws->sched == GFS_STATIC && chunk_size == 0)
+    {
+      unsigned long q = counts[0] / num_ents;
+      unsigned long t = counts[0] % num_ents;
+      doacross->boundary = t * (q + 1);
+      doacross->q = q;
+      doacross->t = t;
+    }
+  ws->doacross = doacross;
+}
+
+/* DOACROSS POST operation.  */
+
+void
+GOMP_doacross_post (long *counts)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_work_share *ws = thr->ts.work_share;
+  struct gomp_doacross_work_share *doacross = ws->doacross;
+  unsigned long ent;
+  unsigned int i;
+
+  if (__builtin_expect (doacross == NULL, 0))
+    {
+      __sync_synchronize ();
+      return;
+    }
+
+  if (__builtin_expect (ws->sched == GFS_STATIC, 1))
+    ent = thr->ts.team_id;
+  else if (ws->sched == GFS_GUIDED)
+    ent = counts[0];
+  else
+    ent = counts[0] / doacross->chunk_size;
+  unsigned long *array = (unsigned long *) (doacross->array
+					    + ent * doacross->elt_sz);
+
+  if (__builtin_expect (doacross->flattened, 1))
+    {
+      unsigned long flattened
+	= (unsigned long) counts[0] << doacross->shift_counts[0];
+
+      for (i = 1; i < doacross->ncounts; i++)
+	flattened |= (unsigned long) counts[i]
+		     << doacross->shift_counts[i];
+      flattened++;
+      if (flattened == __atomic_load_n (array, MEMMODEL_ACQUIRE))
+	__atomic_thread_fence (MEMMODEL_RELEASE);
+      else
+	__atomic_store_n (array, flattened, MEMMODEL_RELEASE);
+      return;
+    }
+
+  __atomic_thread_fence (MEMMODEL_ACQUIRE);
+  for (i = doacross->ncounts; i-- > 0; )
+    {
+      if (counts[i] + 1UL != __atomic_load_n (&array[i], MEMMODEL_RELAXED))
+	__atomic_store_n (&array[i], counts[i] + 1UL, MEMMODEL_RELEASE);
+    }
+}
+
+/* DOACROSS WAIT operation.  */
+
+void
+GOMP_doacross_wait (long first, ...)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_work_share *ws = thr->ts.work_share;
+  struct gomp_doacross_work_share *doacross = ws->doacross;
+  va_list ap;
+  unsigned long ent;
+  unsigned int i;
+
+  if (__builtin_expect (doacross == NULL, 0))
+    {
+      __sync_synchronize ();
+      return;
+    }
+
+  if (__builtin_expect (ws->sched == GFS_STATIC, 1))
+    {
+      if (ws->chunk_size == 0)
+	{
+	  if (first < doacross->boundary)
+	    ent = first / (doacross->q + 1);
+	  else
+	    ent = (first - doacross->boundary) / doacross->q
+		  + doacross->t;
+	}
+      else
+	ent = first / ws->chunk_size % thr->ts.team->nthreads;
+    }
+  else if (ws->sched == GFS_GUIDED)
+    ent = first;
+  else
+    ent = first / doacross->chunk_size;
+  unsigned long *array = (unsigned long *) (doacross->array
+					    + ent * doacross->elt_sz);
+
+  if (__builtin_expect (doacross->flattened, 1))
+    {
+      unsigned long flattened
+	= (unsigned long) first << doacross->shift_counts[0];
+      unsigned long cur;
+
+      va_start (ap, first);
+      for (i = 1; i < doacross->ncounts; i++)
+	flattened |= (unsigned long) va_arg (ap, long)
+		     << doacross->shift_counts[i];
+      cur = __atomic_load_n (array, MEMMODEL_ACQUIRE);
+      if (flattened < cur)
+	{
+	  __atomic_thread_fence (MEMMODEL_RELEASE);
+	  va_end (ap);
+	  return;
+	}
+      doacross_spin (array, flattened, cur);
+      __atomic_thread_fence (MEMMODEL_RELEASE);
+      va_end (ap);
+      return;
+    }
+
+  do
+    {
+      va_start (ap, first);
+      for (i = 0; i < doacross->ncounts; i++)
+	{
+	  unsigned long thisv
+	    = (unsigned long) (i ? va_arg (ap, long) : first) + 1;
+	  unsigned long cur = __atomic_load_n (&array[i], MEMMODEL_RELAXED);
+	  if (thisv < cur)
+	    {
+	      i = doacross->ncounts;
+	      break;
+	    }
+	  if (thisv > cur)
+	    break;
+	}
+      va_end (ap);
+      if (i == doacross->ncounts)
+	break;
+      cpu_relax ();
+    }
+  while (1);
+  __sync_synchronize ();
+}
+
+typedef unsigned long long gomp_ull;
+
+void
+gomp_doacross_ull_init (unsigned ncounts, gomp_ull *counts, gomp_ull chunk_size)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_team *team = thr->ts.team;
+  struct gomp_work_share *ws = thr->ts.work_share;
+  unsigned int i, bits[MAX_COLLAPSED_BITS], num_bits = 0;
+  unsigned long ent, num_ents, elt_sz, shift_sz;
+  struct gomp_doacross_work_share *doacross;
+
+  if (team == NULL || team->nthreads == 1)
+    return;
+
+  for (i = 0; i < ncounts; i++)
+    {
+      /* If any count is 0, GOMP_doacross_{post,wait} can't be called.  */
+      if (counts[i] == 0)
+	return;
+
+      if (num_bits <= MAX_COLLAPSED_BITS)
+	{
+	  unsigned int this_bits;
+	  if (counts[i] == 1)
+	    this_bits = 1;
+	  else
+	    this_bits = __SIZEOF_LONG_LONG__ * __CHAR_BIT__
+			- __builtin_clzll (counts[i] - 1);
+	  if (num_bits + this_bits <= MAX_COLLAPSED_BITS)
+	    {
+	      bits[i] = this_bits;
+	      num_bits += this_bits;
+	    }
+	  else
+	    num_bits = MAX_COLLAPSED_BITS + 1;
+	}
+    }
+
+  if (ws->sched == GFS_STATIC)
+    num_ents = team->nthreads;
+  else if (ws->sched == GFS_GUIDED)
+    num_ents = counts[0];
+  else
+    num_ents = (counts[0] - 1) / chunk_size + 1;
+  if (num_bits <= MAX_COLLAPSED_BITS)
+    {
+      elt_sz = sizeof (unsigned long);
+      shift_sz = ncounts * sizeof (unsigned int);
+    }
+  else
+    {
+      if (sizeof (gomp_ull) == sizeof (unsigned long))
+	elt_sz = sizeof (gomp_ull) * ncounts;
+      else if (sizeof (gomp_ull) == 2 * sizeof (unsigned long))
+	elt_sz = sizeof (unsigned long) * 2 * ncounts;
+      else
+	abort ();
+      shift_sz = 0;
+    }
+  elt_sz = (elt_sz + 63) & ~63UL;
+
+  doacross = gomp_malloc (sizeof (*doacross) + 63 + num_ents * elt_sz
+			  + shift_sz);
+  doacross->chunk_size_ull = chunk_size;
+  doacross->elt_sz = elt_sz;
+  doacross->ncounts = ncounts;
+  doacross->flattened = false;
+  doacross->boundary = 0;
+  doacross->array = (unsigned char *)
+		    ((((uintptr_t) (doacross + 1)) + 63 + shift_sz)
+		     & ~(uintptr_t) 63);
+  if (num_bits <= MAX_COLLAPSED_BITS)
+    {
+      unsigned int shift_count = 0;
+      doacross->flattened = true;
+      for (i = ncounts; i > 0; i--)
+	{
+	  doacross->shift_counts[i - 1] = shift_count;
+	  shift_count += bits[i - 1];
+	}
+      for (ent = 0; ent < num_ents; ent++)
+	*(unsigned long *) (doacross->array + ent * elt_sz) = 0;
+    }
+  else
+    for (ent = 0; ent < num_ents; ent++)
+      memset (doacross->array + ent * elt_sz, '\0',
+	      sizeof (unsigned long) * ncounts);
+  if (ws->sched == GFS_STATIC && chunk_size == 0)
+    {
+      gomp_ull q = counts[0] / num_ents;
+      gomp_ull t = counts[0] % num_ents;
+      doacross->boundary_ull = t * (q + 1);
+      doacross->q_ull = q;
+      doacross->t = t;
+    }
+  ws->doacross = doacross;
+}
+
+/* DOACROSS POST operation.  */
+
+void
+GOMP_doacross_ull_post (gomp_ull *counts)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_work_share *ws = thr->ts.work_share;
+  struct gomp_doacross_work_share *doacross = ws->doacross;
+  unsigned long ent;
+  unsigned int i;
+
+  if (__builtin_expect (doacross == NULL, 0))
+    {
+      __sync_synchronize ();
+      return;
+    }
+
+  if (__builtin_expect (ws->sched == GFS_STATIC, 1))
+    ent = thr->ts.team_id;
+  else if (ws->sched == GFS_GUIDED)
+    ent = counts[0];
+  else
+    ent = counts[0] / doacross->chunk_size_ull;
+
+  if (__builtin_expect (doacross->flattened, 1))
+    {
+      unsigned long *array = (unsigned long *) (doacross->array
+			      + ent * doacross->elt_sz);
+      gomp_ull flattened
+	= counts[0] << doacross->shift_counts[0];
+
+      for (i = 1; i < doacross->ncounts; i++)
+	flattened |= counts[i] << doacross->shift_counts[i];
+      flattened++;
+      if (flattened == __atomic_load_n (array, MEMMODEL_ACQUIRE))
+	__atomic_thread_fence (MEMMODEL_RELEASE);
+      else
+	__atomic_store_n (array, flattened, MEMMODEL_RELEASE);
+      return;
+    }
+
+  __atomic_thread_fence (MEMMODEL_ACQUIRE);
+  if (sizeof (gomp_ull) == sizeof (unsigned long))
+    {
+      gomp_ull *array = (gomp_ull *) (doacross->array
+				      + ent * doacross->elt_sz);
+
+      for (i = doacross->ncounts; i-- > 0; )
+	{
+	  if (counts[i] + 1UL != __atomic_load_n (&array[i], MEMMODEL_RELAXED))
+	    __atomic_store_n (&array[i], counts[i] + 1UL, MEMMODEL_RELEASE);
+	}
+    }
+  else
+    {
+      unsigned long *array = (unsigned long *) (doacross->array
+						+ ent * doacross->elt_sz);
+
+      for (i = doacross->ncounts; i-- > 0; )
+	{
+	  gomp_ull cull = counts[i] + 1UL;
+	  unsigned long c = (unsigned long) cull;
+	  if (c != __atomic_load_n (&array[2 * i + 1], MEMMODEL_RELAXED))
+	    __atomic_store_n (&array[2 * i + 1], c, MEMMODEL_RELEASE);
+	  c = cull >> (__SIZEOF_LONG_LONG__ * __CHAR_BIT__ / 2);
+	  if (c != __atomic_load_n (&array[2 * i], MEMMODEL_RELAXED))
+	    __atomic_store_n (&array[2 * i], c, MEMMODEL_RELEASE);
+	}
+    }
+}
+
+/* DOACROSS WAIT operation.  */
+
+void
+GOMP_doacross_ull_wait (gomp_ull first, ...)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_work_share *ws = thr->ts.work_share;
+  struct gomp_doacross_work_share *doacross = ws->doacross;
+  va_list ap;
+  unsigned long ent;
+  unsigned int i;
+
+  if (__builtin_expect (doacross == NULL, 0))
+    {
+      __sync_synchronize ();
+      return;
+    }
+
+  if (__builtin_expect (ws->sched == GFS_STATIC, 1))
+    {
+      if (ws->chunk_size_ull == 0)
+	{
+	  if (first < doacross->boundary_ull)
+	    ent = first / (doacross->q_ull + 1);
+	  else
+	    ent = (first - doacross->boundary_ull) / doacross->q_ull
+		  + doacross->t;
+	}
+      else
+	ent = first / ws->chunk_size_ull % thr->ts.team->nthreads;
+    }
+  else if (ws->sched == GFS_GUIDED)
+    ent = first;
+  else
+    ent = first / doacross->chunk_size_ull;
+
+  if (__builtin_expect (doacross->flattened, 1))
+    {
+      unsigned long *array = (unsigned long *) (doacross->array
+						+ ent * doacross->elt_sz);
+      gomp_ull flattened = first << doacross->shift_counts[0];
+      unsigned long cur;
+
+      va_start (ap, first);
+      for (i = 1; i < doacross->ncounts; i++)
+	flattened |= va_arg (ap, gomp_ull)
+		     << doacross->shift_counts[i];
+      cur = __atomic_load_n (array, MEMMODEL_ACQUIRE);
+      if (flattened < cur)
+	{
+	  __atomic_thread_fence (MEMMODEL_RELEASE);
+	  va_end (ap);
+	  return;
+	}
+      doacross_spin (array, flattened, cur);
+      __atomic_thread_fence (MEMMODEL_RELEASE);
+      va_end (ap);
+      return;
+    }
+
+  if (sizeof (gomp_ull) == sizeof (unsigned long))
+    {
+      gomp_ull *array = (gomp_ull *) (doacross->array
+				      + ent * doacross->elt_sz);
+      do
+	{
+	  va_start (ap, first);
+	  for (i = 0; i < doacross->ncounts; i++)
+	    {
+	      gomp_ull thisv
+		= (i ? va_arg (ap, gomp_ull) : first) + 1;
+	      gomp_ull cur = __atomic_load_n (&array[i], MEMMODEL_RELAXED);
+	      if (thisv < cur)
+		{
+		  i = doacross->ncounts;
+		  break;
+		}
+	      if (thisv > cur)
+		break;
+	    }
+	  va_end (ap);
+	  if (i == doacross->ncounts)
+	    break;
+	  cpu_relax ();
+	}
+      while (1);
+    }
+  else
+    {
+      unsigned long *array = (unsigned long *) (doacross->array
+						+ ent * doacross->elt_sz);
+      do
+	{
+	  va_start (ap, first);
+	  for (i = 0; i < doacross->ncounts; i++)
+	    {
+	      gomp_ull thisv
+		= (i ? va_arg (ap, gomp_ull) : first) + 1;
+	      unsigned long t
+		= thisv >> (__SIZEOF_LONG_LONG__ * __CHAR_BIT__ / 2);
+	      unsigned long cur
+		= __atomic_load_n (&array[2 * i], MEMMODEL_RELAXED);
+	      if (t < cur)
+		{
+		  i = doacross->ncounts;
+		  break;
+		}
+	      if (t > cur)
+		break;
+	      t = thisv;
+	      cur = __atomic_load_n (&array[2 * i + 1], MEMMODEL_RELAXED);
+	      if (t < cur)
+		{
+		  i = doacross->ncounts;
+		  break;
+		}
+	      if (t > cur)
+		break;
+	    }
+	  va_end (ap);
+	  if (i == doacross->ncounts)
+	    break;
+	  cpu_relax ();
+	}
+      while (1);
+    }
+  __sync_synchronize ();
+}
--- libgomp/loop.c.jj	2014-05-15 10:56:36.487505570 +0200
+++ libgomp/loop.c	2016-07-13 16:57:13.488423109 +0200
@@ -110,6 +110,11 @@ gomp_loop_static_start (long start, long
   return !gomp_iter_static_next (istart, iend);
 }

+/* The current dynamic implementation is always monotonic.  The
+   entrypoints without nonmonotonic in them have to be always monotonic,
+   but the nonmonotonic ones could be changed to use work-stealing for
+   improved scalability.  */
+
 static bool
 gomp_loop_dynamic_start (long start, long end, long incr, long chunk_size,
			 long *istart, long *iend)
@@ -135,6 +140,9 @@ gomp_loop_dynamic_start (long start, lon
   return ret;
 }

+/* Similarly as for dynamic, though the question is how can the chunk sizes
+   be decreased without a central locking or atomics.  */
+
 static bool
 gomp_loop_guided_start (long start, long end, long incr, long chunk_size,
			long *istart, long *iend)
@@ -168,13 +176,16 @@ GOMP_loop_runtime_start (long start, lon
   switch (icv->run_sched_var)
     {
     case GFS_STATIC:
-      return gomp_loop_static_start (start, end, incr, icv->run_sched_modifier,
+      return gomp_loop_static_start (start, end, incr,
+				     icv->run_sched_chunk_size,
				     istart, iend);
     case GFS_DYNAMIC:
-      return gomp_loop_dynamic_start (start, end, incr, icv->run_sched_modifier,
+      return gomp_loop_dynamic_start (start, end, incr,
+				      icv->run_sched_chunk_size,
				      istart, iend);
     case GFS_GUIDED:
-      return gomp_loop_guided_start (start, end, incr, icv->run_sched_modifier,
+      return gomp_loop_guided_start (start, end, incr,
+				     icv->run_sched_chunk_size,
				     istart, iend);
     case GFS_AUTO:
       /* For now map to schedule(static), later on we could play with feedback
@@ -265,15 +276,15 @@ GOMP_loop_ordered_runtime_start (long st
     {
     case GFS_STATIC:
       return gomp_loop_ordered_static_start (start, end, incr,
-					     icv->run_sched_modifier,
+					     icv->run_sched_chunk_size,
					     istart, iend);
     case GFS_DYNAMIC:
       return gomp_loop_ordered_dynamic_start (start, end, incr,
-					      icv->run_sched_modifier,
+					      icv->run_sched_chunk_size,
					      istart, iend);
     case GFS_GUIDED:
       return gomp_loop_ordered_guided_start (start, end, incr,
-					     icv->run_sched_modifier,
+					     icv->run_sched_chunk_size,
					     istart, iend);
     case GFS_AUTO:
       /* For now map to schedule(static), later on we could play with feedback
@@ -285,6 +296,111 @@ GOMP_loop_ordered_runtime_start (long st
     }
 }

+/* The *_doacross_*_start routines are similar.  The only difference is that
+   this work-share construct is initialized to expect an ORDERED(N) - DOACROSS
+   section, and the worksharing loop iterates always from 0 to COUNTS[0] - 1
+   and other COUNTS array elements tell the library number of iterations
+   in the ordered inner loops.  */
+
+static bool
+gomp_loop_doacross_static_start (unsigned ncounts, long *counts,
+				 long chunk_size, long *istart, long *iend)
+{
+  struct gomp_thread *thr = gomp_thread ();
+
+  thr->ts.static_trip = 0;
+  if (gomp_work_share_start (false))
+    {
+      gomp_loop_init (thr->ts.work_share, 0, counts[0], 1,
+		      GFS_STATIC, chunk_size);
+      gomp_doacross_init (ncounts, counts, chunk_size);
+      gomp_work_share_init_done ();
+    }
+
+  return !gomp_iter_static_next (istart, iend);
+}
+
+static bool
+gomp_loop_doacross_dynamic_start (unsigned ncounts, long *counts,
+				  long chunk_size, long *istart, long *iend)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  bool ret;
+
+  if (gomp_work_share_start (false))
+    {
+      gomp_loop_init (thr->ts.work_share, 0, counts[0], 1,
+		      GFS_DYNAMIC, chunk_size);
+      gomp_doacross_init (ncounts, counts, chunk_size);
+      gomp_work_share_init_done ();
+    }
+
+#ifdef HAVE_SYNC_BUILTINS
+  ret = gomp_iter_dynamic_next (istart, iend);
+#else
+  gomp_mutex_lock (&thr->ts.work_share->lock);
+  ret = gomp_iter_dynamic_next_locked (istart, iend);
+  gomp_mutex_unlock (&thr->ts.work_share->lock);
+#endif
+
+  return ret;
+}
+
+static bool
+gomp_loop_doacross_guided_start (unsigned ncounts, long *counts,
+				 long chunk_size, long *istart, long *iend)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  bool ret;
+
+  if (gomp_work_share_start (false))
+    {
+      gomp_loop_init (thr->ts.work_share, 0, counts[0], 1,
+		      GFS_GUIDED, chunk_size);
+      gomp_doacross_init (ncounts, counts, chunk_size);
+      gomp_work_share_init_done ();
+    }
+
+#ifdef HAVE_SYNC_BUILTINS
+  ret = gomp_iter_guided_next (istart, iend);
+#else
+  gomp_mutex_lock (&thr->ts.work_share->lock);
+  ret = gomp_iter_guided_next_locked (istart, iend);
+  gomp_mutex_unlock (&thr->ts.work_share->lock);
+#endif
+
+  return ret;
+}
+
+bool
+GOMP_loop_doacross_runtime_start (unsigned ncounts, long *counts,
+				  long *istart, long *iend)
+{
+  struct gomp_task_icv *icv = gomp_icv (false);
+  switch (icv->run_sched_var)
+    {
+    case GFS_STATIC:
+      return gomp_loop_doacross_static_start (ncounts, counts,
+					      icv->run_sched_chunk_size,
+					      istart, iend);
+    case GFS_DYNAMIC:
+      return gomp_loop_doacross_dynamic_start (ncounts, counts,
+					       icv->run_sched_chunk_size,
+					       istart, iend);
+    case GFS_GUIDED:
+      return gomp_loop_doacross_guided_start (ncounts, counts,
+					      icv->run_sched_chunk_size,
+					      istart, iend);
+    case GFS_AUTO:
+      /* For now map to schedule(static), later on we could play with feedback
+	 driven choice.  */
+      return gomp_loop_doacross_static_start (ncounts, counts,
+					      0, istart, iend);
+    default:
+      abort ();
+    }
+}
+
 /* The *_next routines are called when the thread completes processing of
    the iteration block currently assigned to it.  If the work-share
    construct is bound directly to a parallel construct, then the iteration
@@ -483,7 +599,7 @@ GOMP_parallel_loop_runtime_start (void (
 {
   struct gomp_task_icv *icv = gomp_icv (false);
   gomp_parallel_loop_start (fn, data, num_threads, start, end, incr,
-			    icv->run_sched_var, icv->run_sched_modifier, 0);
+			    icv->run_sched_var, icv->run_sched_chunk_size, 0);
 }

 ialias_redirect (GOMP_parallel_end)
@@ -521,6 +637,37 @@ GOMP_parallel_loop_guided (void (*fn) (v
   GOMP_parallel_end ();
 }

+#ifdef HAVE_ATTRIBUTE_ALIAS
+extern __typeof(GOMP_parallel_loop_dynamic) GOMP_parallel_loop_nonmonotonic_dynamic
+	__attribute__((alias ("GOMP_parallel_loop_dynamic")));
+extern __typeof(GOMP_parallel_loop_guided) GOMP_parallel_loop_nonmonotonic_guided
+	__attribute__((alias ("GOMP_parallel_loop_guided")));
+#else
+void
+GOMP_parallel_loop_nonmonotonic_dynamic (void (*fn) (void *), void *data,
+					 unsigned num_threads, long start,
+					 long end, long incr, long chunk_size,
+					 unsigned flags)
+{
+  gomp_parallel_loop_start (fn, data, num_threads, start, end, incr,
+			    GFS_DYNAMIC, chunk_size, flags);
+  fn (data);
+  GOMP_parallel_end ();
+}
+
+void
+GOMP_parallel_loop_nonmonotonic_guided (void (*fn) (void *), void *data,
+					unsigned num_threads, long start,
+					long end, long incr, long chunk_size,
+					unsigned flags)
+{
+  gomp_parallel_loop_start (fn, data, num_threads, start, end, incr,
+			    GFS_GUIDED, chunk_size, flags);
+  fn (data);
+  GOMP_parallel_end ();
+}
+#endif
+
 void
 GOMP_parallel_loop_runtime (void (*fn) (void *), void *data,
			    unsigned num_threads, long start, long end,
@@ -528,7 +675,7 @@ GOMP_parallel_loop_runtime (void (*fn) (
 {
   struct gomp_task_icv *icv = gomp_icv (false);
   gomp_parallel_loop_start (fn, data, num_threads, start, end, incr,
-			    icv->run_sched_var, icv->run_sched_modifier,
+			    icv->run_sched_var, icv->run_sched_chunk_size,
			    flags);
   fn (data);
   GOMP_parallel_end ();
@@ -569,6 +716,10 @@ extern __typeof(gomp_loop_dynamic_start)
	__attribute__((alias ("gomp_loop_dynamic_start")));
 extern __typeof(gomp_loop_guided_start) GOMP_loop_guided_start
	__attribute__((alias ("gomp_loop_guided_start")));
+extern __typeof(gomp_loop_dynamic_start) GOMP_loop_nonmonotonic_dynamic_start
+	__attribute__((alias ("gomp_loop_dynamic_start")));
+extern __typeof(gomp_loop_guided_start) GOMP_loop_nonmonotonic_guided_start
+	__attribute__((alias ("gomp_loop_guided_start")));

 extern __typeof(gomp_loop_ordered_static_start) GOMP_loop_ordered_static_start
	__attribute__((alias ("gomp_loop_ordered_static_start")));
@@ -577,12 +728,23 @@ extern __typeof(gomp_loop_ordered_dynami
 extern __typeof(gomp_loop_ordered_guided_start) GOMP_loop_ordered_guided_start
	__attribute__((alias ("gomp_loop_ordered_guided_start")));

+extern __typeof(gomp_loop_doacross_static_start) GOMP_loop_doacross_static_start
+	__attribute__((alias ("gomp_loop_doacross_static_start")));
+extern __typeof(gomp_loop_doacross_dynamic_start) GOMP_loop_doacross_dynamic_start
+	__attribute__((alias ("gomp_loop_doacross_dynamic_start")));
+extern __typeof(gomp_loop_doacross_guided_start) GOMP_loop_doacross_guided_start
+	__attribute__((alias ("gomp_loop_doacross_guided_start")));
+
 extern __typeof(gomp_loop_static_next) GOMP_loop_static_next
	__attribute__((alias ("gomp_loop_static_next")));
 extern __typeof(gomp_loop_dynamic_next) GOMP_loop_dynamic_next
	__attribute__((alias ("gomp_loop_dynamic_next")));
 extern __typeof(gomp_loop_guided_next) GOMP_loop_guided_next
	__attribute__((alias ("gomp_loop_guided_next")));
+extern __typeof(gomp_loop_dynamic_next) GOMP_loop_nonmonotonic_dynamic_next
+	__attribute__((alias ("gomp_loop_dynamic_next")));
+extern __typeof(gomp_loop_guided_next) GOMP_loop_nonmonotonic_guided_next
+	__attribute__((alias ("gomp_loop_guided_next")));

 extern __typeof(gomp_loop_ordered_static_next) GOMP_loop_ordered_static_next
	__attribute__((alias ("gomp_loop_ordered_static_next")));
@@ -613,6 +775,21 @@ GOMP_loop_guided_start (long start, long
 }

 bool
+GOMP_loop_nonmonotonic_dynamic_start (long start, long end, long incr,
+				      long chunk_size, long *istart,
+				      long *iend)
+{
+  return gomp_loop_dynamic_start (start, end, incr, chunk_size, istart, iend);
+}
+
+bool
+GOMP_loop_nonmonotonic_guided_start (long start, long end, long incr,
+				     long chunk_size, long *istart, long *iend)
+{
+  return gomp_loop_guided_start (start, end, incr, chunk_size, istart, iend);
+}
+
+bool
 GOMP_loop_ordered_static_start (long start, long end, long incr,
				long chunk_size, long *istart, long *iend)
 {
@@ -637,6 +814,30 @@ GOMP_loop_ordered_guided_start (long sta
 }

 bool
+GOMP_loop_doacross_static_start (unsigned ncounts, long *counts,
+				 long chunk_size, long *istart, long *iend)
+{
+  return gomp_loop_doacross_static_start (ncounts, counts, chunk_size,
+					  istart, iend);
+}
+
+bool
+GOMP_loop_doacross_dynamic_start (unsigned ncounts, long *counts,
+				  long chunk_size, long *istart, long *iend)
+{
+  return gomp_loop_doacross_dynamic_start (ncounts, counts, chunk_size,
+					   istart, iend);
+}
+
+bool
+GOMP_loop_doacross_guided_start (unsigned ncounts, long *counts,
+				 long chunk_size, long *istart, long *iend)
+{
+  return gomp_loop_doacross_guided_start (ncounts, counts, chunk_size,
+					  istart, iend);
+}
+
+bool
 GOMP_loop_static_next (long *istart, long *iend)
 {
   return gomp_loop_static_next (istart, iend);
@@ -653,6 +854,18 @@ GOMP_loop_guided_next (long *istart, lon
 {
   return gomp_loop_guided_next (istart, iend);
 }
+
+bool
+GOMP_loop_nonmonotonic_dynamic_next (long *istart, long *iend)
+{
+  return gomp_loop_dynamic_next (istart, iend);
+}
+
+bool
+GOMP_loop_nonmonotonic_guided_next (long *istart, long *iend)
+{
+  return gomp_loop_guided_next (istart, iend);
+}

 bool
 GOMP_loop_ordered_static_next (long *istart, long *iend)
--- libgomp/error.c.jj	2013-01-21 16:00:31.834953566 +0100
+++ libgomp/error.c	2016-07-13 16:57:04.437535335 +0200
@@ -35,7 +35,26 @@
 #include <stdlib.h>


-static void
+#undef gomp_vdebug
+void
+gomp_vdebug (int kind __attribute__ ((unused)), const char *msg, va_list list)
+{
+  if (gomp_debug_var)
+    vfprintf (stderr, msg, list);
+}
+
+#undef gomp_debug
+void
+gomp_debug (int kind, const char *msg, ...)
+{
+  va_list list;
+
+  va_start (list, msg);
+  gomp_vdebug (kind, msg, list);
+  va_end (list);
+}
+
+void
 gomp_verror (const char *fmt, va_list list)
 {
   fputs ("\nlibgomp: ", stderr);
@@ -54,13 +73,18 @@ gomp_error (const char *fmt, ...)
 }

 void
+gomp_vfatal (const char *fmt, va_list list)
+{
+  gomp_verror (fmt, list);
+  exit (EXIT_FAILURE);
+}
+
+void
 gomp_fatal (const char *fmt, ...)
 {
   va_list list;

   va_start (list, fmt);
-  gomp_verror (fmt, list);
+  gomp_vfatal (fmt, list);
   va_end (list);
-
-  exit (EXIT_FAILURE);
 }
--- libgomp/Makefile.am.jj	2014-05-15 11:12:10.000000000 +0200
+++ libgomp/Makefile.am	2016-07-14 16:10:51.968202878 +0200
@@ -60,7 +60,13 @@ libgomp_la_LINK = $(LINK) $(libgomp_la_L
 libgomp_la_SOURCES = alloc.c barrier.c critical.c env.c error.c iter.c \
	iter_ull.c loop.c loop_ull.c ordered.c parallel.c sections.c single.c \
	task.c team.c work.c lock.c mutex.c proc.c sem.c bar.c ptrlock.c \
-	time.c fortran.c affinity.c target.c
+	time.c fortran.c affinity.c target.c splay-tree.c libgomp-plugin.c \
+	oacc-parallel.c oacc-host.c oacc-init.c oacc-mem.c oacc-async.c \
+	oacc-plugin.c oacc-cuda.c priority_queue.c
+
+if USE_FORTRAN
+libgomp_la_SOURCES += openacc.f90
+endif

 nodist_noinst_HEADERS = libgomp_f.h
 nodist_libsubinclude_HEADERS = omp.h
--- libgomp/Makefile.in.jj	2014-05-15 11:12:10.000000000 +0200
+++ libgomp/Makefile.in	2016-07-14 16:11:10.981954087 +0200
@@ -36,6 +36,7 @@ POST_UNINSTALL = :
 build_triplet = @build@
 host_triplet = @host@
 target_triplet = @target@
+@USE_FORTRAN_TRUE@am__append_1 = openacc.f90
 subdir = .
 DIST_COMMON = ChangeLog $(srcdir)/Makefile.in $(srcdir)/Makefile.am \
	$(top_srcdir)/configure $(am__configure_deps) \
@@ -92,11 +93,15 @@ am__installdirs = "$(DESTDIR)$(toolexecl
	"$(DESTDIR)$(toolexeclibdir)"
 LTLIBRARIES = $(toolexeclib_LTLIBRARIES)
 libgomp_la_LIBADD =
+@USE_FORTRAN_TRUE@am__objects_1 = openacc.lo
 am_libgomp_la_OBJECTS = alloc.lo barrier.lo critical.lo env.lo \
	error.lo iter.lo iter_ull.lo loop.lo loop_ull.lo ordered.lo \
	parallel.lo sections.lo single.lo task.lo team.lo work.lo \
	lock.lo mutex.lo proc.lo sem.lo bar.lo ptrlock.lo time.lo \
-	fortran.lo affinity.lo target.lo
+	fortran.lo affinity.lo target.lo splay-tree.lo \
+	libgomp-plugin.lo oacc-parallel.lo oacc-host.lo oacc-init.lo \
+	oacc-mem.lo oacc-async.lo oacc-plugin.lo oacc-cuda.lo \
+	priority_queue.lo $(am__objects_1)
 libgomp_la_OBJECTS = $(am_libgomp_la_OBJECTS)
 DEFAULT_INCLUDES = -I.@am__isrc@
 depcomp = $(SHELL) $(top_srcdir)/../depcomp
@@ -108,6 +113,13 @@ LTCOMPILE = $(LIBTOOL) --tag=CC $(AM_LIB
	--mode=compile $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) \
	$(AM_CPPFLAGS) $(CPPFLAGS) $(AM_CFLAGS) $(CFLAGS)
 CCLD = $(CC)
+FCCOMPILE = $(FC) $(AM_FCFLAGS) $(FCFLAGS)
+LTFCCOMPILE = $(LIBTOOL) --tag=FC $(AM_LIBTOOLFLAGS) $(LIBTOOLFLAGS) \
+	--mode=compile $(FC) $(AM_FCFLAGS) $(FCFLAGS)
+FCLD = $(FC)
+FCLINK = $(LIBTOOL) --tag=FC $(AM_LIBTOOLFLAGS) $(LIBTOOLFLAGS) \
+	--mode=link $(FCLD) $(AM_FCFLAGS) $(FCFLAGS) $(AM_LDFLAGS) \
+	$(LDFLAGS) -o $@
 SOURCES = $(libgomp_la_SOURCES)
 MULTISRCTOP =
 MULTIBUILDTOP =
@@ -315,10 +327,12 @@ libgomp_la_LDFLAGS = $(libgomp_version_i
 libgomp_la_DEPENDENCIES = $(libgomp_version_dep)
 libgomp_la_LINK = $(LINK) $(libgomp_la_LDFLAGS)
 libgomp_la_SOURCES = alloc.c barrier.c critical.c env.c error.c iter.c \
-	iter_ull.c loop.c loop_ull.c ordered.c parallel.c sections.c single.c \
-	task.c team.c work.c lock.c mutex.c proc.c sem.c bar.c ptrlock.c \
-	time.c fortran.c affinity.c target.c
-
+	iter_ull.c loop.c loop_ull.c ordered.c parallel.c sections.c \
+	single.c task.c team.c work.c lock.c mutex.c proc.c sem.c \
+	bar.c ptrlock.c time.c fortran.c affinity.c target.c \
+	splay-tree.c libgomp-plugin.c oacc-parallel.c oacc-host.c \
+	oacc-init.c oacc-mem.c oacc-async.c oacc-plugin.c oacc-cuda.c \
+	priority_queue.c $(am__append_1)
 nodist_noinst_HEADERS = libgomp_f.h
 nodist_libsubinclude_HEADERS = omp.h
 @USE_FORTRAN_TRUE@nodist_finclude_HEADERS = omp_lib.h omp_lib.f90 omp_lib.mod omp_lib_kinds.mod
@@ -351,7 +365,7 @@ all: config.h
	$(MAKE) $(AM_MAKEFLAGS) all-recursive

 .SUFFIXES:
-.SUFFIXES: .c .dvi .lo .o .obj .ps
+.SUFFIXES: .c .dvi .f90 .lo .o .obj .ps
 am--refresh:
	@:
 $(srcdir)/Makefile.in: @MAINTAINER_MODE_TRUE@ $(srcdir)/Makefile.am  $(am__configure_deps)
@@ -463,17 +477,27 @@ distclean-compile:
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/fortran.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/iter.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/iter_ull.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/libgomp-plugin.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/lock.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/loop.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/loop_ull.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/mutex.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/oacc-async.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/oacc-cuda.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/oacc-host.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/oacc-init.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/oacc-mem.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/oacc-parallel.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/oacc-plugin.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/ordered.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/parallel.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/priority_queue.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/proc.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/ptrlock.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/sections.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/sem.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/single.Plo@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/splay-tree.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/target.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/task.Plo@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/team.Plo@am__quote@
@@ -501,6 +525,15 @@ distclean-compile:
 @AMDEP_TRUE@@am__fastdepCC_FALSE@	DEPDIR=$(DEPDIR) $(CCDEPMODE) $(depcomp) @AMDEPBACKSLASH@
 @am__fastdepCC_FALSE@	$(LTCOMPILE) -c -o $@ $<

+.f90.o:
+	$(FCCOMPILE) -c -o $@ $<
+
+.f90.obj:
+	$(FCCOMPILE) -c -o $@ `$(CYGPATH_W) '$<'`
+
+.f90.lo:
+	$(LTFCCOMPILE) -c -o $@ $<
+
 mostlyclean-libtool:
	-rm -f *.lo

--- libgomp/task.c.jj	2014-08-06 16:25:16.575091658 +0200
+++ libgomp/task.c	2016-07-13 17:47:58.722758497 +0200
@@ -28,6 +28,7 @@
 #include "libgomp.h"
 #include <stdlib.h>
 #include <string.h>
+#include "gomp-constants.h"

 typedef struct gomp_task_depend_entry *hash_entry_type;

@@ -63,6 +64,14 @@ void
 gomp_init_task (struct gomp_task *task, struct gomp_task *parent_task,
		struct gomp_task_icv *prev_icv)
 {
+  /* It would seem that using memset here would be a win, but it turns
+     out that partially filling gomp_task allows us to keep the
+     overhead of task creation low.  In the nqueens-1.c test, for a
+     sufficiently large N, we drop the overhead from 5-6% to 1%.
+
+     Note, the nqueens-1.c test in serial mode is a good test to
+     benchmark the overhead of creating tasks as there are millions of
+     tiny tasks created that all run undeferred.  */
   task->parent = parent_task;
   task->icv = *prev_icv;
   task->kind = GOMP_TASK_IMPLICIT;
@@ -71,7 +80,7 @@ gomp_init_task (struct gomp_task *task,
   task->final_task = false;
   task->copy_ctors_done = false;
   task->parent_depends_on = false;
-  task->children = NULL;
+  priority_queue_init (&task->children_queue);
   task->taskgroup = NULL;
   task->dependers = NULL;
   task->depend_hash = NULL;
@@ -90,30 +99,194 @@ gomp_end_task (void)
   thr->task = task->parent;
 }

+/* Clear the parent field of every task in LIST.  */
+
 static inline void
-gomp_clear_parent (struct gomp_task *children)
+gomp_clear_parent_in_list (struct priority_list *list)
 {
-  struct gomp_task *task = children;
-
-  if (task)
+  struct priority_node *p = list->tasks;
+  if (p)
     do
       {
-	task->parent = NULL;
-	task = task->next_child;
+	priority_node_to_task (PQ_CHILDREN, p)->parent = NULL;
+	p = p->next;
       }
-    while (task != children);
+    while (p != list->tasks);
+}
+
+/* Splay tree version of gomp_clear_parent_in_list.
+
+   Clear the parent field of every task in NODE within SP, and free
+   the node when done.  */
+
+static void
+gomp_clear_parent_in_tree (prio_splay_tree sp, prio_splay_tree_node node)
+{
+  if (!node)
+    return;
+  prio_splay_tree_node left = node->left, right = node->right;
+  gomp_clear_parent_in_list (&node->key.l);
+#if _LIBGOMP_CHECKING_
+  memset (node, 0xaf, sizeof (*node));
+#endif
+  /* No need to remove the node from the tree.  We're nuking
+     everything, so just free the nodes and our caller can clear the
+     entire splay tree.  */
+  free (node);
+  gomp_clear_parent_in_tree (sp, left);
+  gomp_clear_parent_in_tree (sp, right);
+}
+
+/* Clear the parent field of every task in Q and remove every task
+   from Q.  */
+
+static inline void
+gomp_clear_parent (struct priority_queue *q)
+{
+  if (priority_queue_multi_p (q))
+    {
+      gomp_clear_parent_in_tree (&q->t, q->t.root);
+      /* All the nodes have been cleared in gomp_clear_parent_in_tree.
+	 No need to remove anything.  We can just nuke everything.  */
+      q->t.root = NULL;
+    }
+  else
+    gomp_clear_parent_in_list (&q->l);
 }

-static void gomp_task_maybe_wait_for_dependencies (void **depend);
+/* Helper function for GOMP_task and gomp_create_target_task.
+
+   For a TASK with in/out dependencies, fill in the various dependency
+   queues.  PARENT is the parent of said task.  DEPEND is as in
+   GOMP_task.  */
+
+static void
+gomp_task_handle_depend (struct gomp_task *task, struct gomp_task *parent,
+			 void **depend)
+{
+  size_t ndepend = (uintptr_t) depend[0];
+  size_t nout = (uintptr_t) depend[1];
+  size_t i;
+  hash_entry_type ent;
+
+  task->depend_count = ndepend;
+  task->num_dependees = 0;
+  if (parent->depend_hash == NULL)
+    parent->depend_hash = htab_create (2 * ndepend > 12 ? 2 * ndepend : 12);
+  for (i = 0; i < ndepend; i++)
+    {
+      task->depend[i].addr = depend[2 + i];
+      task->depend[i].next = NULL;
+      task->depend[i].prev = NULL;
+      task->depend[i].task = task;
+      task->depend[i].is_in = i >= nout;
+      task->depend[i].redundant = false;
+      task->depend[i].redundant_out = false;
+
+      hash_entry_type *slot = htab_find_slot (&parent->depend_hash,
+					      &task->depend[i], INSERT);
+      hash_entry_type out = NULL, last = NULL;
+      if (*slot)
+	{
+	  /* If multiple depends on the same task are the same, all but the
+	     first one are redundant.  As inout/out come first, if any of them
+	     is inout/out, it will win, which is the right semantics.  */
+	  if ((*slot)->task == task)
+	    {
+	      task->depend[i].redundant = true;
+	      continue;
+	    }
+	  for (ent = *slot; ent; ent = ent->next)
+	    {
+	      if (ent->redundant_out)
+		break;
+
+	      last = ent;
+
+	      /* depend(in:...) doesn't depend on earlier depend(in:...).  */
+	      if (i >= nout && ent->is_in)
+		continue;
+
+	      if (!ent->is_in)
+		out = ent;
+
+	      struct gomp_task *tsk = ent->task;
+	      if (tsk->dependers == NULL)
+		{
+		  tsk->dependers
+		    = gomp_malloc (sizeof (struct gomp_dependers_vec)
+				   + 6 * sizeof (struct gomp_task *));
+		  tsk->dependers->n_elem = 1;
+		  tsk->dependers->allocated = 6;
+		  tsk->dependers->elem[0] = task;
+		  task->num_dependees++;
+		  continue;
+		}
+	      /* We already have some other dependency on tsk from earlier
+		 depend clause.  */
+	      else if (tsk->dependers->n_elem
+		       && (tsk->dependers->elem[tsk->dependers->n_elem - 1]
+			   == task))
+		continue;
+	      else if (tsk->dependers->n_elem == tsk->dependers->allocated)
+		{
+		  tsk->dependers->allocated
+		    = tsk->dependers->allocated * 2 + 2;
+		  tsk->dependers
+		    = gomp_realloc (tsk->dependers,
+				    sizeof (struct gomp_dependers_vec)
+				    + (tsk->dependers->allocated
+				       * sizeof (struct gomp_task *)));
+		}
+	      tsk->dependers->elem[tsk->dependers->n_elem++] = task;
+	      task->num_dependees++;
+	    }
+	  task->depend[i].next = *slot;
+	  (*slot)->prev = &task->depend[i];
+	}
+      *slot = &task->depend[i];
+
+      /* There is no need to store more than one depend({,in}out:) task per
+	 address in the hash table chain for the purpose of creation of
+	 deferred tasks, because each out depends on all earlier outs, thus it
+	 is enough to record just the last depend({,in}out:).  For depend(in:),
+	 we need to keep all of the previous ones not terminated yet, because
+	 a later depend({,in}out:) might need to depend on all of them.  So, if
+	 the new task's clause is depend({,in}out:), we know there is at most
+	 one other depend({,in}out:) clause in the list (out).  For
+	 non-deferred tasks we want to see all outs, so they are moved to the
+	 end of the chain, after first redundant_out entry all following
+	 entries should be redundant_out.  */
+      if (!task->depend[i].is_in && out)
+	{
+	  if (out != last)
+	    {
+	      out->next->prev = out->prev;
+	      out->prev->next = out->next;
+	      out->next = last->next;
+	      out->prev = last;
+	      last->next = out;
+	      if (out->next)
+		out->next->prev = out;
+	    }
+	  out->redundant_out = true;
+	}
+    }
+}

 /* Called when encountering an explicit task directive.  If IF_CLAUSE is
    false, then we must not delay in executing the task.  If UNTIED is true,
-   then the task may be executed by any member of the team.  */
+   then the task may be executed by any member of the team.
+
+   DEPEND is an array containing:
+	depend[0]: number of depend elements.
+	depend[1]: number of depend elements of type "out".
+	depend[2..N+1]: address of [1..N]th depend element.  */

 void
 GOMP_task (void (*fn) (void *), void *data, void (*cpyfn) (void *, void *),
	   long arg_size, long arg_align, bool if_clause, unsigned flags,
-	   void **depend)
+	   void **depend, int priority)
 {
   struct gomp_thread *thr = gomp_thread ();
   struct gomp_team *team = thr->ts.team;
@@ -125,8 +298,7 @@ GOMP_task (void (*fn) (void *), void *da
      might be running on different thread than FN.  */
   if (cpyfn)
     if_clause = false;
-  if (flags & 1)
-    flags &= ~1;
+  flags &= ~GOMP_TASK_FLAG_UNTIED;
 #endif

   /* If parallel or taskgroup has been cancelled, don't start new tasks.  */
@@ -135,6 +307,11 @@ GOMP_task (void (*fn) (void *), void *da
	  || (thr->task->taskgroup && thr->task->taskgroup->cancelled)))
     return;

+  if ((flags & GOMP_TASK_FLAG_PRIORITY) == 0)
+    priority = 0;
+  else if (priority > gomp_max_task_priority_var)
+    priority = gomp_max_task_priority_var;
+
   if (!if_clause || team == NULL
       || (thr->task && thr->task->final_task)
       || team->task_count > 64 * team->nthreads)
@@ -147,12 +324,15 @@ GOMP_task (void (*fn) (void *), void *da
	 depend clauses for non-deferred tasks other than this, because
	 the parent task is suspended until the child task finishes and thus
	 it can't start further child tasks.  */
-      if ((flags & 8) && thr->task && thr->task->depend_hash)
+      if ((flags & GOMP_TASK_FLAG_DEPEND)
+	  && thr->task && thr->task->depend_hash)
	gomp_task_maybe_wait_for_dependencies (depend);

       gomp_init_task (&task, thr->task, gomp_icv (false));
-      task.kind = GOMP_TASK_IFFALSE;
-      task.final_task = (thr->task && thr->task->final_task) || (flags & 2);
+      task.kind = GOMP_TASK_UNDEFERRED;
+      task.final_task = (thr->task && thr->task->final_task)
+			|| (flags & GOMP_TASK_FLAG_FINAL);
+      task.priority = priority;
       if (thr->task)
	{
	  task.in_tied_task = thr->task->in_tied_task;
@@ -178,10 +358,10 @@ GOMP_task (void (*fn) (void *), void *da
	 child thread, but seeing a stale non-NULL value is not a
	 problem.  Once past the task_lock acquisition, this thread
	 will see the real value of task.children.  */
-      if (task.children != NULL)
+      if (!priority_queue_empty_p (&task.children_queue, MEMMODEL_RELAXED))
	{
	  gomp_mutex_lock (&team->task_lock);
-	  gomp_clear_parent (task.children);
+	  gomp_clear_parent (&task.children_queue);
	  gomp_mutex_unlock (&team->task_lock);
	}
       gomp_end_task ();
@@ -195,7 +375,7 @@ GOMP_task (void (*fn) (void *), void *da
       bool do_wake;
       size_t depend_size = 0;

-      if (flags & 8)
+      if (flags & GOMP_TASK_FLAG_DEPEND)
	depend_size = ((uintptr_t) depend[0]
		       * sizeof (struct gomp_task_depend_entry));
       task = gomp_malloc (sizeof (*task) + depend_size
@@ -203,7 +383,8 @@ GOMP_task (void (*fn) (void *), void *da
       arg = (char *) (((uintptr_t) (task + 1) + depend_size + arg_align - 1)
		      & ~(uintptr_t) (arg_align - 1));
       gomp_init_task (task, parent, gomp_icv (false));
-      task->kind = GOMP_TASK_IFFALSE;
+      task->priority = priority;
+      task->kind = GOMP_TASK_UNDEFERRED;
       task->in_tied_task = parent->in_tied_task;
       task->taskgroup = taskgroup;
       thr->task = task;
@@ -218,7 +399,7 @@ GOMP_task (void (*fn) (void *), void *da
       task->kind = GOMP_TASK_WAITING;
       task->fn = fn;
       task->fn_data = arg;
-      task->final_task = (flags & 2) >> 1;
+      task->final_task = (flags & GOMP_TASK_FLAG_FINAL) >> 1;
       gomp_mutex_lock (&team->task_lock);
       /* If parallel or taskgroup has been cancelled, don't start new
	 tasks.  */
@@ -235,171 +416,39 @@ GOMP_task (void (*fn) (void *), void *da
	taskgroup->num_children++;
       if (depend_size)
	{
-	  size_t ndepend = (uintptr_t) depend[0];
-	  size_t nout = (uintptr_t) depend[1];
-	  size_t i;
-	  hash_entry_type ent;
-
-	  task->depend_count = ndepend;
-	  task->num_dependees = 0;
-	  if (parent->depend_hash == NULL)
-	    parent->depend_hash
-	      = htab_create (2 * ndepend > 12 ? 2 * ndepend : 12);
-	  for (i = 0; i < ndepend; i++)
-	    {
-	      task->depend[i].addr = depend[2 + i];
-	      task->depend[i].next = NULL;
-	      task->depend[i].prev = NULL;
-	      task->depend[i].task = task;
-	      task->depend[i].is_in = i >= nout;
-	      task->depend[i].redundant = false;
-	      task->depend[i].redundant_out = false;
-
-	      hash_entry_type *slot
-		= htab_find_slot (&parent->depend_hash, &task->depend[i],
-				  INSERT);
-	      hash_entry_type out = NULL, last = NULL;
-	      if (*slot)
-		{
-		  /* If multiple depends on the same task are the
-		     same, all but the first one are redundant.
-		     As inout/out come first, if any of them is
-		     inout/out, it will win, which is the right
-		     semantics.  */
-		  if ((*slot)->task == task)
-		    {
-		      task->depend[i].redundant = true;
-		      continue;
-		    }
-		  for (ent = *slot; ent; ent = ent->next)
-		    {
-		      if (ent->redundant_out)
-			break;
-
-		      last = ent;
-
-		      /* depend(in:...) doesn't depend on earlier
-			 depend(in:...).  */
-		      if (i >= nout && ent->is_in)
-			continue;
-
-		      if (!ent->is_in)
-			out = ent;
-
-		      struct gomp_task *tsk = ent->task;
-		      if (tsk->dependers == NULL)
-			{
-			  tsk->dependers
-			    = gomp_malloc (sizeof (struct gomp_dependers_vec)
-					   + 6 * sizeof (struct gomp_task *));
-			  tsk->dependers->n_elem = 1;
-			  tsk->dependers->allocated = 6;
-			  tsk->dependers->elem[0] = task;
-			  task->num_dependees++;
-			  continue;
-			}
-		      /* We already have some other dependency on tsk
-			 from earlier depend clause.  */
-		      else if (tsk->dependers->n_elem
-			       && (tsk->dependers->elem[tsk->dependers->n_elem
-							- 1]
-				   == task))
-			continue;
-		      else if (tsk->dependers->n_elem
-			       == tsk->dependers->allocated)
-			{
-			  tsk->dependers->allocated
-			    = tsk->dependers->allocated * 2 + 2;
-			  tsk->dependers
-			    = gomp_realloc (tsk->dependers,
-					    sizeof (struct gomp_dependers_vec)
-					    + (tsk->dependers->allocated
-					       * sizeof (struct gomp_task *)));
-			}
-		      tsk->dependers->elem[tsk->dependers->n_elem++] = task;
-		      task->num_dependees++;
-		    }
-		  task->depend[i].next = *slot;
-		  (*slot)->prev = &task->depend[i];
-		}
-	      *slot = &task->depend[i];
-
-	      /* There is no need to store more than one depend({,in}out:)
-		 task per address in the hash table chain for the purpose
-		 of creation of deferred tasks, because each out
-		 depends on all earlier outs, thus it is enough to record
-		 just the last depend({,in}out:).  For depend(in:), we need
-		 to keep all of the previous ones not terminated yet, because
-		 a later depend({,in}out:) might need to depend on all of
-		 them.  So, if the new task's clause is depend({,in}out:),
-		 we know there is at most one other depend({,in}out:) clause
-		 in the list (out).  For non-deferred tasks we want to see
-		 all outs, so they are moved to the end of the chain,
-		 after first redundant_out entry all following entries
-		 should be redundant_out.  */
-	      if (!task->depend[i].is_in && out)
-		{
-		  if (out != last)
-		    {
-		      out->next->prev = out->prev;
-		      out->prev->next = out->next;
-		      out->next = last->next;
-		      out->prev = last;
-		      last->next = out;
-		      if (out->next)
-			out->next->prev = out;
-		    }
-		  out->redundant_out = true;
-		}
-	    }
+	  gomp_task_handle_depend (task, parent, depend);
	  if (task->num_dependees)
	    {
+	      /* Tasks that depend on other tasks are not put into the
+		 various waiting queues, so we are done for now.  Said
+		 tasks are instead put into the queues via
+		 gomp_task_run_post_handle_dependers() after their
+		 dependencies have been satisfied.  After which, they
+		 can be picked up by the various scheduling
+		 points.  */
	      gomp_mutex_unlock (&team->task_lock);
	      return;
	    }
	}
-      if (parent->children)
-	{
-	  task->next_child = parent->children;
-	  task->prev_child = parent->children->prev_child;
-	  task->next_child->prev_child = task;
-	  task->prev_child->next_child = task;
-	}
-      else
-	{
-	  task->next_child = task;
-	  task->prev_child = task;
-	}
-      parent->children = task;
+
+      priority_queue_insert (PQ_CHILDREN, &parent->children_queue,
+			     task, priority,
+			     PRIORITY_INSERT_BEGIN,
+			     /*adjust_parent_depends_on=*/false,
+			     task->parent_depends_on);
       if (taskgroup)
-	{
-	  if (taskgroup->children)
-	    {
-	      task->next_taskgroup = taskgroup->children;
-	      task->prev_taskgroup = taskgroup->children->prev_taskgroup;
-	      task->next_taskgroup->prev_taskgroup = task;
-	      task->prev_taskgroup->next_taskgroup = task;
-	    }
-	  else
-	    {
-	      task->next_taskgroup = task;
-	      task->prev_taskgroup = task;
-	    }
-	  taskgroup->children = task;
-	}
-      if (team->task_queue)
-	{
-	  task->next_queue = team->task_queue;
-	  task->prev_queue = team->task_queue->prev_queue;
-	  task->next_queue->prev_queue = task;
-	  task->prev_queue->next_queue = task;
-	}
-      else
-	{
-	  task->next_queue = task;
-	  task->prev_queue = task;
-	  team->task_queue = task;
-	}
+	priority_queue_insert (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
+			       task, priority,
+			       PRIORITY_INSERT_BEGIN,
+			       /*adjust_parent_depends_on=*/false,
+			       task->parent_depends_on);
+
+      priority_queue_insert (PQ_TEAM, &team->task_queue,
+			     task, priority,
+			     PRIORITY_INSERT_END,
+			     /*adjust_parent_depends_on=*/false,
+			     task->parent_depends_on);
+
       ++team->task_count;
       ++team->task_queued_count;
       gomp_team_barrier_set_task_pending (&team->barrier);
@@ -411,36 +460,529 @@ GOMP_task (void (*fn) (void *), void *da
     }
 }

-static inline bool
-gomp_task_run_pre (struct gomp_task *child_task, struct gomp_task *parent,
-		   struct gomp_taskgroup *taskgroup, struct gomp_team *team)
+ialias (GOMP_taskgroup_start)
+ialias (GOMP_taskgroup_end)
+
+#define TYPE long
+#define UTYPE unsigned long
+#define TYPE_is_long 1
+#include "taskloop.c"
+#undef TYPE
+#undef UTYPE
+#undef TYPE_is_long
+
+#define TYPE unsigned long long
+#define UTYPE TYPE
+#define GOMP_taskloop GOMP_taskloop_ull
+#include "taskloop.c"
+#undef TYPE
+#undef UTYPE
+#undef GOMP_taskloop
+
+static void inline
+priority_queue_move_task_first (enum priority_queue_type type,
+				struct priority_queue *head,
+				struct gomp_task *task)
 {
+#if _LIBGOMP_CHECKING_
+  if (!priority_queue_task_in_queue_p (type, head, task))
+    gomp_fatal ("Attempt to move first missing task %p", task);
+#endif
+  struct priority_list *list;
+  if (priority_queue_multi_p (head))
+    {
+      list = priority_queue_lookup_priority (head, task->priority);
+#if _LIBGOMP_CHECKING_
+      if (!list)
+	gomp_fatal ("Unable to find priority %d", task->priority);
+#endif
+    }
+  else
+    list = &head->l;
+  priority_list_remove (list, task_to_priority_node (type, task), 0);
+  priority_list_insert (type, list, task, task->priority,
+			PRIORITY_INSERT_BEGIN, type == PQ_CHILDREN,
+			task->parent_depends_on);
+}
+
+/* Actual body of GOMP_PLUGIN_target_task_completion that is executed
+   with team->task_lock held, or is executed in the thread that called
+   gomp_target_task_fn if GOMP_PLUGIN_target_task_completion has been
+   run before it acquires team->task_lock.  */
+
+static void
+gomp_target_task_completion (struct gomp_team *team, struct gomp_task *task)
+{
+  struct gomp_task *parent = task->parent;
   if (parent)
+    priority_queue_move_task_first (PQ_CHILDREN, &parent->children_queue,
+				    task);
+
+  struct gomp_taskgroup *taskgroup = task->taskgroup;
+  if (taskgroup)
+    priority_queue_move_task_first (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
+				    task);
+
+  priority_queue_insert (PQ_TEAM, &team->task_queue, task, task->priority,
+			 PRIORITY_INSERT_BEGIN, false,
+			 task->parent_depends_on);
+  task->kind = GOMP_TASK_WAITING;
+  if (parent && parent->taskwait)
     {
-      if (parent->children == child_task)
-	parent->children = child_task->next_child;
-      if (__builtin_expect (child_task->parent_depends_on, 0)
-	  && parent->taskwait->last_parent_depends_on == child_task)
-	{
-	  if (child_task->prev_child->kind == GOMP_TASK_WAITING
-	      && child_task->prev_child->parent_depends_on)
-	    parent->taskwait->last_parent_depends_on = child_task->prev_child;
-	  else
-	    parent->taskwait->last_parent_depends_on = NULL;
+      if (parent->taskwait->in_taskwait)
+	{
+	  /* One more task has had its dependencies met.
+	     Inform any waiters.  */
+	  parent->taskwait->in_taskwait = false;
+	  gomp_sem_post (&parent->taskwait->taskwait_sem);
	}
+      else if (parent->taskwait->in_depend_wait)
+	{
+	  /* One more task has had its dependencies met.
+	     Inform any waiters.  */
+	  parent->taskwait->in_depend_wait = false;
+	  gomp_sem_post (&parent->taskwait->taskwait_sem);
+	}
+    }
+  if (taskgroup && taskgroup->in_taskgroup_wait)
+    {
+      /* One more task has had its dependencies met.
+	 Inform any waiters.  */
+      taskgroup->in_taskgroup_wait = false;
+      gomp_sem_post (&taskgroup->taskgroup_sem);
     }
-  if (taskgroup && taskgroup->children == child_task)
-    taskgroup->children = child_task->next_taskgroup;
-  child_task->prev_queue->next_queue = child_task->next_queue;
-  child_task->next_queue->prev_queue = child_task->prev_queue;
-  if (team->task_queue == child_task)
+
+  ++team->task_queued_count;
+  gomp_team_barrier_set_task_pending (&team->barrier);
+  /* I'm afraid this can't be done after releasing team->task_lock,
+     as gomp_target_task_completion is run from unrelated thread and
+     therefore in between gomp_mutex_unlock and gomp_team_barrier_wake
+     the team could be gone already.  */
+  if (team->nthreads > team->task_running_count)
+    gomp_team_barrier_wake (&team->barrier, 1);
+}
+
+/* Signal that a target task TTASK has completed the asynchronously
+   running phase and should be requeued as a task to handle the
+   variable unmapping.  */
+
+void
+GOMP_PLUGIN_target_task_completion (void *data)
+{
+  struct gomp_target_task *ttask = (struct gomp_target_task *) data;
+  struct gomp_task *task = ttask->task;
+  struct gomp_team *team = ttask->team;
+
+  gomp_mutex_lock (&team->task_lock);
+  if (ttask->state == GOMP_TARGET_TASK_READY_TO_RUN)
     {
-      if (child_task->next_queue != child_task)
-	team->task_queue = child_task->next_queue;
+      ttask->state = GOMP_TARGET_TASK_FINISHED;
+      gomp_mutex_unlock (&team->task_lock);
+      return;
+    }
+  ttask->state = GOMP_TARGET_TASK_FINISHED;
+  gomp_target_task_completion (team, task);
+  gomp_mutex_unlock (&team->task_lock);
+}
+
+static void gomp_task_run_post_handle_depend_hash (struct gomp_task *);
+
+/* Called for nowait target tasks.  */
+
+bool
+gomp_create_target_task (struct gomp_device_descr *devicep,
+			 void (*fn) (void *), size_t mapnum, void **hostaddrs,
+			 size_t *sizes, unsigned short *kinds,
+			 unsigned int flags, void **depend, void **args,
+			 enum gomp_target_task_state state)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_team *team = thr->ts.team;
+
+  /* If parallel or taskgroup has been cancelled, don't start new tasks.  */
+  if (team
+      && (gomp_team_barrier_cancelled (&team->barrier)
+	  || (thr->task->taskgroup && thr->task->taskgroup->cancelled)))
+    return true;
+
+  struct gomp_target_task *ttask;
+  struct gomp_task *task;
+  struct gomp_task *parent = thr->task;
+  struct gomp_taskgroup *taskgroup = parent->taskgroup;
+  bool do_wake;
+  size_t depend_size = 0;
+  uintptr_t depend_cnt = 0;
+  size_t tgt_align = 0, tgt_size = 0;
+
+  if (depend != NULL)
+    {
+      depend_cnt = (uintptr_t) depend[0];
+      depend_size = depend_cnt * sizeof (struct gomp_task_depend_entry);
+    }
+  if (fn)
+    {
+      /* GOMP_MAP_FIRSTPRIVATE need to be copied first, as they are
+	 firstprivate on the target task.  */
+      size_t i;
+      for (i = 0; i < mapnum; i++)
+	if ((kinds[i] & 0xff) == GOMP_MAP_FIRSTPRIVATE)
+	  {
+	    size_t align = (size_t) 1 << (kinds[i] >> 8);
+	    if (tgt_align < align)
+	      tgt_align = align;
+	    tgt_size = (tgt_size + align - 1) & ~(align - 1);
+	    tgt_size += sizes[i];
+	  }
+      if (tgt_align)
+	tgt_size += tgt_align - 1;
       else
-	team->task_queue = NULL;
+	tgt_size = 0;
     }
+
+  task = gomp_malloc (sizeof (*task) + depend_size
+		      + sizeof (*ttask)
+		      + mapnum * (sizeof (void *) + sizeof (size_t)
+				  + sizeof (unsigned short))
+		      + tgt_size);
+  gomp_init_task (task, parent, gomp_icv (false));
+  task->priority = 0;
+  task->kind = GOMP_TASK_WAITING;
+  task->in_tied_task = parent->in_tied_task;
+  task->taskgroup = taskgroup;
+  ttask = (struct gomp_target_task *) &task->depend[depend_cnt];
+  ttask->devicep = devicep;
+  ttask->fn = fn;
+  ttask->mapnum = mapnum;
+  ttask->args = args;
+  memcpy (ttask->hostaddrs, hostaddrs, mapnum * sizeof (void *));
+  ttask->sizes = (size_t *) &ttask->hostaddrs[mapnum];
+  memcpy (ttask->sizes, sizes, mapnum * sizeof (size_t));
+  ttask->kinds = (unsigned short *) &ttask->sizes[mapnum];
+  memcpy (ttask->kinds, kinds, mapnum * sizeof (unsigned short));
+  if (tgt_align)
+    {
+      char *tgt = (char *) &ttask->kinds[mapnum];
+      size_t i;
+      uintptr_t al = (uintptr_t) tgt & (tgt_align - 1);
+      if (al)
+	tgt += tgt_align - al;
+      tgt_size = 0;
+      for (i = 0; i < mapnum; i++)
+	if ((kinds[i] & 0xff) == GOMP_MAP_FIRSTPRIVATE)
+	  {
+	    size_t align = (size_t) 1 << (kinds[i] >> 8);
+	    tgt_size = (tgt_size + align - 1) & ~(align - 1);
+	    memcpy (tgt + tgt_size, hostaddrs[i], sizes[i]);
+	    ttask->hostaddrs[i] = tgt + tgt_size;
+	    tgt_size = tgt_size + sizes[i];
+	  }
+    }
+  ttask->flags = flags;
+  ttask->state = state;
+  ttask->task = task;
+  ttask->team = team;
+  task->fn = NULL;
+  task->fn_data = ttask;
+  task->final_task = 0;
+  gomp_mutex_lock (&team->task_lock);
+  /* If parallel or taskgroup has been cancelled, don't start new tasks.  */
+  if (__builtin_expect (gomp_team_barrier_cancelled (&team->barrier)
+			|| (taskgroup && taskgroup->cancelled), 0))
+    {
+      gomp_mutex_unlock (&team->task_lock);
+      gomp_finish_task (task);
+      free (task);
+      return true;
+    }
+  if (depend_size)
+    {
+      gomp_task_handle_depend (task, parent, depend);
+      if (task->num_dependees)
+	{
+	  if (taskgroup)
+	    taskgroup->num_children++;
+	  gomp_mutex_unlock (&team->task_lock);
+	  return true;
+	}
+    }
+  if (state == GOMP_TARGET_TASK_DATA)
+    {
+      gomp_task_run_post_handle_depend_hash (task);
+      gomp_mutex_unlock (&team->task_lock);
+      gomp_finish_task (task);
+      free (task);
+      return false;
+    }
+  if (taskgroup)
+    taskgroup->num_children++;
+  /* For async offloading, if we don't need to wait for dependencies,
+     run the gomp_target_task_fn right away, essentially schedule the
+     mapping part of the task in the current thread.  */
+  if (devicep != NULL
+      && (devicep->capabilities & GOMP_OFFLOAD_CAP_OPENMP_400))
+    {
+      priority_queue_insert (PQ_CHILDREN, &parent->children_queue, task, 0,
+			     PRIORITY_INSERT_END,
+			     /*adjust_parent_depends_on=*/false,
+			     task->parent_depends_on);
+      if (taskgroup)
+	priority_queue_insert (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
+			       task, 0, PRIORITY_INSERT_END,
+			       /*adjust_parent_depends_on=*/false,
+			       task->parent_depends_on);
+      task->pnode[PQ_TEAM].next = NULL;
+      task->pnode[PQ_TEAM].prev = NULL;
+      task->kind = GOMP_TASK_TIED;
+      ++team->task_count;
+      gomp_mutex_unlock (&team->task_lock);
+
+      thr->task = task;
+      gomp_target_task_fn (task->fn_data);
+      thr->task = parent;
+
+      gomp_mutex_lock (&team->task_lock);
+      task->kind = GOMP_TASK_ASYNC_RUNNING;
+      /* If GOMP_PLUGIN_target_task_completion has run already
+	 in between gomp_target_task_fn and the mutex lock,
+	 perform the requeuing here.  */
+      if (ttask->state == GOMP_TARGET_TASK_FINISHED)
+	gomp_target_task_completion (team, task);
+      else
+	ttask->state = GOMP_TARGET_TASK_RUNNING;
+      gomp_mutex_unlock (&team->task_lock);
+      return true;
+    }
+  priority_queue_insert (PQ_CHILDREN, &parent->children_queue, task, 0,
+			 PRIORITY_INSERT_BEGIN,
+			 /*adjust_parent_depends_on=*/false,
+			 task->parent_depends_on);
+  if (taskgroup)
+    priority_queue_insert (PQ_TASKGROUP, &taskgroup->taskgroup_queue, task, 0,
+			   PRIORITY_INSERT_BEGIN,
+			   /*adjust_parent_depends_on=*/false,
+			   task->parent_depends_on);
+  priority_queue_insert (PQ_TEAM, &team->task_queue, task, 0,
+			 PRIORITY_INSERT_END,
+			 /*adjust_parent_depends_on=*/false,
+			 task->parent_depends_on);
+  ++team->task_count;
+  ++team->task_queued_count;
+  gomp_team_barrier_set_task_pending (&team->barrier);
+  do_wake = team->task_running_count + !parent->in_tied_task
+	    < team->nthreads;
+  gomp_mutex_unlock (&team->task_lock);
+  if (do_wake)
+    gomp_team_barrier_wake (&team->barrier, 1);
+  return true;
+}
+
+/* Given a parent_depends_on task in LIST, move it to the front of its
+   priority so it is run as soon as possible.
+
+   Care is taken to update the list's LAST_PARENT_DEPENDS_ON field.
+
+   We rearrange the queue such that all parent_depends_on tasks are
+   first, and last_parent_depends_on points to the last such task we
+   rearranged.  For example, given the following tasks in a queue
+   where PD[123] are the parent_depends_on tasks:
+
+	task->children
+	|
+	V
+	C1 -> C2 -> C3 -> PD1 -> PD2 -> PD3 -> C4
+
+	We rearrange such that:
+
+	task->children
+	|	       +--- last_parent_depends_on
+	|	       |
+	V	       V
+	PD1 -> PD2 -> PD3 -> C1 -> C2 -> C3 -> C4.  */
+
+static void inline
+priority_list_upgrade_task (struct priority_list *list,
+			    struct priority_node *node)
+{
+  struct priority_node *last_parent_depends_on
+    = list->last_parent_depends_on;
+  if (last_parent_depends_on)
+    {
+      node->prev->next = node->next;
+      node->next->prev = node->prev;
+      node->prev = last_parent_depends_on;
+      node->next = last_parent_depends_on->next;
+      node->prev->next = node;
+      node->next->prev = node;
+    }
+  else if (node != list->tasks)
+    {
+      node->prev->next = node->next;
+      node->next->prev = node->prev;
+      node->prev = list->tasks->prev;
+      node->next = list->tasks;
+      list->tasks = node;
+      node->prev->next = node;
+      node->next->prev = node;
+    }
+  list->last_parent_depends_on = node;
+}
+
+/* Given a parent_depends_on TASK in its parent's children_queue, move
+   it to the front of its priority so it is run as soon as possible.
+
+   PARENT is passed as an optimization.
+
+   (This function could be defined in priority_queue.c, but we want it
+   inlined, and putting it in priority_queue.h is not an option, given
+   that gomp_task has not been properly defined at that point).  */
+
+static void inline
+priority_queue_upgrade_task (struct gomp_task *task,
+			     struct gomp_task *parent)
+{
+  struct priority_queue *head = &parent->children_queue;
+  struct priority_node *node = &task->pnode[PQ_CHILDREN];
+#if _LIBGOMP_CHECKING_
+  if (!task->parent_depends_on)
+    gomp_fatal ("priority_queue_upgrade_task: task must be a "
+		"parent_depends_on task");
+  if (!priority_queue_task_in_queue_p (PQ_CHILDREN, head, task))
+    gomp_fatal ("priority_queue_upgrade_task: cannot find task=%p", task);
+#endif
+  if (priority_queue_multi_p (head))
+    {
+      struct priority_list *list
+	= priority_queue_lookup_priority (head, task->priority);
+      priority_list_upgrade_task (list, node);
+    }
+  else
+    priority_list_upgrade_task (&head->l, node);
+}
+
+/* Given a CHILD_TASK in LIST that is about to be executed, move it out of
+   the way in LIST so that other tasks can be considered for
+   execution.  LIST contains tasks of type TYPE.
+
+   Care is taken to update the queue's LAST_PARENT_DEPENDS_ON field
+   if applicable.  */
+
+static void inline
+priority_list_downgrade_task (enum priority_queue_type type,
+			      struct priority_list *list,
+			      struct gomp_task *child_task)
+{
+  struct priority_node *node = task_to_priority_node (type, child_task);
+  if (list->tasks == node)
+    list->tasks = node->next;
+  else if (node->next != list->tasks)
+    {
+      /* The task in NODE is about to become TIED and TIED tasks
+	 cannot come before WAITING tasks.  If we're about to
+	 leave the queue in such an indeterminate state, rewire
+	 things appropriately.  However, a TIED task at the end is
+	 perfectly fine.  */
+      struct gomp_task *next_task = priority_node_to_task (type, node->next);
+      if (next_task->kind == GOMP_TASK_WAITING)
+	{
+	  /* Remove from list.  */
+	  node->prev->next = node->next;
+	  node->next->prev = node->prev;
+	  /* Rewire at the end.  */
+	  node->next = list->tasks;
+	  node->prev = list->tasks->prev;
+	  list->tasks->prev->next = node;
+	  list->tasks->prev = node;
+	}
+    }
+
+  /* If the current task is the last_parent_depends_on for its
+     priority, adjust last_parent_depends_on appropriately.  */
+  if (__builtin_expect (child_task->parent_depends_on, 0)
+      && list->last_parent_depends_on == node)
+    {
+      struct gomp_task *prev_child = priority_node_to_task (type, node->prev);
+      if (node->prev != node
+	  && prev_child->kind == GOMP_TASK_WAITING
+	  && prev_child->parent_depends_on)
+	list->last_parent_depends_on = node->prev;
+      else
+	{
+	  /* There are no more parent_depends_on entries waiting
+	     to run, clear the list.  */
+	  list->last_parent_depends_on = NULL;
+	}
+    }
+}
+
+/* Given a TASK in HEAD that is about to be executed, move it out of
+   the way so that other tasks can be considered for execution.  HEAD
+   contains tasks of type TYPE.
+
+   Care is taken to update the queue's LAST_PARENT_DEPENDS_ON field
+   if applicable.
+
+   (This function could be defined in priority_queue.c, but we want it
+   inlined, and putting it in priority_queue.h is not an option, given
+   that gomp_task has not been properly defined at that point).  */
+
+static void inline
+priority_queue_downgrade_task (enum priority_queue_type type,
+			       struct priority_queue *head,
+			       struct gomp_task *task)
+{
+#if _LIBGOMP_CHECKING_
+  if (!priority_queue_task_in_queue_p (type, head, task))
+    gomp_fatal ("Attempt to downgrade missing task %p", task);
+#endif
+  if (priority_queue_multi_p (head))
+    {
+      struct priority_list *list
+	= priority_queue_lookup_priority (head, task->priority);
+      priority_list_downgrade_task (type, list, task);
+    }
+  else
+    priority_list_downgrade_task (type, &head->l, task);
+}
+
+/* Setup CHILD_TASK to execute.  This is done by setting the task to
+   TIED, and updating all relevant queues so that CHILD_TASK is no
+   longer chosen for scheduling.  Also, remove CHILD_TASK from the
+   overall team task queue entirely.
+
+   Return TRUE if task or its containing taskgroup has been
+   cancelled.  */
+
+static inline bool
+gomp_task_run_pre (struct gomp_task *child_task, struct gomp_task *parent,
+		   struct gomp_team *team)
+{
+#if _LIBGOMP_CHECKING_
+  if (child_task->parent)
+    priority_queue_verify (PQ_CHILDREN,
+			   &child_task->parent->children_queue, true);
+  if (child_task->taskgroup)
+    priority_queue_verify (PQ_TASKGROUP,
+			   &child_task->taskgroup->taskgroup_queue, false);
+  priority_queue_verify (PQ_TEAM, &team->task_queue, false);
+#endif
+
+  /* Task is about to go tied, move it out of the way.  */
+  if (parent)
+    priority_queue_downgrade_task (PQ_CHILDREN, &parent->children_queue,
+				   child_task);
+
+  /* Task is about to go tied, move it out of the way.  */
+  struct gomp_taskgroup *taskgroup = child_task->taskgroup;
+  if (taskgroup)
+    priority_queue_downgrade_task (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
+				   child_task);
+
+  priority_queue_remove (PQ_TEAM, &team->task_queue, child_task,
+			 MEMMODEL_RELAXED);
+  child_task->pnode[PQ_TEAM].next = NULL;
+  child_task->pnode[PQ_TEAM].prev = NULL;
   child_task->kind = GOMP_TASK_TIED;
+
   if (--team->task_queued_count == 0)
     gomp_team_barrier_clear_task_pending (&team->barrier);
   if ((gomp_team_barrier_cancelled (&team->barrier)
@@ -478,6 +1020,14 @@ gomp_task_run_post_handle_depend_hash (s
       }
 }

+/* After a CHILD_TASK has been run, adjust the dependency queue for
+   each task that depends on CHILD_TASK, to record the fact that there
+   is one less dependency to worry about.  If a task that depended on
+   CHILD_TASK now has no dependencies, place it in the various queues
+   so it gets scheduled to run.
+
+   TEAM is the team to which CHILD_TASK belongs to.  */
+
 static size_t
 gomp_task_run_post_handle_dependers (struct gomp_task *child_task,
				     struct gomp_team *team)
@@ -487,91 +1037,60 @@ gomp_task_run_post_handle_dependers (str
   for (i = 0; i < count; i++)
     {
       struct gomp_task *task = child_task->dependers->elem[i];
+
+      /* CHILD_TASK satisfies a dependency for TASK.  Keep track of
+	 TASK's remaining dependencies.  Once TASK has no other
+	 depenencies, put it into the various queues so it will get
+	 scheduled for execution.  */
       if (--task->num_dependees != 0)
	continue;

       struct gomp_taskgroup *taskgroup = task->taskgroup;
       if (parent)
	{
-	  if (parent->children)
-	    {
-	      /* If parent is in gomp_task_maybe_wait_for_dependencies
-		 and it doesn't need to wait for this task, put it after
-		 all ready to run tasks it needs to wait for.  */
-	      if (parent->taskwait && parent->taskwait->last_parent_depends_on
-		  && !task->parent_depends_on)
-		{
-		  struct gomp_task *last_parent_depends_on
-		    = parent->taskwait->last_parent_depends_on;
-		  task->next_child = last_parent_depends_on->next_child;
-		  task->prev_child = last_parent_depends_on;
-		}
-	      else
-		{
-		  task->next_child = parent->children;
-		  task->prev_child = parent->children->prev_child;
-		  parent->children = task;
-		}
-	      task->next_child->prev_child = task;
-	      task->prev_child->next_child = task;
-	    }
-	  else
-	    {
-	      task->next_child = task;
-	      task->prev_child = task;
-	      parent->children = task;
-	    }
+	  priority_queue_insert (PQ_CHILDREN, &parent->children_queue,
+				 task, task->priority,
+				 PRIORITY_INSERT_BEGIN,
+				 /*adjust_parent_depends_on=*/true,
+				 task->parent_depends_on);
	  if (parent->taskwait)
	    {
	      if (parent->taskwait->in_taskwait)
		{
+		  /* One more task has had its dependencies met.
+		     Inform any waiters.  */
		  parent->taskwait->in_taskwait = false;
		  gomp_sem_post (&parent->taskwait->taskwait_sem);
		}
	      else if (parent->taskwait->in_depend_wait)
		{
+		  /* One more task has had its dependencies met.
+		     Inform any waiters.  */
		  parent->taskwait->in_depend_wait = false;
		  gomp_sem_post (&parent->taskwait->taskwait_sem);
		}
-	      if (parent->taskwait->last_parent_depends_on == NULL
-		  && task->parent_depends_on)
-		parent->taskwait->last_parent_depends_on = task;
	    }
	}
       if (taskgroup)
	{
-	  if (taskgroup->children)
-	    {
-	      task->next_taskgroup = taskgroup->children;
-	      task->prev_taskgroup = taskgroup->children->prev_taskgroup;
-	      task->next_taskgroup->prev_taskgroup = task;
-	      task->prev_taskgroup->next_taskgroup = task;
-	    }
-	  else
-	    {
-	      task->next_taskgroup = task;
-	      task->prev_taskgroup = task;
-	    }
-	  taskgroup->children = task;
+	  priority_queue_insert (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
+				 task, task->priority,
+				 PRIORITY_INSERT_BEGIN,
+				 /*adjust_parent_depends_on=*/false,
+				 task->parent_depends_on);
	  if (taskgroup->in_taskgroup_wait)
	    {
+	      /* One more task has had its dependencies met.
+		 Inform any waiters.  */
	      taskgroup->in_taskgroup_wait = false;
	      gomp_sem_post (&taskgroup->taskgroup_sem);
	    }
	}
-      if (team->task_queue)
-	{
-	  task->next_queue = team->task_queue;
-	  task->prev_queue = team->task_queue->prev_queue;
-	  task->next_queue->prev_queue = task;
-	  task->prev_queue->next_queue = task;
-	}
-      else
-	{
-	  task->next_queue = task;
-	  task->prev_queue = task;
-	  team->task_queue = task;
-	}
+      priority_queue_insert (PQ_TEAM, &team->task_queue,
+			     task, task->priority,
+			     PRIORITY_INSERT_END,
+			     /*adjust_parent_depends_on=*/false,
+			     task->parent_depends_on);
       ++team->task_count;
       ++team->task_queued_count;
       ++ret;
@@ -601,12 +1120,18 @@ gomp_task_run_post_handle_depend (struct
   return gomp_task_run_post_handle_dependers (child_task, team);
 }

+/* Remove CHILD_TASK from its parent.  */
+
 static inline void
 gomp_task_run_post_remove_parent (struct gomp_task *child_task)
 {
   struct gomp_task *parent = child_task->parent;
   if (parent == NULL)
     return;
+
+  /* If this was the last task the parent was depending on,
+     synchronize with gomp_task_maybe_wait_for_dependencies so it can
+     clean up and return.  */
   if (__builtin_expect (child_task->parent_depends_on, 0)
       && --parent->taskwait->n_depend == 0
       && parent->taskwait->in_depend_wait)
@@ -614,36 +1139,31 @@ gomp_task_run_post_remove_parent (struct
       parent->taskwait->in_depend_wait = false;
       gomp_sem_post (&parent->taskwait->taskwait_sem);
     }
-  child_task->prev_child->next_child = child_task->next_child;
-  child_task->next_child->prev_child = child_task->prev_child;
-  if (parent->children != child_task)
-    return;
-  if (child_task->next_child != child_task)
-    parent->children = child_task->next_child;
-  else
+
+  if (priority_queue_remove (PQ_CHILDREN, &parent->children_queue,
+			     child_task, MEMMODEL_RELEASE)
+      && parent->taskwait && parent->taskwait->in_taskwait)
     {
-      /* We access task->children in GOMP_taskwait
-	 outside of the task lock mutex region, so
-	 need a release barrier here to ensure memory
-	 written by child_task->fn above is flushed
-	 before the NULL is written.  */
-      __atomic_store_n (&parent->children, NULL, MEMMODEL_RELEASE);
-      if (parent->taskwait && parent->taskwait->in_taskwait)
-	{
-	  parent->taskwait->in_taskwait = false;
-	  gomp_sem_post (&parent->taskwait->taskwait_sem);
-	}
+      parent->taskwait->in_taskwait = false;
+      gomp_sem_post (&parent->taskwait->taskwait_sem);
     }
+  child_task->pnode[PQ_CHILDREN].next = NULL;
+  child_task->pnode[PQ_CHILDREN].prev = NULL;
 }

+/* Remove CHILD_TASK from its taskgroup.  */
+
 static inline void
 gomp_task_run_post_remove_taskgroup (struct gomp_task *child_task)
 {
   struct gomp_taskgroup *taskgroup = child_task->taskgroup;
   if (taskgroup == NULL)
     return;
-  child_task->prev_taskgroup->next_taskgroup = child_task->next_taskgroup;
-  child_task->next_taskgroup->prev_taskgroup = child_task->prev_taskgroup;
+  bool empty = priority_queue_remove (PQ_TASKGROUP,
+				      &taskgroup->taskgroup_queue,
+				      child_task, MEMMODEL_RELAXED);
+  child_task->pnode[PQ_TASKGROUP].next = NULL;
+  child_task->pnode[PQ_TASKGROUP].prev = NULL;
   if (taskgroup->num_children > 1)
     --taskgroup->num_children;
   else
@@ -655,18 +1175,10 @@ gomp_task_run_post_remove_taskgroup (str
	 before the NULL is written.  */
       __atomic_store_n (&taskgroup->num_children, 0, MEMMODEL_RELEASE);
     }
-  if (taskgroup->children != child_task)
-    return;
-  if (child_task->next_taskgroup != child_task)
-    taskgroup->children = child_task->next_taskgroup;
-  else
+  if (empty && taskgroup->in_taskgroup_wait)
     {
-      taskgroup->children = NULL;
-      if (taskgroup->in_taskgroup_wait)
-	{
-	  taskgroup->in_taskgroup_wait = false;
-	  gomp_sem_post (&taskgroup->taskgroup_sem);
-	}
+      taskgroup->in_taskgroup_wait = false;
+      gomp_sem_post (&taskgroup->taskgroup_sem);
     }
 }

@@ -696,11 +1208,15 @@ gomp_barrier_handle_tasks (gomp_barrier_
   while (1)
     {
       bool cancelled = false;
-      if (team->task_queue != NULL)
+      if (!priority_queue_empty_p (&team->task_queue, MEMMODEL_RELAXED))
	{
-	  child_task = team->task_queue;
+	  bool ignored;
+	  child_task
+	    = priority_queue_next_task (PQ_TEAM, &team->task_queue,
+					PQ_IGNORED, NULL,
+					&ignored);
	  cancelled = gomp_task_run_pre (child_task, child_task->parent,
-					 child_task->taskgroup, team);
+					 team);
	  if (__builtin_expect (cancelled, 0))
	    {
	      if (to_free)
@@ -729,7 +1245,29 @@ gomp_barrier_handle_tasks (gomp_barrier_
       if (child_task)
	{
	  thr->task = child_task;
-	  child_task->fn (child_task->fn_data);
+	  if (__builtin_expect (child_task->fn == NULL, 0))
+	    {
+	      if (gomp_target_task_fn (child_task->fn_data))
+		{
+		  thr->task = task;
+		  gomp_mutex_lock (&team->task_lock);
+		  child_task->kind = GOMP_TASK_ASYNC_RUNNING;
+		  team->task_running_count--;
+		  struct gomp_target_task *ttask
+		    = (struct gomp_target_task *) child_task->fn_data;
+		  /* If GOMP_PLUGIN_target_task_completion has run already
+		     in between gomp_target_task_fn and the mutex lock,
+		     perform the requeuing here.  */
+		  if (ttask->state == GOMP_TARGET_TASK_FINISHED)
+		    gomp_target_task_completion (team, child_task);
+		  else
+		    ttask->state = GOMP_TARGET_TASK_RUNNING;
+		  child_task = NULL;
+		  continue;
+		}
+	    }
+	  else
+	    child_task->fn (child_task->fn_data);
	  thr->task = task;
	}
       else
@@ -741,7 +1279,7 @@ gomp_barrier_handle_tasks (gomp_barrier_
	  size_t new_tasks
	    = gomp_task_run_post_handle_depend (child_task, team);
	  gomp_task_run_post_remove_parent (child_task);
-	  gomp_clear_parent (child_task->children);
+	  gomp_clear_parent (&child_task->children_queue);
	  gomp_task_run_post_remove_taskgroup (child_task);
	  to_free = child_task;
	  child_task = NULL;
@@ -765,7 +1303,9 @@ gomp_barrier_handle_tasks (gomp_barrier_
     }
 }

-/* Called when encountering a taskwait directive.  */
+/* Called when encountering a taskwait directive.
+
+   Wait for all children of the current task.  */

 void
 GOMP_taskwait (void)
@@ -785,15 +1325,16 @@ GOMP_taskwait (void)
      child thread task work function are seen before we exit from
      GOMP_taskwait.  */
   if (task == NULL
-      || __atomic_load_n (&task->children, MEMMODEL_ACQUIRE) == NULL)
+      || priority_queue_empty_p (&task->children_queue, MEMMODEL_ACQUIRE))
     return;

   memset (&taskwait, 0, sizeof (taskwait));
+  bool child_q = false;
   gomp_mutex_lock (&team->task_lock);
   while (1)
     {
       bool cancelled = false;
-      if (task->children == NULL)
+      if (priority_queue_empty_p (&task->children_queue, MEMMODEL_RELAXED))
	{
	  bool destroy_taskwait = task->taskwait != NULL;
	  task->taskwait = NULL;
@@ -807,12 +1348,14 @@ GOMP_taskwait (void)
	    gomp_sem_destroy (&taskwait.taskwait_sem);
	  return;
	}
-      if (task->children->kind == GOMP_TASK_WAITING)
+      struct gomp_task *next_task
+	= priority_queue_next_task (PQ_CHILDREN, &task->children_queue,
+				    PQ_TEAM, &team->task_queue, &child_q);
+      if (next_task->kind == GOMP_TASK_WAITING)
	{
-	  child_task = task->children;
+	  child_task = next_task;
	  cancelled
-	    = gomp_task_run_pre (child_task, task, child_task->taskgroup,
-				 team);
+	    = gomp_task_run_pre (child_task, task, team);
	  if (__builtin_expect (cancelled, 0))
	    {
	      if (to_free)
@@ -826,8 +1369,10 @@ GOMP_taskwait (void)
	}
       else
	{
-	  /* All tasks we are waiting for are already running
-	     in other threads.  Wait for them.  */
+	/* All tasks we are waiting for are either running in other
+	   threads, or they are tasks that have not had their
+	   dependencies met (so they're not even in the queue).  Wait
+	   for them.  */
	  if (task->taskwait == NULL)
	    {
	      taskwait.in_depend_wait = false;
@@ -851,7 +1396,28 @@ GOMP_taskwait (void)
       if (child_task)
	{
	  thr->task = child_task;
-	  child_task->fn (child_task->fn_data);
+	  if (__builtin_expect (child_task->fn == NULL, 0))
+	    {
+	      if (gomp_target_task_fn (child_task->fn_data))
+		{
+		  thr->task = task;
+		  gomp_mutex_lock (&team->task_lock);
+		  child_task->kind = GOMP_TASK_ASYNC_RUNNING;
+		  struct gomp_target_task *ttask
+		    = (struct gomp_target_task *) child_task->fn_data;
+		  /* If GOMP_PLUGIN_target_task_completion has run already
+		     in between gomp_target_task_fn and the mutex lock,
+		     perform the requeuing here.  */
+		  if (ttask->state == GOMP_TARGET_TASK_FINISHED)
+		    gomp_target_task_completion (team, child_task);
+		  else
+		    ttask->state = GOMP_TARGET_TASK_RUNNING;
+		  child_task = NULL;
+		  continue;
+		}
+	    }
+	  else
+	    child_task->fn (child_task->fn_data);
	  thr->task = task;
	}
       else
@@ -862,17 +1428,19 @@ GOMP_taskwait (void)
	 finish_cancelled:;
	  size_t new_tasks
	    = gomp_task_run_post_handle_depend (child_task, team);
-	  child_task->prev_child->next_child = child_task->next_child;
-	  child_task->next_child->prev_child = child_task->prev_child;
-	  if (task->children == child_task)
-	    {
-	      if (child_task->next_child != child_task)
-		task->children = child_task->next_child;
-	      else
-		task->children = NULL;
+
+	  if (child_q)
+	    {
+	      priority_queue_remove (PQ_CHILDREN, &task->children_queue,
+				     child_task, MEMMODEL_RELAXED);
+	      child_task->pnode[PQ_CHILDREN].next = NULL;
+	      child_task->pnode[PQ_CHILDREN].prev = NULL;
	    }
-	  gomp_clear_parent (child_task->children);
+
+	  gomp_clear_parent (&child_task->children_queue);
+
	  gomp_task_run_post_remove_taskgroup (child_task);
+
	  to_free = child_task;
	  child_task = NULL;
	  team->task_count--;
@@ -887,10 +1455,20 @@ GOMP_taskwait (void)
     }
 }

-/* This is like GOMP_taskwait, but we only wait for tasks that the
-   upcoming task depends on.  */
+/* An undeferred task is about to run.  Wait for all tasks that this
+   undeferred task depends on.

-static void
+   This is done by first putting all known ready dependencies
+   (dependencies that have their own dependencies met) at the top of
+   the scheduling queues.  Then we iterate through these imminently
+   ready tasks (and possibly other high priority tasks), and run them.
+   If we run out of ready dependencies to execute, we either wait for
+   the reamining dependencies to finish, or wait for them to get
+   scheduled so we can run them.
+
+   DEPEND is as in GOMP_task.  */
+
+void
 gomp_task_maybe_wait_for_dependencies (void **depend)
 {
   struct gomp_thread *thr = gomp_thread ();
@@ -898,7 +1476,6 @@ gomp_task_maybe_wait_for_dependencies (v
   struct gomp_team *team = thr->ts.team;
   struct gomp_task_depend_entry elem, *ent = NULL;
   struct gomp_taskwait taskwait;
-  struct gomp_task *last_parent_depends_on = NULL;
   size_t ndepend = (uintptr_t) depend[0];
   size_t nout = (uintptr_t) depend[1];
   size_t i;
@@ -922,32 +1499,11 @@ gomp_task_maybe_wait_for_dependencies (v
	      {
		tsk->parent_depends_on = true;
		++num_awaited;
+		/* If depenency TSK itself has no dependencies and is
+		   ready to run, move it up front so that we run it as
+		   soon as possible.  */
		if (tsk->num_dependees == 0 && tsk->kind == GOMP_TASK_WAITING)
-		  {
-		    /* If a task we need to wait for is not already
-		       running and is ready to be scheduled, move it
-		       to front, so that we run it as soon as possible.  */
-		    if (last_parent_depends_on)
-		      {
-			tsk->prev_child->next_child = tsk->next_child;
-			tsk->next_child->prev_child = tsk->prev_child;
-			tsk->prev_child = last_parent_depends_on;
-			tsk->next_child = last_parent_depends_on->next_child;
-			tsk->prev_child->next_child = tsk;
-			tsk->next_child->prev_child = tsk;
-		      }
-		    else if (tsk != task->children)
-		      {
-			tsk->prev_child->next_child = tsk->next_child;
-			tsk->next_child->prev_child = tsk->prev_child;
-			tsk->prev_child = task->children;
-			tsk->next_child = task->children->next_child;
-			task->children = tsk;
-			tsk->prev_child->next_child = tsk;
-			tsk->next_child->prev_child = tsk;
-		      }
-		    last_parent_depends_on = tsk;
-		  }
+		  priority_queue_upgrade_task (tsk, task);
	      }
	  }
     }
@@ -959,7 +1515,6 @@ gomp_task_maybe_wait_for_dependencies (v

   memset (&taskwait, 0, sizeof (taskwait));
   taskwait.n_depend = num_awaited;
-  taskwait.last_parent_depends_on = last_parent_depends_on;
   gomp_sem_init (&taskwait.taskwait_sem, 0);
   task->taskwait = &taskwait;

@@ -978,12 +1533,30 @@ gomp_task_maybe_wait_for_dependencies (v
	  gomp_sem_destroy (&taskwait.taskwait_sem);
	  return;
	}
-      if (task->children->kind == GOMP_TASK_WAITING)
+
+      /* Theoretically when we have multiple priorities, we should
+	 chose between the highest priority item in
+	 task->children_queue and team->task_queue here, so we should
+	 use priority_queue_next_task().  However, since we are
+	 running an undeferred task, perhaps that makes all tasks it
+	 depends on undeferred, thus a priority of INF?  This would
+	 make it unnecessary to take anything into account here,
+	 but the dependencies.
+
+	 On the other hand, if we want to use priority_queue_next_task(),
+	 care should be taken to only use priority_queue_remove()
+	 below if the task was actually removed from the children
+	 queue.  */
+      bool ignored;
+      struct gomp_task *next_task
+	= priority_queue_next_task (PQ_CHILDREN, &task->children_queue,
+				    PQ_IGNORED, NULL, &ignored);
+
+      if (next_task->kind == GOMP_TASK_WAITING)
	{
-	  child_task = task->children;
+	  child_task = next_task;
	  cancelled
-	    = gomp_task_run_pre (child_task, task, child_task->taskgroup,
-				 team);
+	    = gomp_task_run_pre (child_task, task, team);
	  if (__builtin_expect (cancelled, 0))
	    {
	      if (to_free)
@@ -996,8 +1569,10 @@ gomp_task_maybe_wait_for_dependencies (v
	    }
	}
       else
-	/* All tasks we are waiting for are already running
-	   in other threads.  Wait for them.  */
+	/* All tasks we are waiting for are either running in other
+	   threads, or they are tasks that have not had their
+	   dependencies met (so they're not even in the queue).  Wait
+	   for them.  */
	taskwait.in_depend_wait = true;
       gomp_mutex_unlock (&team->task_lock);
       if (do_wake)
@@ -1014,7 +1589,28 @@ gomp_task_maybe_wait_for_dependencies (v
       if (child_task)
	{
	  thr->task = child_task;
-	  child_task->fn (child_task->fn_data);
+	  if (__builtin_expect (child_task->fn == NULL, 0))
+	    {
+	      if (gomp_target_task_fn (child_task->fn_data))
+		{
+		  thr->task = task;
+		  gomp_mutex_lock (&team->task_lock);
+		  child_task->kind = GOMP_TASK_ASYNC_RUNNING;
+		  struct gomp_target_task *ttask
+		    = (struct gomp_target_task *) child_task->fn_data;
+		  /* If GOMP_PLUGIN_target_task_completion has run already
+		     in between gomp_target_task_fn and the mutex lock,
+		     perform the requeuing here.  */
+		  if (ttask->state == GOMP_TARGET_TASK_FINISHED)
+		    gomp_target_task_completion (team, child_task);
+		  else
+		    ttask->state = GOMP_TARGET_TASK_RUNNING;
+		  child_task = NULL;
+		  continue;
+		}
+	    }
+	  else
+	    child_task->fn (child_task->fn_data);
	  thr->task = task;
	}
       else
@@ -1027,16 +1623,13 @@ gomp_task_maybe_wait_for_dependencies (v
	    = gomp_task_run_post_handle_depend (child_task, team);
	  if (child_task->parent_depends_on)
	    --taskwait.n_depend;
-	  child_task->prev_child->next_child = child_task->next_child;
-	  child_task->next_child->prev_child = child_task->prev_child;
-	  if (task->children == child_task)
-	    {
-	      if (child_task->next_child != child_task)
-		task->children = child_task->next_child;
-	      else
-		task->children = NULL;
-	    }
-	  gomp_clear_parent (child_task->children);
+
+	  priority_queue_remove (PQ_CHILDREN, &task->children_queue,
+				 child_task, MEMMODEL_RELAXED);
+	  child_task->pnode[PQ_CHILDREN].next = NULL;
+	  child_task->pnode[PQ_CHILDREN].prev = NULL;
+
+	  gomp_clear_parent (&child_task->children_queue);
	  gomp_task_run_post_remove_taskgroup (child_task);
	  to_free = child_task;
	  child_task = NULL;
@@ -1069,14 +1662,14 @@ GOMP_taskgroup_start (void)
   struct gomp_taskgroup *taskgroup;

   /* If team is NULL, all tasks are executed as
-     GOMP_TASK_IFFALSE tasks and thus all children tasks of
+     GOMP_TASK_UNDEFERRED tasks and thus all children tasks of
      taskgroup and their descendant tasks will be finished
      by the time GOMP_taskgroup_end is called.  */
   if (team == NULL)
     return;
   taskgroup = gomp_malloc (sizeof (struct gomp_taskgroup));
   taskgroup->prev = task->taskgroup;
-  taskgroup->children = NULL;
+  priority_queue_init (&taskgroup->taskgroup_queue);
   taskgroup->in_taskgroup_wait = false;
   taskgroup->cancelled = false;
   taskgroup->num_children = 0;
@@ -1098,6 +1691,17 @@ GOMP_taskgroup_end (void)
   if (team == NULL)
     return;
   taskgroup = task->taskgroup;
+  if (__builtin_expect (taskgroup == NULL, 0)
+      && thr->ts.level == 0)
+    {
+      /* This can happen if GOMP_taskgroup_start is called when
+	 thr->ts.team == NULL, but inside of the taskgroup there
+	 is #pragma omp target nowait that creates an implicit
+	 team with a single thread.  In this case, we want to wait
+	 for all outstanding tasks in this team.  */
+      gomp_team_barrier_wait (&team->barrier);
+      return;
+    }

   /* The acquire barrier on load of taskgroup->num_children here
      synchronizes with the write of 0 in gomp_task_run_post_remove_taskgroup.
@@ -1108,19 +1712,25 @@ GOMP_taskgroup_end (void)
   if (__atomic_load_n (&taskgroup->num_children, MEMMODEL_ACQUIRE) == 0)
     goto finish;

+  bool unused;
   gomp_mutex_lock (&team->task_lock);
   while (1)
     {
       bool cancelled = false;
-      if (taskgroup->children == NULL)
+      if (priority_queue_empty_p (&taskgroup->taskgroup_queue,
+				  MEMMODEL_RELAXED))
	{
	  if (taskgroup->num_children)
	    {
-	      if (task->children == NULL)
+	      if (priority_queue_empty_p (&task->children_queue,
+					  MEMMODEL_RELAXED))
		goto do_wait;
-	      child_task = task->children;
-            }
-          else
+	      child_task
+		= priority_queue_next_task (PQ_CHILDREN, &task->children_queue,
+					    PQ_TEAM, &team->task_queue,
+					    &unused);
+	    }
+	  else
	    {
	      gomp_mutex_unlock (&team->task_lock);
	      if (to_free)
@@ -1132,12 +1742,13 @@ GOMP_taskgroup_end (void)
	    }
	}
       else
-	child_task = taskgroup->children;
+	child_task
+	  = priority_queue_next_task (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
+				      PQ_TEAM, &team->task_queue, &unused);
       if (child_task->kind == GOMP_TASK_WAITING)
	{
	  cancelled
-	    = gomp_task_run_pre (child_task, child_task->parent, taskgroup,
-				 team);
+	    = gomp_task_run_pre (child_task, child_task->parent, team);
	  if (__builtin_expect (cancelled, 0))
	    {
	      if (to_free)
@@ -1153,8 +1764,10 @@ GOMP_taskgroup_end (void)
	{
	  child_task = NULL;
	 do_wait:
-	  /* All tasks we are waiting for are already running
-	     in other threads.  Wait for them.  */
+	/* All tasks we are waiting for are either running in other
+	   threads, or they are tasks that have not had their
+	   dependencies met (so they're not even in the queue).  Wait
+	   for them.  */
	  taskgroup->in_taskgroup_wait = true;
	}
       gomp_mutex_unlock (&team->task_lock);
@@ -1172,7 +1785,28 @@ GOMP_taskgroup_end (void)
       if (child_task)
	{
	  thr->task = child_task;
-	  child_task->fn (child_task->fn_data);
+	  if (__builtin_expect (child_task->fn == NULL, 0))
+	    {
+	      if (gomp_target_task_fn (child_task->fn_data))
+		{
+		  thr->task = task;
+		  gomp_mutex_lock (&team->task_lock);
+		  child_task->kind = GOMP_TASK_ASYNC_RUNNING;
+		  struct gomp_target_task *ttask
+		    = (struct gomp_target_task *) child_task->fn_data;
+		  /* If GOMP_PLUGIN_target_task_completion has run already
+		     in between gomp_target_task_fn and the mutex lock,
+		     perform the requeuing here.  */
+		  if (ttask->state == GOMP_TARGET_TASK_FINISHED)
+		    gomp_target_task_completion (team, child_task);
+		  else
+		    ttask->state = GOMP_TARGET_TASK_RUNNING;
+		  child_task = NULL;
+		  continue;
+		}
+	    }
+	  else
+	    child_task->fn (child_task->fn_data);
	  thr->task = task;
	}
       else
@@ -1184,7 +1818,7 @@ GOMP_taskgroup_end (void)
	  size_t new_tasks
	    = gomp_task_run_post_handle_depend (child_task, team);
	  gomp_task_run_post_remove_parent (child_task);
-	  gomp_clear_parent (child_task->children);
+	  gomp_clear_parent (&child_task->children_queue);
	  gomp_task_run_post_remove_taskgroup (child_task);
	  to_free = child_task;
	  child_task = NULL;
--- libgomp/libgomp_g.h.jj	2014-05-15 10:56:31.429532978 +0200
+++ libgomp/libgomp_g.h	2016-07-13 16:57:04.422535521 +0200
@@ -29,6 +29,7 @@
 #define LIBGOMP_G_H 1

 #include <stdbool.h>
+#include <stddef.h>

 /* barrier.c */

@@ -50,6 +51,10 @@ extern bool GOMP_loop_static_start (long
 extern bool GOMP_loop_dynamic_start (long, long, long, long, long *, long *);
 extern bool GOMP_loop_guided_start (long, long, long, long, long *, long *);
 extern bool GOMP_loop_runtime_start (long, long, long, long *, long *);
+extern bool GOMP_loop_nonmonotonic_dynamic_start (long, long, long, long,
+						  long *, long *);
+extern bool GOMP_loop_nonmonotonic_guided_start (long, long, long, long,
+						 long *, long *);

 extern bool GOMP_loop_ordered_static_start (long, long, long, long,
					    long *, long *);
@@ -63,12 +68,23 @@ extern bool GOMP_loop_static_next (long
 extern bool GOMP_loop_dynamic_next (long *, long *);
 extern bool GOMP_loop_guided_next (long *, long *);
 extern bool GOMP_loop_runtime_next (long *, long *);
+extern bool GOMP_loop_nonmonotonic_dynamic_next (long *, long *);
+extern bool GOMP_loop_nonmonotonic_guided_next (long *, long *);

 extern bool GOMP_loop_ordered_static_next (long *, long *);
 extern bool GOMP_loop_ordered_dynamic_next (long *, long *);
 extern bool GOMP_loop_ordered_guided_next (long *, long *);
 extern bool GOMP_loop_ordered_runtime_next (long *, long *);

+extern bool GOMP_loop_doacross_static_start (unsigned, long *, long, long *,
+					     long *);
+extern bool GOMP_loop_doacross_dynamic_start (unsigned, long *, long, long *,
+					      long *);
+extern bool GOMP_loop_doacross_guided_start (unsigned, long *, long, long *,
+					     long *);
+extern bool GOMP_loop_doacross_runtime_start (unsigned, long *, long *,
+					      long *);
+
 extern void GOMP_parallel_loop_static_start (void (*)(void *), void *,
					     unsigned, long, long, long, long);
 extern void GOMP_parallel_loop_dynamic_start (void (*)(void *), void *,
@@ -89,6 +105,12 @@ extern void GOMP_parallel_loop_guided (v
 extern void GOMP_parallel_loop_runtime (void (*)(void *), void *,
					unsigned, long, long, long,
					unsigned);
+extern void GOMP_parallel_loop_nonmonotonic_dynamic (void (*)(void *), void *,
+						     unsigned, long, long,
+						     long, long, unsigned);
+extern void GOMP_parallel_loop_nonmonotonic_guided (void (*)(void *), void *,
+						    unsigned, long, long,
+						    long, long, unsigned);

 extern void GOMP_loop_end (void);
 extern void GOMP_loop_end_nowait (void);
@@ -119,6 +141,18 @@ extern bool GOMP_loop_ull_runtime_start
					 unsigned long long,
					 unsigned long long *,
					 unsigned long long *);
+extern bool GOMP_loop_ull_nonmonotonic_dynamic_start (bool, unsigned long long,
+						      unsigned long long,
+						      unsigned long long,
+						      unsigned long long,
+						      unsigned long long *,
+						      unsigned long long *);
+extern bool GOMP_loop_ull_nonmonotonic_guided_start (bool, unsigned long long,
+						     unsigned long long,
+						     unsigned long long,
+						     unsigned long long,
+						     unsigned long long *,
+						     unsigned long long *);

 extern bool GOMP_loop_ull_ordered_static_start (bool, unsigned long long,
						unsigned long long,
@@ -152,6 +186,10 @@ extern bool GOMP_loop_ull_guided_next (u
				       unsigned long long *);
 extern bool GOMP_loop_ull_runtime_next (unsigned long long *,
					unsigned long long *);
+extern bool GOMP_loop_ull_nonmonotonic_dynamic_next (unsigned long long *,
+						     unsigned long long *);
+extern bool GOMP_loop_ull_nonmonotonic_guided_next (unsigned long long *,
+						    unsigned long long *);

 extern bool GOMP_loop_ull_ordered_static_next (unsigned long long *,
					       unsigned long long *);
@@ -162,10 +200,34 @@ extern bool GOMP_loop_ull_ordered_guided
 extern bool GOMP_loop_ull_ordered_runtime_next (unsigned long long *,
						unsigned long long *);

+extern bool GOMP_loop_ull_doacross_static_start (unsigned,
+						 unsigned long long *,
+						 unsigned long long,
+						 unsigned long long *,
+						 unsigned long long *);
+extern bool GOMP_loop_ull_doacross_dynamic_start (unsigned,
+						  unsigned long long *,
+						  unsigned long long,
+						  unsigned long long *,
+						  unsigned long long *);
+extern bool GOMP_loop_ull_doacross_guided_start (unsigned,
+						 unsigned long long *,
+						 unsigned long long,
+						 unsigned long long *,
+						 unsigned long long *);
+extern bool GOMP_loop_ull_doacross_runtime_start (unsigned,
+						  unsigned long long *,
+						  unsigned long long *,
+						  unsigned long long *);
+
 /* ordered.c */

 extern void GOMP_ordered_start (void);
 extern void GOMP_ordered_end (void);
+extern void GOMP_doacross_post (long *);
+extern void GOMP_doacross_wait (long, ...);
+extern void GOMP_doacross_ull_post (unsigned long long *);
+extern void GOMP_doacross_ull_wait (unsigned long long, ...);

 /* parallel.c */

@@ -178,7 +240,15 @@ extern bool GOMP_cancellation_point (int
 /* task.c */

 extern void GOMP_task (void (*) (void *), void *, void (*) (void *, void *),
-		       long, long, bool, unsigned, void **);
+		       long, long, bool, unsigned, void **, int);
+extern void GOMP_taskloop (void (*) (void *), void *,
+			   void (*) (void *, void *), long, long, unsigned,
+			   unsigned long, int, long, long, long);
+extern void GOMP_taskloop_ull (void (*) (void *), void *,
+			       void (*) (void *, void *), long, long,
+			       unsigned, unsigned long, int,
+			       unsigned long long, unsigned long long,
+			       unsigned long long);
 extern void GOMP_taskwait (void);
 extern void GOMP_taskyield (void);
 extern void GOMP_taskgroup_start (void);
@@ -206,11 +276,38 @@ extern void GOMP_single_copy_end (void *

 extern void GOMP_target (int, void (*) (void *), const void *,
			 size_t, void **, size_t *, unsigned char *);
+extern void GOMP_target_ext (int, void (*) (void *), size_t, void **, size_t *,
+			     unsigned short *, unsigned int, void **, void **);
 extern void GOMP_target_data (int, const void *,
			      size_t, void **, size_t *, unsigned char *);
+extern void GOMP_target_data_ext (int, size_t, void **, size_t *,
+				  unsigned short *);
 extern void GOMP_target_end_data (void);
 extern void GOMP_target_update (int, const void *,
				size_t, void **, size_t *, unsigned char *);
+extern void GOMP_target_update_ext (int, size_t, void **, size_t *,
+				    unsigned short *, unsigned int, void **);
+extern void GOMP_target_enter_exit_data (int, size_t, void **, size_t *,
+					 unsigned short *, unsigned int,
+					 void **);
 extern void GOMP_teams (unsigned int, unsigned int);

+/* oacc-parallel.c */
+
+extern void GOACC_parallel_keyed (int, void (*) (void *), size_t,
+				  void **, size_t *, unsigned short *, ...);
+extern void GOACC_parallel (int, void (*) (void *), size_t, void **, size_t *,
+			    unsigned short *, int, int, int, int, int, ...);
+extern void GOACC_data_start (int, size_t, void **, size_t *,
+			      unsigned short *);
+extern void GOACC_data_end (void);
+extern void GOACC_enter_exit_data (int, size_t, void **,
+				   size_t *, unsigned short *, int, int, ...);
+extern void GOACC_update (int, size_t, void **, size_t *,
+			  unsigned short *, int, int, ...);
+extern void GOACC_wait (int, int, ...);
+extern int GOACC_get_num_threads (void);
+extern int GOACC_get_thread_num (void);
+extern void GOACC_declare (int, size_t, void **, size_t *, unsigned short *);
+
 #endif /* LIBGOMP_G_H */
--- libgomp/libgomp.h.jj	2014-08-01 15:59:49.145188127 +0200
+++ libgomp/libgomp.h	2016-07-14 17:40:24.038243456 +0200
@@ -34,12 +34,35 @@
 #ifndef LIBGOMP_H
 #define LIBGOMP_H 1

+#ifndef _LIBGOMP_CHECKING_
+/* Define to 1 to perform internal sanity checks.  */
+#define _LIBGOMP_CHECKING_ 0
+#endif
+
 #include "config.h"
 #include "gstdint.h"
+#include "libgomp-plugin.h"

 #include <pthread.h>
 #include <stdbool.h>
 #include <stdlib.h>
+#include <stdarg.h>
+
+/* Needed for memset in priority_queue.c.  */
+#if _LIBGOMP_CHECKING_
+# ifdef STRING_WITH_STRINGS
+#  include <string.h>
+#  include <strings.h>
+# else
+#  ifdef HAVE_STRING_H
+#   include <string.h>
+#  else
+#   ifdef HAVE_STRINGS_H
+#    include <strings.h>
+#   endif
+#  endif
+# endif
+#endif

 #ifdef HAVE_ATTRIBUTE_VISIBILITY
 # pragma GCC visibility push(hidden)
@@ -56,6 +79,44 @@ enum memmodel
   MEMMODEL_SEQ_CST = 5
 };

+/* alloc.c */
+
+extern void *gomp_malloc (size_t) __attribute__((malloc));
+extern void *gomp_malloc_cleared (size_t) __attribute__((malloc));
+extern void *gomp_realloc (void *, size_t);
+
+/* Avoid conflicting prototypes of alloca() in system headers by using
+   GCC's builtin alloca().  */
+#define gomp_alloca(x)  __builtin_alloca(x)
+
+/* error.c */
+
+extern void gomp_vdebug (int, const char *, va_list);
+extern void gomp_debug (int, const char *, ...)
+	__attribute__ ((format (printf, 2, 3)));
+#define gomp_vdebug(KIND, FMT, VALIST) \
+  do { \
+    if (__builtin_expect (gomp_debug_var, 0)) \
+      (gomp_vdebug) ((KIND), (FMT), (VALIST)); \
+  } while (0)
+#define gomp_debug(KIND, ...) \
+  do { \
+    if (__builtin_expect (gomp_debug_var, 0)) \
+      (gomp_debug) ((KIND), __VA_ARGS__); \
+  } while (0)
+extern void gomp_verror (const char *, va_list);
+extern void gomp_error (const char *, ...)
+	__attribute__ ((format (printf, 1, 2)));
+extern void gomp_vfatal (const char *, va_list)
+	__attribute__ ((noreturn));
+extern void gomp_fatal (const char *, ...)
+	__attribute__ ((noreturn, format (printf, 1, 2)));
+
+struct gomp_task;
+struct gomp_taskgroup;
+struct htab;
+
+#include "priority_queue.h"
 #include "sem.h"
 #include "mutex.h"
 #include "bar.h"
@@ -74,6 +135,44 @@ enum gomp_schedule_type
   GFS_AUTO
 };

+struct gomp_doacross_work_share
+{
+  union {
+    /* chunk_size copy, as ws->chunk_size is multiplied by incr for
+       GFS_DYNAMIC.  */
+    long chunk_size;
+    /* Likewise, but for ull implementation.  */
+    unsigned long long chunk_size_ull;
+    /* For schedule(static,0) this is the number
+       of iterations assigned to the last thread, i.e. number of
+       iterations / number of threads.  */
+    long q;
+    /* Likewise, but for ull implementation.  */
+    unsigned long long q_ull;
+  };
+  /* Size of each array entry (padded to cache line size).  */
+  unsigned long elt_sz;
+  /* Number of dimensions in sink vectors.  */
+  unsigned int ncounts;
+  /* True if the iterations can be flattened.  */
+  bool flattened;
+  /* Actual array (of elt_sz sized units), aligned to cache line size.
+     This is indexed by team_id for GFS_STATIC and outermost iteration
+     / chunk_size for other schedules.  */
+  unsigned char *array;
+  /* These two are only used for schedule(static,0).  */
+  /* This one is number of iterations % number of threads.  */
+  long t;
+  union {
+    /* And this one is cached t * (q + 1).  */
+    long boundary;
+    /* Likewise, but for the ull implementation.  */
+    unsigned long long boundary_ull;
+  };
+  /* Array of shift counts for each dimension if they can be flattened.  */
+  unsigned int shift_counts[];
+};
+
 struct gomp_work_share
 {
   /* This member records the SCHEDULE clause to be used for this construct.
@@ -105,13 +204,18 @@ struct gomp_work_share
     };
   };

-  /* This is a circular queue that details which threads will be allowed
-     into the ordered region and in which order.  When a thread allocates
-     iterations on which it is going to work, it also registers itself at
-     the end of the array.  When a thread reaches the ordered region, it
-     checks to see if it is the one at the head of the queue.  If not, it
-     blocks on its RELEASE semaphore.  */
-  unsigned *ordered_team_ids;
+  union {
+    /* This is a circular queue that details which threads will be allowed
+       into the ordered region and in which order.  When a thread allocates
+       iterations on which it is going to work, it also registers itself at
+       the end of the array.  When a thread reaches the ordered region, it
+       checks to see if it is the one at the head of the queue.  If not, it
+       blocks on its RELEASE semaphore.  */
+    unsigned *ordered_team_ids;
+
+    /* This is a pointer to DOACROSS work share data.  */
+    struct gomp_doacross_work_share *doacross;
+  };

   /* This is the number of threads that have registered themselves in
      the circular queue ordered_team_ids.  */
@@ -230,7 +334,7 @@ struct gomp_task_icv
 {
   unsigned long nthreads_var;
   enum gomp_schedule_type run_sched_var;
-  int run_sched_modifier;
+  int run_sched_chunk_size;
   int default_device_var;
   unsigned int thread_limit_var;
   bool dyn_var;
@@ -246,6 +350,7 @@ extern gomp_mutex_t gomp_managed_threads
 #endif
 extern unsigned long gomp_max_active_levels_var;
 extern bool gomp_cancel_var;
+extern int gomp_max_task_priority_var;
 extern unsigned long long gomp_spin_count_var, gomp_throttled_spin_count_var;
 extern unsigned long gomp_available_cpus, gomp_managed_threads;
 extern unsigned long *gomp_nthreads_var_list, gomp_nthreads_var_list_len;
@@ -253,25 +358,36 @@ extern char *gomp_bind_var_list;
 extern unsigned long gomp_bind_var_list_len;
 extern void **gomp_places_list;
 extern unsigned long gomp_places_list_len;
+extern int gomp_debug_var;
+extern int goacc_device_num;
+extern char *goacc_device_type;

 enum gomp_task_kind
 {
+  /* Implicit task.  */
   GOMP_TASK_IMPLICIT,
-  GOMP_TASK_IFFALSE,
+  /* Undeferred task.  */
+  GOMP_TASK_UNDEFERRED,
+  /* Task created by GOMP_task and waiting to be run.  */
   GOMP_TASK_WAITING,
-  GOMP_TASK_TIED
+  /* Task currently executing or scheduled and about to execute.  */
+  GOMP_TASK_TIED,
+  /* Used for target tasks that have vars mapped and async run started,
+     but not yet completed.  Once that completes, they will be readded
+     into the queues as GOMP_TASK_WAITING in order to perform the var
+     unmapping.  */
+  GOMP_TASK_ASYNC_RUNNING
 };

-struct gomp_task;
-struct gomp_taskgroup;
-struct htab;
-
 struct gomp_task_depend_entry
 {
+  /* Address of dependency.  */
   void *addr;
   struct gomp_task_depend_entry *next;
   struct gomp_task_depend_entry *prev;
+  /* Task that provides the dependency in ADDR.  */
   struct gomp_task *task;
+  /* Depend entry is of type "IN".  */
   bool is_in;
   bool redundant;
   bool redundant_out;
@@ -290,8 +406,8 @@ struct gomp_taskwait
 {
   bool in_taskwait;
   bool in_depend_wait;
+  /* Number of tasks we are waiting for.  */
   size_t n_depend;
-  struct gomp_task *last_parent_depends_on;
   gomp_sem_t taskwait_sem;
 };

@@ -299,20 +415,31 @@ struct gomp_taskwait

 struct gomp_task
 {
+  /* Parent of this task.  */
   struct gomp_task *parent;
-  struct gomp_task *children;
-  struct gomp_task *next_child;
-  struct gomp_task *prev_child;
-  struct gomp_task *next_queue;
-  struct gomp_task *prev_queue;
-  struct gomp_task *next_taskgroup;
-  struct gomp_task *prev_taskgroup;
+  /* Children of this task.  */
+  struct priority_queue children_queue;
+  /* Taskgroup this task belongs in.  */
   struct gomp_taskgroup *taskgroup;
+  /* Tasks that depend on this task.  */
   struct gomp_dependers_vec *dependers;
   struct htab *depend_hash;
   struct gomp_taskwait *taskwait;
+  /* Number of items in DEPEND.  */
   size_t depend_count;
+  /* Number of tasks this task depends on.  Once this counter reaches
+     0, we have no unsatisfied dependencies, and this task can be put
+     into the various queues to be scheduled.  */
   size_t num_dependees;
+
+  /* Priority of this task.  */
+  int priority;
+  /* The priority node for this task in each of the different queues.
+     We put this here to avoid allocating space for each priority
+     node.  Then we play offsetof() games to convert between pnode[]
+     entries and the gomp_task in which they reside.  */
+  struct priority_node pnode[3];
+
   struct gomp_task_icv icv;
   void (*fn) (void *);
   void *fn_data;
@@ -320,20 +447,58 @@ struct gomp_task
   bool in_tied_task;
   bool final_task;
   bool copy_ctors_done;
+  /* Set for undeferred tasks with unsatisfied dependencies which
+     block further execution of their parent until the dependencies
+     are satisfied.  */
   bool parent_depends_on;
+  /* Dependencies provided and/or needed for this task.  DEPEND_COUNT
+     is the number of items available.  */
   struct gomp_task_depend_entry depend[];
 };

+/* This structure describes a single #pragma omp taskgroup.  */
+
 struct gomp_taskgroup
 {
   struct gomp_taskgroup *prev;
-  struct gomp_task *children;
+  /* Queue of tasks that belong in this taskgroup.  */
+  struct priority_queue taskgroup_queue;
   bool in_taskgroup_wait;
   bool cancelled;
   gomp_sem_t taskgroup_sem;
   size_t num_children;
 };

+/* Various state of OpenMP async offloading tasks.  */
+enum gomp_target_task_state
+{
+  GOMP_TARGET_TASK_DATA,
+  GOMP_TARGET_TASK_BEFORE_MAP,
+  GOMP_TARGET_TASK_FALLBACK,
+  GOMP_TARGET_TASK_READY_TO_RUN,
+  GOMP_TARGET_TASK_RUNNING,
+  GOMP_TARGET_TASK_FINISHED
+};
+
+/* This structure describes a target task.  */
+
+struct gomp_target_task
+{
+  struct gomp_device_descr *devicep;
+  void (*fn) (void *);
+  size_t mapnum;
+  size_t *sizes;
+  unsigned short *kinds;
+  unsigned int flags;
+  enum gomp_target_task_state state;
+  struct target_mem_desc *tgt;
+  struct gomp_task *task;
+  struct gomp_team *team;
+  /* Device-specific target arguments.  */
+  void **args;
+  void *hostaddrs[];
+};
+
 /* This structure describes a "team" of threads.  These are the threads
    that are spawned by a PARALLEL constructs, as well as the work sharing
    constructs that the team encounters.  */
@@ -396,7 +561,8 @@ struct gomp_team
   struct gomp_work_share work_shares[8];

   gomp_mutex_t task_lock;
-  struct gomp_task *task_queue;
+  /* Scheduled tasks.  */
+  struct priority_queue task_queue;
   /* Number of all GOMP_TASK_{WAITING,TIED} tasks in the team.  */
   unsigned int task_count;
   /* Number of GOMP_TASK_WAITING tasks currently waiting to be scheduled.  */
@@ -451,6 +617,9 @@ struct gomp_thread_pool
   struct gomp_thread **threads;
   unsigned threads_size;
   unsigned threads_used;
+  /* The last team is used for non-nested teams to delay their destruction to
+     make sure all the threads in the team move on to the pool's barrier before
+     the team's barrier is destroyed.  */
   struct gomp_team *last_team;
   /* Number of threads running in this contention group.  */
   unsigned long threads_busy;
@@ -519,23 +688,7 @@ extern bool gomp_affinity_same_place (vo
 extern bool gomp_affinity_finalize_place_list (bool);
 extern bool gomp_affinity_init_level (int, unsigned long, bool);
 extern void gomp_affinity_print_place (void *);
-
-/* alloc.c */
-
-extern void *gomp_malloc (size_t) __attribute__((malloc));
-extern void *gomp_malloc_cleared (size_t) __attribute__((malloc));
-extern void *gomp_realloc (void *, size_t);
-
-/* Avoid conflicting prototypes of alloca() in system headers by using
-   GCC's builtin alloca().  */
-#define gomp_alloca(x)  __builtin_alloca(x)
-
-/* error.c */
-
-extern void gomp_error (const char *, ...)
-	__attribute__((format (printf, 1, 2)));
-extern void gomp_fatal (const char *, ...)
-	__attribute__((noreturn, format (printf, 1, 2)));
+extern void gomp_get_place_proc_ids_8 (int, int64_t *);

 /* iter.c */

@@ -572,6 +725,9 @@ extern void gomp_ordered_next (void);
 extern void gomp_ordered_static_init (void);
 extern void gomp_ordered_static_next (void);
 extern void gomp_ordered_sync (void);
+extern void gomp_doacross_init (unsigned, long *, long);
+extern void gomp_doacross_ull_init (unsigned, unsigned long long *,
+				    unsigned long long);

 /* parallel.c */

@@ -588,6 +744,12 @@ extern void gomp_init_task (struct gomp_
			    struct gomp_task_icv *);
 extern void gomp_end_task (void);
 extern void gomp_barrier_handle_tasks (gomp_barrier_state_t);
+extern void gomp_task_maybe_wait_for_dependencies (void **);
+extern bool gomp_create_target_task (struct gomp_device_descr *,
+				     void (*) (void *), size_t, void **,
+				     size_t *, unsigned short *, unsigned int,
+				     void **, void **,
+				     enum gomp_target_task_state);

 static void inline
 gomp_finish_task (struct gomp_task *task)
@@ -606,7 +768,213 @@ extern void gomp_free_thread (void *);

 /* target.c */

+extern void gomp_init_targets_once (void);
 extern int gomp_get_num_devices (void);
+extern bool gomp_target_task_fn (void *);
+
+/* Splay tree definitions.  */
+typedef struct splay_tree_node_s *splay_tree_node;
+typedef struct splay_tree_s *splay_tree;
+typedef struct splay_tree_key_s *splay_tree_key;
+
+struct target_var_desc {
+  /* Splay key.  */
+  splay_tree_key key;
+  /* True if data should be copied from device to host at the end.  */
+  bool copy_from;
+  /* True if data always should be copied from device to host at the end.  */
+  bool always_copy_from;
+  /* Relative offset against key host_start.  */
+  uintptr_t offset;
+  /* Actual length.  */
+  uintptr_t length;
+};
+
+struct target_mem_desc {
+  /* Reference count.  */
+  uintptr_t refcount;
+  /* All the splay nodes allocated together.  */
+  splay_tree_node array;
+  /* Start of the target region.  */
+  uintptr_t tgt_start;
+  /* End of the targer region.  */
+  uintptr_t tgt_end;
+  /* Handle to free.  */
+  void *to_free;
+  /* Previous target_mem_desc.  */
+  struct target_mem_desc *prev;
+  /* Number of items in following list.  */
+  size_t list_count;
+
+  /* Corresponding target device descriptor.  */
+  struct gomp_device_descr *device_descr;
+
+  /* List of target items to remove (or decrease refcount)
+     at the end of region.  */
+  struct target_var_desc list[];
+};
+
+/* Special value for refcount - infinity.  */
+#define REFCOUNT_INFINITY (~(uintptr_t) 0)
+/* Special value for refcount - tgt_offset contains target address of the
+   artificial pointer to "omp declare target link" object.  */
+#define REFCOUNT_LINK (~(uintptr_t) 1)
+
+struct splay_tree_key_s {
+  /* Address of the host object.  */
+  uintptr_t host_start;
+  /* Address immediately after the host object.  */
+  uintptr_t host_end;
+  /* Descriptor of the target memory.  */
+  struct target_mem_desc *tgt;
+  /* Offset from tgt->tgt_start to the start of the target object.  */
+  uintptr_t tgt_offset;
+  /* Reference count.  */
+  uintptr_t refcount;
+  /* Pointer to the original mapping of "omp declare target link" object.  */
+  splay_tree_key link_key;
+};
+
+/* The comparison function.  */
+
+static inline int
+splay_compare (splay_tree_key x, splay_tree_key y)
+{
+  if (x->host_start == x->host_end
+      && y->host_start == y->host_end)
+    return 0;
+  if (x->host_end <= y->host_start)
+    return -1;
+  if (x->host_start >= y->host_end)
+    return 1;
+  return 0;
+}
+
+#include "splay-tree.h"
+
+typedef struct acc_dispatch_t
+{
+  /* This is a linked list of data mapped using the
+     acc_map_data/acc_unmap_data or "acc enter data"/"acc exit data" pragmas.
+     Unlike mapped_data in the goacc_thread struct, unmapping can
+     happen out-of-order with respect to mapping.  */
+  /* This is guarded by the lock in the "outer" struct gomp_device_descr.  */
+  struct target_mem_desc *data_environ;
+
+  /* Execute.  */
+  void (*exec_func) (void (*) (void *), size_t, void **, void **, int,
+		     unsigned *, void *);
+
+  /* Async cleanup callback registration.  */
+  void (*register_async_cleanup_func) (void *, int);
+
+  /* Asynchronous routines.  */
+  int (*async_test_func) (int);
+  int (*async_test_all_func) (void);
+  void (*async_wait_func) (int);
+  void (*async_wait_async_func) (int, int);
+  void (*async_wait_all_func) (void);
+  void (*async_wait_all_async_func) (int);
+  void (*async_set_async_func) (int);
+
+  /* Create/destroy TLS data.  */
+  void *(*create_thread_data_func) (int);
+  void (*destroy_thread_data_func) (void *);
+
+  /* NVIDIA target specific routines.  */
+  struct {
+    void *(*get_current_device_func) (void);
+    void *(*get_current_context_func) (void);
+    void *(*get_stream_func) (int);
+    int (*set_stream_func) (int, void *);
+  } cuda;
+} acc_dispatch_t;
+
+/* Various state of the accelerator device.  */
+enum gomp_device_state
+{
+  GOMP_DEVICE_UNINITIALIZED,
+  GOMP_DEVICE_INITIALIZED,
+  GOMP_DEVICE_FINALIZED
+};
+
+/* This structure describes accelerator device.
+   It contains name of the corresponding libgomp plugin, function handlers for
+   interaction with the device, ID-number of the device, and information about
+   mapped memory.  */
+struct gomp_device_descr
+{
+  /* Immutable data, which is only set during initialization, and which is not
+     guarded by the lock.  */
+
+  /* The name of the device.  */
+  const char *name;
+
+  /* Capabilities of device (supports OpenACC, OpenMP).  */
+  unsigned int capabilities;
+
+  /* This is the ID number of device among devices of the same type.  */
+  int target_id;
+
+  /* This is the TYPE of device.  */
+  enum offload_target_type type;
+
+  /* Function handlers.  */
+  const char *(*get_name_func) (void);
+  unsigned int (*get_caps_func) (void);
+  int (*get_type_func) (void);
+  int (*get_num_devices_func) (void);
+  bool (*init_device_func) (int);
+  bool (*fini_device_func) (int);
+  unsigned (*version_func) (void);
+  int (*load_image_func) (int, unsigned, const void *, struct addr_pair **);
+  bool (*unload_image_func) (int, unsigned, const void *);
+  void *(*alloc_func) (int, size_t);
+  bool (*free_func) (int, void *);
+  bool (*dev2host_func) (int, void *, const void *, size_t);
+  bool (*host2dev_func) (int, void *, const void *, size_t);
+  bool (*dev2dev_func) (int, void *, const void *, size_t);
+  bool (*can_run_func) (void *);
+  void (*run_func) (int, void *, void *, void **);
+  void (*async_run_func) (int, void *, void *, void **, void *);
+
+  /* Splay tree containing information about mapped memory regions.  */
+  struct splay_tree_s mem_map;
+
+  /* Mutex for the mutable data.  */
+  gomp_mutex_t lock;
+
+  /* Current state of the device.  OpenACC allows to move from INITIALIZED state
+     back to UNINITIALIZED state.  OpenMP allows only to move from INITIALIZED
+     to FINALIZED state (at program shutdown).  */
+  enum gomp_device_state state;
+
+  /* OpenACC-specific data and functions.  */
+  /* This is mutable because of its mutable data_environ and target_data
+     members.  */
+  acc_dispatch_t openacc;
+};
+
+/* Kind of the pragma, for which gomp_map_vars () is called.  */
+enum gomp_map_vars_kind
+{
+  GOMP_MAP_VARS_OPENACC,
+  GOMP_MAP_VARS_TARGET,
+  GOMP_MAP_VARS_DATA,
+  GOMP_MAP_VARS_ENTER_DATA
+};
+
+extern void gomp_acc_insert_pointer (size_t, void **, size_t *, void *);
+extern void gomp_acc_remove_pointer (void *, bool, int, int);
+
+extern struct target_mem_desc *gomp_map_vars (struct gomp_device_descr *,
+					      size_t, void **, void **,
+					      size_t *, void *, bool,
+					      enum gomp_map_vars_kind);
+extern void gomp_unmap_vars (struct target_mem_desc *, bool);
+extern void gomp_init_device (struct gomp_device_descr *);
+extern void gomp_free_memmap (struct splay_tree_s *);
+extern void gomp_unload_device (struct gomp_device_descr *);

 /* work.c */

@@ -646,8 +1014,28 @@ typedef enum omp_proc_bind_t
   omp_proc_bind_spread = 4
 } omp_proc_bind_t;

+typedef enum omp_lock_hint_t
+{
+  omp_lock_hint_none = 0,
+  omp_lock_hint_uncontended = 1,
+  omp_lock_hint_contended = 2,
+  omp_lock_hint_nonspeculative = 4,
+  omp_lock_hint_speculative = 8,
+} omp_lock_hint_t;
+
+extern void omp_init_lock_with_hint (omp_lock_t *, omp_lock_hint_t)
+  __GOMP_NOTHROW;
+extern void omp_init_nest_lock_with_hint (omp_lock_t *, omp_lock_hint_t)
+  __GOMP_NOTHROW;
+
 extern int omp_get_cancellation (void) __GOMP_NOTHROW;
 extern omp_proc_bind_t omp_get_proc_bind (void) __GOMP_NOTHROW;
+extern int omp_get_num_places (void) __GOMP_NOTHROW;
+extern int omp_get_place_num_procs (int) __GOMP_NOTHROW;
+extern void omp_get_place_proc_ids (int, int *) __GOMP_NOTHROW;
+extern int omp_get_place_num (void) __GOMP_NOTHROW;
+extern int omp_get_partition_num_places (void) __GOMP_NOTHROW;
+extern void omp_get_partition_place_nums (int *) __GOMP_NOTHROW;

 extern void omp_set_default_device (int) __GOMP_NOTHROW;
 extern int omp_get_default_device (void) __GOMP_NOTHROW;
@@ -656,6 +1044,24 @@ extern int omp_get_num_teams (void) __GO
 extern int omp_get_team_num (void) __GOMP_NOTHROW;

 extern int omp_is_initial_device (void) __GOMP_NOTHROW;
+extern int omp_get_initial_device (void) __GOMP_NOTHROW;
+extern int omp_get_max_task_priority (void) __GOMP_NOTHROW;
+
+extern void *omp_target_alloc (__SIZE_TYPE__, int) __GOMP_NOTHROW;
+extern void omp_target_free (void *, int) __GOMP_NOTHROW;
+extern int omp_target_is_present (void *, int) __GOMP_NOTHROW;
+extern int omp_target_memcpy (void *, void *, __SIZE_TYPE__, __SIZE_TYPE__,
+			      __SIZE_TYPE__, int, int) __GOMP_NOTHROW;
+extern int omp_target_memcpy_rect (void *, void *, __SIZE_TYPE__, int,
+				   const __SIZE_TYPE__ *,
+				   const __SIZE_TYPE__ *,
+				   const __SIZE_TYPE__ *,
+				   const __SIZE_TYPE__ *,
+				   const __SIZE_TYPE__ *, int, int)
+  __GOMP_NOTHROW;
+extern int omp_target_associate_ptr (void *, void *, __SIZE_TYPE__,
+				     __SIZE_TYPE__, int) __GOMP_NOTHROW;
+extern int omp_target_disassociate_ptr (void *, int) __GOMP_NOTHROW;

 #if !defined (HAVE_ATTRIBUTE_VISIBILITY) \
     || !defined (HAVE_ATTRIBUTE_ALIAS) \
@@ -728,4 +1134,34 @@ extern int gomp_test_nest_lock_25 (omp_n
 # define ialias_call(fn) fn
 #endif

+/* Helper function for priority_node_to_task() and
+   task_to_priority_node().
+
+   Return the offset from a task to its priority_node entry.  The
+   priority_node entry is has a type of TYPE.  */
+
+static inline size_t
+priority_queue_offset (enum priority_queue_type type)
+{
+  return offsetof (struct gomp_task, pnode[(int) type]);
+}
+
+/* Return the task associated with a priority NODE of type TYPE.  */
+
+static inline struct gomp_task *
+priority_node_to_task (enum priority_queue_type type,
+		       struct priority_node *node)
+{
+  return (struct gomp_task *) ((char *) node - priority_queue_offset (type));
+}
+
+/* Return the priority node of type TYPE for a given TASK.  */
+
+static inline struct priority_node *
+task_to_priority_node (enum priority_queue_type type,
+		       struct gomp_task *task)
+{
+  return (struct priority_node *) ((char *) task
+				   + priority_queue_offset (type));
+}
 #endif /* LIBGOMP_H */
--- libgomp/env.c.jj	2014-05-15 10:56:32.420522486 +0200
+++ libgomp/env.c	2016-07-13 16:57:04.437535335 +0200
@@ -27,6 +27,8 @@

 #include "libgomp.h"
 #include "libgomp_f.h"
+#include "oacc-int.h"
+#include "gomp-constants.h"
 #include <ctype.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -56,7 +58,7 @@ struct gomp_task_icv gomp_global_icv = {
   .nthreads_var = 1,
   .thread_limit_var = UINT_MAX,
   .run_sched_var = GFS_DYNAMIC,
-  .run_sched_modifier = 1,
+  .run_sched_chunk_size = 1,
   .default_device_var = 0,
   .dyn_var = false,
   .nest_var = false,
@@ -66,6 +68,7 @@ struct gomp_task_icv gomp_global_icv = {

 unsigned long gomp_max_active_levels_var = INT_MAX;
 bool gomp_cancel_var = false;
+int gomp_max_task_priority_var = 0;
 #ifndef HAVE_SYNC_BUILTINS
 gomp_mutex_t gomp_managed_threads_lock;
 #endif
@@ -76,6 +79,9 @@ char *gomp_bind_var_list;
 unsigned long gomp_bind_var_list_len;
 void **gomp_places_list;
 unsigned long gomp_places_list_len;
+int gomp_debug_var;
+char *goacc_device_type;
+int goacc_device_num;

 /* Parse the OMP_SCHEDULE environment variable.  */

@@ -118,7 +124,7 @@ parse_schedule (void)
     ++env;
   if (*env == '\0')
     {
-      gomp_global_icv.run_sched_modifier
+      gomp_global_icv.run_sched_chunk_size
	= gomp_global_icv.run_sched_var != GFS_STATIC;
       return;
     }
@@ -144,7 +150,7 @@ parse_schedule (void)

   if (value == 0 && gomp_global_icv.run_sched_var != GFS_STATIC)
     value = 1;
-  gomp_global_icv.run_sched_modifier = value;
+  gomp_global_icv.run_sched_chunk_size = value;
   return;

  unknown:
@@ -1011,6 +1017,16 @@ parse_affinity (bool ignore)
   return false;
 }

+static void
+parse_acc_device_type (void)
+{
+  const char *env = getenv ("ACC_DEVICE_TYPE");
+
+  if (env && *env != '\0')
+    goacc_device_type = strdup (env);
+  else
+    goacc_device_type = NULL;
+}

 static void
 handle_omp_display_env (unsigned long stacksize, int wait_policy)
@@ -1054,7 +1070,7 @@ handle_omp_display_env (unsigned long st

   fputs ("\nOPENMP DISPLAY ENVIRONMENT BEGIN\n", stderr);

-  fputs ("  _OPENMP = '201307'\n", stderr);
+  fputs ("  _OPENMP = '201511'\n", stderr);
   fprintf (stderr, "  OMP_DYNAMIC = '%s'\n",
	   gomp_global_icv.dyn_var ? "TRUE" : "FALSE");
   fprintf (stderr, "  OMP_NESTED = '%s'\n",
@@ -1142,6 +1158,8 @@ handle_omp_display_env (unsigned long st
	   gomp_cancel_var ? "TRUE" : "FALSE");
   fprintf (stderr, "  OMP_DEFAULT_DEVICE = '%d'\n",
	   gomp_global_icv.default_device_var);
+  fprintf (stderr, "  OMP_MAX_TASK_PRIORITY = '%d'\n",
+	   gomp_max_task_priority_var);

   if (verbose)
     {
@@ -1174,6 +1192,7 @@ initialize_env (void)
   parse_boolean ("OMP_NESTED", &gomp_global_icv.nest_var);
   parse_boolean ("OMP_CANCELLATION", &gomp_cancel_var);
   parse_int ("OMP_DEFAULT_DEVICE", &gomp_global_icv.default_device_var, true);
+  parse_int ("OMP_MAX_TASK_PRIORITY", &gomp_max_task_priority_var, true);
   parse_unsigned_long ("OMP_MAX_ACTIVE_LEVELS", &gomp_max_active_levels_var,
		       true);
   if (parse_unsigned_long ("OMP_THREAD_LIMIT", &thread_limit_var, false))
@@ -1181,6 +1200,7 @@ initialize_env (void)
       gomp_global_icv.thread_limit_var
	= thread_limit_var > INT_MAX ? UINT_MAX : thread_limit_var;
     }
+  parse_int ("GOMP_DEBUG", &gomp_debug_var, true);
 #ifndef HAVE_SYNC_BUILTINS
   gomp_mutex_init (&gomp_managed_threads_lock);
 #endif
@@ -1271,6 +1291,15 @@ initialize_env (void)
     }

   handle_omp_display_env (stacksize, wait_policy);
+
+  /* OpenACC.  */
+
+  if (!parse_int ("ACC_DEVICE_NUM", &goacc_device_num, true))
+    goacc_device_num = 0;
+
+  parse_acc_device_type ();
+
+  goacc_runtime_initialize ();
 }

 
@@ -1312,21 +1341,21 @@ omp_get_nested (void)
 }

 void
-omp_set_schedule (omp_sched_t kind, int modifier)
+omp_set_schedule (omp_sched_t kind, int chunk_size)
 {
   struct gomp_task_icv *icv = gomp_icv (true);
   switch (kind)
     {
     case omp_sched_static:
-      if (modifier < 1)
-	modifier = 0;
-      icv->run_sched_modifier = modifier;
+      if (chunk_size < 1)
+	chunk_size = 0;
+      icv->run_sched_chunk_size = chunk_size;
       break;
     case omp_sched_dynamic:
     case omp_sched_guided:
-      if (modifier < 1)
-	modifier = 1;
-      icv->run_sched_modifier = modifier;
+      if (chunk_size < 1)
+	chunk_size = 1;
+      icv->run_sched_chunk_size = chunk_size;
       break;
     case omp_sched_auto:
       break;
@@ -1337,11 +1366,11 @@ omp_set_schedule (omp_sched_t kind, int
 }

 void
-omp_get_schedule (omp_sched_t *kind, int *modifier)
+omp_get_schedule (omp_sched_t *kind, int *chunk_size)
 {
   struct gomp_task_icv *icv = gomp_icv (false);
   *kind = icv->run_sched_var;
-  *modifier = icv->run_sched_modifier;
+  *chunk_size = icv->run_sched_chunk_size;
 }

 int
@@ -1377,6 +1406,12 @@ omp_get_cancellation (void)
   return gomp_cancel_var;
 }

+int
+omp_get_max_task_priority (void)
+{
+  return gomp_max_task_priority_var;
+}
+
 omp_proc_bind_t
 omp_get_proc_bind (void)
 {
@@ -1425,6 +1460,59 @@ omp_is_initial_device (void)
   return 1;
 }

+int
+omp_get_initial_device (void)
+{
+  return GOMP_DEVICE_HOST_FALLBACK;
+}
+
+int
+omp_get_num_places (void)
+{
+  return gomp_places_list_len;
+}
+
+int
+omp_get_place_num (void)
+{
+  if (gomp_places_list == NULL)
+    return -1;
+
+  struct gomp_thread *thr = gomp_thread ();
+  if (thr->place == 0)
+    gomp_init_affinity ();
+
+  return (int) thr->place - 1;
+}
+
+int
+omp_get_partition_num_places (void)
+{
+  if (gomp_places_list == NULL)
+    return 0;
+
+  struct gomp_thread *thr = gomp_thread ();
+  if (thr->place == 0)
+    gomp_init_affinity ();
+
+  return thr->ts.place_partition_len;
+}
+
+void
+omp_get_partition_place_nums (int *place_nums)
+{
+  if (gomp_places_list == NULL)
+    return;
+
+  struct gomp_thread *thr = gomp_thread ();
+  if (thr->place == 0)
+    gomp_init_affinity ();
+
+  unsigned int i;
+  for (i = 0; i < thr->ts.place_partition_len; i++)
+    *place_nums++ = thr->ts.place_partition_off + i;
+}
+
 ialias (omp_set_dynamic)
 ialias (omp_set_nested)
 ialias (omp_set_num_threads)
@@ -1444,3 +1532,9 @@ ialias (omp_get_num_devices)
 ialias (omp_get_num_teams)
 ialias (omp_get_team_num)
 ialias (omp_is_initial_device)
+ialias (omp_get_initial_device)
+ialias (omp_get_max_task_priority)
+ialias (omp_get_num_places)
+ialias (omp_get_place_num)
+ialias (omp_get_partition_num_places)
+ialias (omp_get_partition_place_nums)
--- libgomp/openacc.h.jj	2016-07-13 16:57:04.432535397 +0200
+++ libgomp/openacc.h	2016-07-13 16:57:04.432535397 +0200
@@ -0,0 +1,131 @@
+/* OpenACC Runtime Library User-facing Declarations
+
+   Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _OPENACC_H
+#define _OPENACC_H 1
+
+/* The OpenACC standard is silent on whether or not including <openacc.h>
+   might or must not include other header files.  We chose to include
+   some.  */
+#include <stddef.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if __cplusplus >= 201103
+# define __GOACC_NOTHROW noexcept
+#elif __cplusplus
+# define __GOACC_NOTHROW throw ()
+#else /* Not C++ */
+# define __GOACC_NOTHROW __attribute__ ((__nothrow__))
+#endif
+
+/* Types */
+typedef enum acc_device_t {
+  /* Keep in sync with include/gomp-constants.h.  */
+  acc_device_none = 0,
+  acc_device_default = 1,
+  acc_device_host = 2,
+  /* acc_device_host_nonshm = 3 removed.  */
+  acc_device_not_host = 4,
+  acc_device_nvidia = 5,
+  _ACC_device_hwm,
+  /* Ensure enumeration is layout compatible with int.  */
+  _ACC_highest = __INT_MAX__,
+  _ACC_neg = -1
+} acc_device_t;
+
+typedef enum acc_async_t {
+  /* Keep in sync with include/gomp-constants.h.  */
+  acc_async_noval = -1,
+  acc_async_sync  = -2
+} acc_async_t;
+
+int acc_get_num_devices (acc_device_t) __GOACC_NOTHROW;
+void acc_set_device_type (acc_device_t) __GOACC_NOTHROW;
+acc_device_t acc_get_device_type (void) __GOACC_NOTHROW;
+void acc_set_device_num (int, acc_device_t) __GOACC_NOTHROW;
+int acc_get_device_num (acc_device_t) __GOACC_NOTHROW;
+int acc_async_test (int) __GOACC_NOTHROW;
+int acc_async_test_all (void) __GOACC_NOTHROW;
+void acc_wait (int) __GOACC_NOTHROW;
+void acc_wait_async (int, int) __GOACC_NOTHROW;
+void acc_wait_all (void) __GOACC_NOTHROW;
+void acc_wait_all_async (int) __GOACC_NOTHROW;
+void acc_init (acc_device_t) __GOACC_NOTHROW;
+void acc_shutdown (acc_device_t) __GOACC_NOTHROW;
+#ifdef __cplusplus
+int acc_on_device (int __arg) __GOACC_NOTHROW;
+#else
+int acc_on_device (acc_device_t __arg) __GOACC_NOTHROW;
+#endif
+void *acc_malloc (size_t) __GOACC_NOTHROW;
+void acc_free (void *) __GOACC_NOTHROW;
+/* Some of these would be more correct with const qualifiers, but
+   the standard specifies otherwise.  */
+void *acc_copyin (void *, size_t) __GOACC_NOTHROW;
+void *acc_present_or_copyin (void *, size_t) __GOACC_NOTHROW;
+void *acc_create (void *, size_t) __GOACC_NOTHROW;
+void *acc_present_or_create (void *, size_t) __GOACC_NOTHROW;
+void acc_copyout (void *, size_t) __GOACC_NOTHROW;
+void acc_delete (void *, size_t) __GOACC_NOTHROW;
+void acc_update_device (void *, size_t) __GOACC_NOTHROW;
+void acc_update_self (void *, size_t) __GOACC_NOTHROW;
+void acc_map_data (void *, void *, size_t) __GOACC_NOTHROW;
+void acc_unmap_data (void *) __GOACC_NOTHROW;
+void *acc_deviceptr (void *) __GOACC_NOTHROW;
+void *acc_hostptr (void *) __GOACC_NOTHROW;
+int acc_is_present (void *, size_t) __GOACC_NOTHROW;
+void acc_memcpy_to_device (void *, void *, size_t) __GOACC_NOTHROW;
+void acc_memcpy_from_device (void *, void *, size_t) __GOACC_NOTHROW;
+
+/* Old names.  OpenACC does not specify whether these can or must
+   not be macros, inlines or aliases for the new names.  */
+#define acc_pcreate acc_present_or_create
+#define acc_pcopyin acc_present_or_copyin
+
+/* CUDA-specific routines.  */
+void *acc_get_current_cuda_device (void) __GOACC_NOTHROW;
+void *acc_get_current_cuda_context (void) __GOACC_NOTHROW;
+void *acc_get_cuda_stream (int) __GOACC_NOTHROW;
+int acc_set_cuda_stream (int, void *) __GOACC_NOTHROW;
+
+#ifdef __cplusplus
+}
+
+/* Forwarding function with correctly typed arg.  */
+
+#pragma acc routine seq
+inline int acc_on_device (acc_device_t __arg) __GOACC_NOTHROW
+{
+  return acc_on_device ((int) __arg);
+}
+#endif
+
+#endif /* _OPENACC_H */
--- libgomp/config/linux/doacross.h.jj	2016-07-13 16:57:18.902355979 +0200
+++ libgomp/config/linux/doacross.h	2016-07-13 16:57:18.902355979 +0200
@@ -0,0 +1,57 @@
+/* Copyright (C) 2015-2016 Free Software Foundation, Inc.
+   Contributed by Jakub Jelinek <jakub@redhat.com>.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* This is a Linux specific implementation of doacross spinning.  */
+
+#ifndef GOMP_DOACROSS_H
+#define GOMP_DOACROSS_H 1
+
+#include "libgomp.h"
+#include <errno.h>
+#include "wait.h"
+
+#ifdef HAVE_ATTRIBUTE_VISIBILITY
+# pragma GCC visibility push(hidden)
+#endif
+
+static inline void doacross_spin (unsigned long *addr, unsigned long expected,
+				  unsigned long cur)
+{
+  /* FIXME: back off depending on how large expected - cur is.  */
+  do
+    {
+      cpu_relax ();
+      cur = __atomic_load_n (addr, MEMMODEL_RELAXED);
+      if (expected < cur)
+	return;
+    }
+  while (1);
+}
+
+#ifdef HAVE_ATTRIBUTE_VISIBILITY
+# pragma GCC visibility pop
+#endif
+
+#endif /* GOMP_DOACROSS_H */
--- libgomp/config/posix/doacross.h.jj	2016-07-13 16:57:18.903355966 +0200
+++ libgomp/config/posix/doacross.h	2016-07-13 16:57:18.903355966 +0200
@@ -0,0 +1,62 @@
+/* Copyright (C) 2015-2016 Free Software Foundation, Inc.
+   Contributed by Jakub Jelinek <jakub@redhat.com>.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* This is a generic implementation of doacross spinning.  */
+
+#ifndef GOMP_DOACROSS_H
+#define GOMP_DOACROSS_H 1
+
+#include "libgomp.h"
+#include <errno.h>
+
+#ifdef HAVE_ATTRIBUTE_VISIBILITY
+# pragma GCC visibility push(hidden)
+#endif
+
+static inline void
+cpu_relax (void)
+{
+  __asm volatile ("" : : : "memory");
+}
+
+static inline void doacross_spin (unsigned long *addr, unsigned long expected,
+				  unsigned long cur)
+{
+  /* FIXME: back off depending on how large expected - cur is.  */
+  do
+    {
+      cpu_relax ();
+      cur = __atomic_load_n (addr, MEMMODEL_RELAXED);
+      if (expected < cur)
+	return;
+    }
+  while (1);
+}
+
+#ifdef HAVE_ATTRIBUTE_VISIBILITY
+# pragma GCC visibility pop
+#endif
+
+#endif /* GOMP_DOACROSS_H */
--- libgomp/splay-tree.c.jj	2016-07-13 16:57:18.919355768 +0200
+++ libgomp/splay-tree.c	2016-07-13 16:57:18.919355768 +0200
@@ -0,0 +1,238 @@
+/* A splay-tree datatype.
+   Copyright (C) 1998-2016 Free Software Foundation, Inc.
+   Contributed by Mark Mitchell (mark@markmitchell.com).
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* The splay tree code copied from include/splay-tree.h and adjusted,
+   so that all the data lives directly in splay_tree_node_s structure
+   and no extra allocations are needed.  */
+
+/* For an easily readable description of splay-trees, see:
+
+     Lewis, Harry R. and Denenberg, Larry.  Data Structures and Their
+     Algorithms.  Harper-Collins, Inc.  1991.
+
+   The major feature of splay trees is that all basic tree operations
+   are amortized O(log n) time for a tree with n nodes.  */
+
+#include "libgomp.h"
+
+/* Rotate the edge joining the left child N with its parent P.  PP is the
+   grandparents' pointer to P.  */
+
+static inline void
+rotate_left (splay_tree_node *pp, splay_tree_node p, splay_tree_node n)
+{
+  splay_tree_node tmp;
+  tmp = n->right;
+  n->right = p;
+  p->left = tmp;
+  *pp = n;
+}
+
+/* Rotate the edge joining the right child N with its parent P.  PP is the
+   grandparents' pointer to P.  */
+
+static inline void
+rotate_right (splay_tree_node *pp, splay_tree_node p, splay_tree_node n)
+{
+  splay_tree_node tmp;
+  tmp = n->left;
+  n->left = p;
+  p->right = tmp;
+  *pp = n;
+}
+
+/* Bottom up splay of KEY.  */
+
+static void
+splay_tree_splay (splay_tree sp, splay_tree_key key)
+{
+  if (sp->root == NULL)
+    return;
+
+  do {
+    int cmp1, cmp2;
+    splay_tree_node n, c;
+
+    n = sp->root;
+    cmp1 = splay_compare (key, &n->key);
+
+    /* Found.  */
+    if (cmp1 == 0)
+      return;
+
+    /* Left or right?  If no child, then we're done.  */
+    if (cmp1 < 0)
+      c = n->left;
+    else
+      c = n->right;
+    if (!c)
+      return;
+
+    /* Next one left or right?  If found or no child, we're done
+       after one rotation.  */
+    cmp2 = splay_compare (key, &c->key);
+    if (cmp2 == 0
+	|| (cmp2 < 0 && !c->left)
+	|| (cmp2 > 0 && !c->right))
+      {
+	if (cmp1 < 0)
+	  rotate_left (&sp->root, n, c);
+	else
+	  rotate_right (&sp->root, n, c);
+	return;
+      }
+
+    /* Now we have the four cases of double-rotation.  */
+    if (cmp1 < 0 && cmp2 < 0)
+      {
+	rotate_left (&n->left, c, c->left);
+	rotate_left (&sp->root, n, n->left);
+      }
+    else if (cmp1 > 0 && cmp2 > 0)
+      {
+	rotate_right (&n->right, c, c->right);
+	rotate_right (&sp->root, n, n->right);
+      }
+    else if (cmp1 < 0 && cmp2 > 0)
+      {
+	rotate_right (&n->left, c, c->right);
+	rotate_left (&sp->root, n, n->left);
+      }
+    else if (cmp1 > 0 && cmp2 < 0)
+      {
+	rotate_left (&n->right, c, c->left);
+	rotate_right (&sp->root, n, n->right);
+      }
+  } while (1);
+}
+
+/* Insert a new NODE into SP.  The NODE shouldn't exist in the tree.  */
+
+attribute_hidden void
+splay_tree_insert (splay_tree sp, splay_tree_node node)
+{
+  int comparison = 0;
+
+  splay_tree_splay (sp, &node->key);
+
+  if (sp->root)
+    comparison = splay_compare (&sp->root->key, &node->key);
+
+  if (sp->root && comparison == 0)
+    gomp_fatal ("Duplicate node");
+  else
+    {
+      /* Insert it at the root.  */
+      if (sp->root == NULL)
+	node->left = node->right = NULL;
+      else if (comparison < 0)
+	{
+	  node->left = sp->root;
+	  node->right = node->left->right;
+	  node->left->right = NULL;
+	}
+      else
+	{
+	  node->right = sp->root;
+	  node->left = node->right->left;
+	  node->right->left = NULL;
+	}
+
+      sp->root = node;
+    }
+}
+
+/* Remove node with KEY from SP.  It is not an error if it did not exist.  */
+
+attribute_hidden void
+splay_tree_remove (splay_tree sp, splay_tree_key key)
+{
+  splay_tree_splay (sp, key);
+
+  if (sp->root && splay_compare (&sp->root->key, key) == 0)
+    {
+      splay_tree_node left, right;
+
+      left = sp->root->left;
+      right = sp->root->right;
+
+      /* One of the children is now the root.  Doesn't matter much
+	 which, so long as we preserve the properties of the tree.  */
+      if (left)
+	{
+	  sp->root = left;
+
+	  /* If there was a right child as well, hang it off the
+	     right-most leaf of the left child.  */
+	  if (right)
+	    {
+	      while (left->right)
+		left = left->right;
+	      left->right = right;
+	    }
+	}
+      else
+	sp->root = right;
+    }
+}
+
+/* Lookup KEY in SP, returning NODE if present, and NULL
+   otherwise.  */
+
+attribute_hidden splay_tree_key
+splay_tree_lookup (splay_tree sp, splay_tree_key key)
+{
+  splay_tree_splay (sp, key);
+
+  if (sp->root && splay_compare (&sp->root->key, key) == 0)
+    return &sp->root->key;
+  else
+    return NULL;
+}
+
+/* Helper function for splay_tree_foreach.
+
+   Run FUNC on every node in KEY.  */
+
+static void
+splay_tree_foreach_internal (splay_tree_node node, splay_tree_callback func,
+			     void *data)
+{
+  if (!node)
+    return;
+  func (&node->key, data);
+  splay_tree_foreach_internal (node->left, func, data);
+  /* Yeah, whatever.  GCC can fix my tail recursion.  */
+  splay_tree_foreach_internal (node->right, func, data);
+}
+
+/* Run FUNC on each of the nodes in SP.  */
+
+attribute_hidden void
+splay_tree_foreach (splay_tree sp, splay_tree_callback func, void *data)
+{
+  splay_tree_foreach_internal (sp->root, func, data);
+}
--- libgomp/libgomp-plugin.c.jj	2016-07-13 16:57:04.435535360 +0200
+++ libgomp/libgomp-plugin.c	2016-07-13 16:57:04.435535360 +0200
@@ -0,0 +1,80 @@
+/* Copyright (C) 2014-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* Exported (non-hidden) functions exposing libgomp interface for plugins.  */
+
+#include <stdlib.h>
+
+#include "libgomp.h"
+#include "libgomp-plugin.h"
+
+void *
+GOMP_PLUGIN_malloc (size_t size)
+{
+  return gomp_malloc (size);
+}
+
+void *
+GOMP_PLUGIN_malloc_cleared (size_t size)
+{
+  return gomp_malloc_cleared (size);
+}
+
+void *
+GOMP_PLUGIN_realloc (void *ptr, size_t size)
+{
+  return gomp_realloc (ptr, size);
+}
+
+void
+GOMP_PLUGIN_debug (int kind, const char *msg, ...)
+{
+  va_list ap;
+
+  va_start (ap, msg);
+  gomp_vdebug (kind, msg, ap);
+  va_end (ap);
+}
+
+void
+GOMP_PLUGIN_error (const char *msg, ...)
+{
+  va_list ap;
+
+  va_start (ap, msg);
+  gomp_verror (msg, ap);
+  va_end (ap);
+}
+
+void
+GOMP_PLUGIN_fatal (const char *msg, ...)
+{
+  va_list ap;
+
+  va_start (ap, msg);
+  gomp_vfatal (msg, ap);
+  va_end (ap);
+}
--- libgomp/libgomp-plugin.h.jj	2016-07-13 16:57:04.438535323 +0200
+++ libgomp/libgomp-plugin.h	2016-07-13 16:57:04.438535323 +0200
@@ -0,0 +1,80 @@
+/* Copyright (C) 2014-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* An interface to various libgomp-internal functions for use by plugins.  */
+
+#ifndef LIBGOMP_PLUGIN_H
+#define LIBGOMP_PLUGIN_H 1
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Capabilities of offloading devices.  */
+#define GOMP_OFFLOAD_CAP_SHARED_MEM	(1 << 0)
+#define GOMP_OFFLOAD_CAP_NATIVE_EXEC	(1 << 1)
+#define GOMP_OFFLOAD_CAP_OPENMP_400	(1 << 2)
+#define GOMP_OFFLOAD_CAP_OPENACC_200	(1 << 3)
+
+/* Type of offload target device.  Keep in sync with include/gomp-constants.h.  */
+enum offload_target_type
+{
+  OFFLOAD_TARGET_TYPE_HOST = 2,
+  /* OFFLOAD_TARGET_TYPE_HOST_NONSHM = 3 removed.  */
+  OFFLOAD_TARGET_TYPE_NVIDIA_PTX = 5,
+  OFFLOAD_TARGET_TYPE_INTEL_MIC = 6,
+  OFFLOAD_TARGET_TYPE_HSA = 7
+};
+
+/* Auxiliary struct, used for transferring pairs of addresses from plugin
+   to libgomp.  */
+struct addr_pair
+{
+  uintptr_t start;
+  uintptr_t end;
+};
+
+/* Miscellaneous functions.  */
+extern void *GOMP_PLUGIN_malloc (size_t) __attribute__ ((malloc));
+extern void *GOMP_PLUGIN_malloc_cleared (size_t) __attribute__ ((malloc));
+extern void *GOMP_PLUGIN_realloc (void *, size_t);
+void GOMP_PLUGIN_target_task_completion (void *);
+
+extern void GOMP_PLUGIN_debug (int, const char *, ...)
+	__attribute__ ((format (printf, 2, 3)));
+extern void GOMP_PLUGIN_error (const char *, ...)
+	__attribute__ ((format (printf, 1, 2)));
+extern void GOMP_PLUGIN_fatal (const char *, ...)
+	__attribute__ ((noreturn, format (printf, 1, 2)));
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
--- libgomp/oacc-async.c.jj	2016-07-13 16:57:13.488423109 +0200
+++ libgomp/oacc-async.c	2016-07-13 16:57:13.488423109 +0200
@@ -0,0 +1,107 @@
+/* OpenACC Runtime Library Definitions.
+
+   Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <assert.h>
+#include "openacc.h"
+#include "libgomp.h"
+#include "oacc-int.h"
+
+int
+acc_async_test (int async)
+{
+  if (async < acc_async_sync)
+    gomp_fatal ("invalid async argument: %d", async);
+
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (!thr || !thr->dev)
+    gomp_fatal ("no device active");
+
+  return thr->dev->openacc.async_test_func (async);
+}
+
+int
+acc_async_test_all (void)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (!thr || !thr->dev)
+    gomp_fatal ("no device active");
+
+  return thr->dev->openacc.async_test_all_func ();
+}
+
+void
+acc_wait (int async)
+{
+  if (async < acc_async_sync)
+    gomp_fatal ("invalid async argument: %d", async);
+
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (!thr || !thr->dev)
+    gomp_fatal ("no device active");
+
+  thr->dev->openacc.async_wait_func (async);
+}
+
+void
+acc_wait_async (int async1, int async2)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (!thr || !thr->dev)
+    gomp_fatal ("no device active");
+
+  thr->dev->openacc.async_wait_async_func (async1, async2);
+}
+
+void
+acc_wait_all (void)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (!thr || !thr->dev)
+    gomp_fatal ("no device active");
+
+  thr->dev->openacc.async_wait_all_func ();
+}
+
+void
+acc_wait_all_async (int async)
+{
+  if (async < acc_async_sync)
+    gomp_fatal ("invalid async argument: %d", async);
+
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (!thr || !thr->dev)
+    gomp_fatal ("no device active");
+
+  thr->dev->openacc.async_wait_all_async_func (async);
+}
--- libgomp/splay-tree.h.jj	2016-07-13 16:57:18.934355582 +0200
+++ libgomp/splay-tree.h	2016-07-13 16:57:18.934355582 +0200
@@ -0,0 +1,130 @@
+/* A splay-tree datatype.
+   Copyright (C) 1998-2016 Free Software Foundation, Inc.
+   Contributed by Mark Mitchell (mark@markmitchell.com).
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* The splay tree code copied from include/splay-tree.h and adjusted,
+   so that all the data lives directly in splay_tree_node_s structure
+   and no extra allocations are needed.
+
+   Files including this header should before including it add:
+typedef struct splay_tree_node_s *splay_tree_node;
+typedef struct splay_tree_s *splay_tree;
+typedef struct splay_tree_key_s *splay_tree_key;
+   define splay_tree_key_s structure, and define
+   splay_compare inline function.
+
+   Alternatively, they can define splay_tree_prefix macro before
+   including this header and then all the above types, the
+   splay_compare function and the splay_tree_{lookup,insert_remove}
+   function will be prefixed by that prefix.  If splay_tree_prefix
+   macro is defined, this header must be included twice: once where
+   you need the header file definitions, and once where you need the
+   .c implementation routines.  In the latter case, you must also
+   define the macro splay_tree_c.  See the include of splay-tree.h in
+   priority_queue.[hc] for an example.  */
+
+/* For an easily readable description of splay-trees, see:
+
+     Lewis, Harry R. and Denenberg, Larry.  Data Structures and Their
+     Algorithms.  Harper-Collins, Inc.  1991.
+
+   The major feature of splay trees is that all basic tree operations
+   are amortized O(log n) time for a tree with n nodes.  */
+
+#ifdef splay_tree_prefix
+# define splay_tree_name_1(prefix, name) prefix ## _ ## name
+# define splay_tree_name(prefix, name) splay_tree_name_1 (prefix, name)
+# define splay_tree_node_s	\
+    splay_tree_name (splay_tree_prefix, splay_tree_node_s)
+# define splay_tree_s		\
+    splay_tree_name (splay_tree_prefix, splay_tree_s)
+# define splay_tree_key_s	\
+    splay_tree_name (splay_tree_prefix, splay_tree_key_s)
+# define splay_tree_node	\
+    splay_tree_name (splay_tree_prefix, splay_tree_node)
+# define splay_tree		\
+    splay_tree_name (splay_tree_prefix, splay_tree)
+# define splay_tree_key		\
+    splay_tree_name (splay_tree_prefix, splay_tree_key)
+# define splay_compare		\
+    splay_tree_name (splay_tree_prefix, splay_compare)
+# define splay_tree_lookup	\
+    splay_tree_name (splay_tree_prefix, splay_tree_lookup)
+# define splay_tree_insert	\
+    splay_tree_name (splay_tree_prefix, splay_tree_insert)
+# define splay_tree_remove	\
+    splay_tree_name (splay_tree_prefix, splay_tree_remove)
+# define splay_tree_foreach	\
+    splay_tree_name (splay_tree_prefix, splay_tree_foreach)
+# define splay_tree_callback	\
+    splay_tree_name (splay_tree_prefix, splay_tree_callback)
+#endif
+
+#ifndef splay_tree_c
+/* Header file definitions and prototypes.  */
+
+/* The nodes in the splay tree.  */
+struct splay_tree_node_s {
+  struct splay_tree_key_s key;
+  /* The left and right children, respectively.  */
+  splay_tree_node left;
+  splay_tree_node right;
+};
+
+/* The splay tree.  */
+struct splay_tree_s {
+  splay_tree_node root;
+};
+
+typedef void (*splay_tree_callback) (splay_tree_key, void *);
+
+extern splay_tree_key splay_tree_lookup (splay_tree, splay_tree_key);
+extern void splay_tree_insert (splay_tree, splay_tree_node);
+extern void splay_tree_remove (splay_tree, splay_tree_key);
+extern void splay_tree_foreach (splay_tree, splay_tree_callback, void *);
+#else  /* splay_tree_c */
+#  ifdef splay_tree_prefix
+#    include "splay-tree.c"
+#    undef splay_tree_name_1
+#    undef splay_tree_name
+#    undef splay_tree_node_s
+#    undef splay_tree_s
+#    undef splay_tree_key_s
+#    undef splay_tree_node
+#    undef splay_tree
+#    undef splay_tree_key
+#    undef splay_compare
+#    undef splay_tree_lookup
+#    undef splay_tree_insert
+#    undef splay_tree_remove
+#    undef splay_tree_foreach
+#    undef splay_tree_callback
+#    undef splay_tree_c
+#  endif
+#endif /* #ifndef splay_tree_c */
+
+#ifdef splay_tree_prefix
+#  undef splay_tree_prefix
+#endif
--- libgomp/oacc-plugin.c.jj	2016-07-13 16:57:13.481423196 +0200
+++ libgomp/oacc-plugin.c	2016-07-14 15:40:21.653151873 +0200
@@ -0,0 +1,44 @@
+/* Copyright (C) 2014-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* Initialize and register OpenACC dispatch table from libgomp plugin.  */
+
+#include "libgomp.h"
+#include "oacc-plugin.h"
+#include "oacc-int.h"
+
+void
+GOMP_PLUGIN_async_unmap_vars (void *ptr, int async)
+{
+}
+
+/* Return the target-specific part of the TLS data for the current thread.  */
+
+void *
+GOMP_PLUGIN_acc_thread (void)
+{
+  return NULL;
+}
--- libgomp/oacc-init.c.jj	2016-07-13 16:57:04.423535509 +0200
+++ libgomp/oacc-init.c	2016-07-14 19:06:41.679575688 +0200
@@ -0,0 +1,640 @@
+/* OpenACC Runtime initialization routines
+
+   Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include "libgomp.h"
+#include "oacc-int.h"
+#include "openacc.h"
+#include <assert.h>
+#include <stdlib.h>
+#include <strings.h>
+#include <stdbool.h>
+#include <string.h>
+
+/* This lock is used to protect access to cached_base_dev, dispatchers and
+   the (abstract) initialisation state of attached offloading devices.  */
+
+static gomp_mutex_t acc_device_lock;
+
+/* A cached version of the dispatcher for the global "current" accelerator type,
+   e.g. used as the default when creating new host threads.  This is the
+   device-type equivalent of goacc_device_num (which specifies which device to
+   use out of potentially several of the same type).  If there are several
+   devices of a given type, this points at the first one.  */
+
+static struct gomp_device_descr *cached_base_dev = NULL;
+
+#if defined HAVE_TLS || defined USE_EMUTLS
+__thread struct goacc_thread *goacc_tls_data;
+#else
+pthread_key_t goacc_tls_key;
+#endif
+static pthread_key_t goacc_cleanup_key;
+
+static struct goacc_thread *goacc_threads;
+static gomp_mutex_t goacc_thread_lock;
+
+/* An array of dispatchers for device types, indexed by the type.  This array
+   only references "base" devices, and other instances of the same type are
+   found by simply indexing from each such device (which are stored linearly,
+   grouped by device in target.c:devices).  */
+static struct gomp_device_descr *dispatchers[_ACC_device_hwm] = { 0 };
+
+attribute_hidden void
+goacc_register (struct gomp_device_descr *disp)
+{
+  /* Only register the 0th device here.  */
+  if (disp->target_id != 0)
+    return;
+
+  gomp_mutex_lock (&acc_device_lock);
+
+  assert (acc_device_type (disp->type) != acc_device_none
+	  && acc_device_type (disp->type) != acc_device_default
+	  && acc_device_type (disp->type) != acc_device_not_host);
+  assert (!dispatchers[disp->type]);
+  dispatchers[disp->type] = disp;
+
+  gomp_mutex_unlock (&acc_device_lock);
+}
+
+static const char *
+name_of_acc_device_t (enum acc_device_t type)
+{
+  switch (type)
+    {
+    case acc_device_none: return "none";
+    case acc_device_default: return "default";
+    case acc_device_host: return "host";
+    case acc_device_not_host: return "not_host";
+    case acc_device_nvidia: return "nvidia";
+    default: gomp_fatal ("unknown device type %u", (unsigned) type);
+    }
+}
+
+/* ACC_DEVICE_LOCK must be held before calling this function.  If FAIL_IS_ERROR
+   is true, this function raises an error if there are no devices of type D,
+   otherwise it returns NULL in that case.  */
+
+static struct gomp_device_descr *
+resolve_device (acc_device_t d, bool fail_is_error)
+{
+  acc_device_t d_arg = d;
+
+  switch (d)
+    {
+    case acc_device_default:
+      {
+	if (goacc_device_type)
+	  {
+	    /* Lookup the named device.  */
+	    if (!strcasecmp (goacc_device_type, "host"))
+	      {
+		d = acc_device_host;
+		goto found;
+	      }
+
+	    if (fail_is_error)
+	      {
+		gomp_mutex_unlock (&acc_device_lock);
+		gomp_fatal ("device type %s not supported", goacc_device_type);
+	      }
+	    else
+	      return NULL;
+	  }
+
+	/* No default device specified, so start scanning for any non-host
+	   device that is available.  */
+	d = acc_device_not_host;
+      }
+      /* FALLTHROUGH */
+
+    case acc_device_not_host:
+      if (d_arg == acc_device_default)
+	{
+	  d = acc_device_host;
+	  goto found;
+	}
+      if (fail_is_error)
+        {
+	  gomp_mutex_unlock (&acc_device_lock);
+	  gomp_fatal ("no device found");
+	}
+      else
+        return NULL;
+      break;
+
+    case acc_device_host:
+      break;
+
+    default:
+      if (d > _ACC_device_hwm)
+	{
+	  if (fail_is_error)
+	    goto unsupported_device;
+	  else
+	    return NULL;
+	}
+      break;
+    }
+ found:
+
+  assert (d != acc_device_none
+	  && d != acc_device_default
+	  && d != acc_device_not_host);
+
+  if (dispatchers[d] == NULL && fail_is_error)
+    {
+    unsupported_device:
+      gomp_mutex_unlock (&acc_device_lock);
+      gomp_fatal ("device type %s not supported", name_of_acc_device_t (d));
+    }
+
+  return dispatchers[d];
+}
+
+/* Emit a suitable error if no device of a particular type is available, or
+   the given device number is out-of-range.  */
+static void
+acc_dev_num_out_of_range (acc_device_t d, int ord, int ndevs)
+{
+  if (ndevs == 0)
+    gomp_fatal ("no devices of type %s available", name_of_acc_device_t (d));
+  else
+    gomp_fatal ("device %u out of range", ord);
+}
+
+/* This is called when plugins have been initialized, and serves to call
+   (indirectly) the target's device_init hook.  Calling multiple times without
+   an intervening acc_shutdown_1 call is an error.  ACC_DEVICE_LOCK must be
+   held before calling this function.  */
+
+static struct gomp_device_descr *
+acc_init_1 (acc_device_t d)
+{
+  struct gomp_device_descr *base_dev, *acc_dev;
+  int ndevs;
+
+  base_dev = resolve_device (d, true);
+
+  ndevs = base_dev->get_num_devices_func ();
+
+  if (ndevs <= 0 || goacc_device_num >= ndevs)
+    acc_dev_num_out_of_range (d, goacc_device_num, ndevs);
+
+  acc_dev = &base_dev[goacc_device_num];
+
+  gomp_mutex_lock (&acc_dev->lock);
+  if (acc_dev->state == GOMP_DEVICE_INITIALIZED)
+    {
+      gomp_mutex_unlock (&acc_dev->lock);
+      gomp_fatal ("device already active");
+    }
+
+  gomp_init_device (acc_dev);
+  gomp_mutex_unlock (&acc_dev->lock);
+
+  return base_dev;
+}
+
+/* ACC_DEVICE_LOCK must be held before calling this function.  */
+
+static void
+acc_shutdown_1 (acc_device_t d)
+{
+  struct gomp_device_descr *base_dev;
+  struct goacc_thread *walk;
+  int ndevs, i;
+  bool devices_active = false;
+
+  /* Get the base device for this device type.  */
+  base_dev = resolve_device (d, true);
+
+  ndevs = base_dev->get_num_devices_func ();
+
+  gomp_mutex_lock (&goacc_thread_lock);
+
+  /* Free target-specific TLS data and close all devices.  */
+  for (walk = goacc_threads; walk != NULL; walk = walk->next)
+    {
+      if (walk->target_tls)
+	base_dev->openacc.destroy_thread_data_func (walk->target_tls);
+
+      walk->target_tls = NULL;
+
+      /* Similarly, if this happens then user code has done something weird.  */
+      if (walk->saved_bound_dev)
+	{
+	  gomp_mutex_unlock (&goacc_thread_lock);
+	  gomp_fatal ("shutdown during host fallback");
+	}
+
+      if (walk->dev)
+	{
+	  gomp_mutex_lock (&walk->dev->lock);
+	  gomp_free_memmap (&walk->dev->mem_map);
+	  gomp_mutex_unlock (&walk->dev->lock);
+
+	  walk->dev = NULL;
+	  walk->base_dev = NULL;
+	}
+    }
+
+  gomp_mutex_unlock (&goacc_thread_lock);
+
+  /* Close all the devices of this type that have been opened.  */
+  bool ret = true;
+  for (i = 0; i < ndevs; i++)
+    {
+      struct gomp_device_descr *acc_dev = &base_dev[i];
+      gomp_mutex_lock (&acc_dev->lock);
+      if (acc_dev->state == GOMP_DEVICE_INITIALIZED)
+        {
+	  devices_active = true;
+	  ret &= acc_dev->fini_device_func (acc_dev->target_id);
+	  acc_dev->state = GOMP_DEVICE_UNINITIALIZED;
+	}
+      gomp_mutex_unlock (&acc_dev->lock);
+    }
+
+  if (!ret)
+    gomp_fatal ("device finalization failed");
+
+  if (!devices_active)
+    gomp_fatal ("no device initialized");
+}
+
+static struct goacc_thread *
+goacc_new_thread (void)
+{
+  struct goacc_thread *thr = gomp_malloc (sizeof (struct gomp_thread));
+
+#if defined HAVE_TLS || defined USE_EMUTLS
+  goacc_tls_data = thr;
+#else
+  pthread_setspecific (goacc_tls_key, thr);
+#endif
+
+  pthread_setspecific (goacc_cleanup_key, thr);
+
+  gomp_mutex_lock (&goacc_thread_lock);
+  thr->next = goacc_threads;
+  goacc_threads = thr;
+  gomp_mutex_unlock (&goacc_thread_lock);
+
+  return thr;
+}
+
+static void
+goacc_destroy_thread (void *data)
+{
+  struct goacc_thread *thr = data, *walk, *prev;
+
+  gomp_mutex_lock (&goacc_thread_lock);
+
+  if (thr)
+    {
+      struct gomp_device_descr *acc_dev = thr->dev;
+
+      if (acc_dev && thr->target_tls)
+	{
+	  acc_dev->openacc.destroy_thread_data_func (thr->target_tls);
+	  thr->target_tls = NULL;
+	}
+
+      assert (!thr->mapped_data);
+
+      /* Remove from thread list.  */
+      for (prev = NULL, walk = goacc_threads; walk;
+	   prev = walk, walk = walk->next)
+	if (walk == thr)
+	  {
+	    if (prev == NULL)
+	      goacc_threads = walk->next;
+	    else
+	      prev->next = walk->next;
+
+	    free (thr);
+
+	    break;
+	  }
+
+      assert (walk);
+    }
+
+  gomp_mutex_unlock (&goacc_thread_lock);
+}
+
+/* Use the ORD'th device instance for the current host thread (or -1 for the
+   current global default).  The device (and the runtime) must be initialised
+   before calling this function.  */
+
+void
+goacc_attach_host_thread_to_device (int ord)
+{
+  struct goacc_thread *thr = goacc_thread ();
+  struct gomp_device_descr *acc_dev = NULL, *base_dev = NULL;
+  int num_devices;
+
+  if (thr && thr->dev && (thr->dev->target_id == ord || ord < 0))
+    return;
+
+  if (ord < 0)
+    ord = goacc_device_num;
+
+  /* Decide which type of device to use.  If the current thread has a device
+     type already (e.g. set by acc_set_device_type), use that, else use the
+     global default.  */
+  if (thr && thr->base_dev)
+    base_dev = thr->base_dev;
+  else
+    {
+      assert (cached_base_dev);
+      base_dev = cached_base_dev;
+    }
+
+  num_devices = base_dev->get_num_devices_func ();
+  if (num_devices <= 0 || ord >= num_devices)
+    acc_dev_num_out_of_range (acc_device_type (base_dev->type), ord,
+			      num_devices);
+
+  if (!thr)
+    thr = goacc_new_thread ();
+
+  thr->base_dev = base_dev;
+  thr->dev = acc_dev = &base_dev[ord];
+  thr->saved_bound_dev = NULL;
+
+  thr->target_tls
+    = acc_dev->openacc.create_thread_data_func (ord);
+
+  acc_dev->openacc.async_set_async_func (acc_async_sync);
+}
+
+/* OpenACC 2.0a (3.2.12, 3.2.13) doesn't specify whether the serialization of
+   init/shutdown is per-process or per-thread.  We choose per-process.  */
+
+void
+acc_init (acc_device_t d)
+{
+  gomp_mutex_lock (&acc_device_lock);
+
+  cached_base_dev = acc_init_1 (d);
+
+  gomp_mutex_unlock (&acc_device_lock);
+
+  goacc_attach_host_thread_to_device (-1);
+}
+
+ialias (acc_init)
+
+void
+acc_shutdown (acc_device_t d)
+{
+  gomp_mutex_lock (&acc_device_lock);
+
+  acc_shutdown_1 (d);
+
+  gomp_mutex_unlock (&acc_device_lock);
+}
+
+ialias (acc_shutdown)
+
+int
+acc_get_num_devices (acc_device_t d)
+{
+  int n = 0;
+  struct gomp_device_descr *acc_dev;
+
+  if (d == acc_device_none)
+    return 0;
+
+  gomp_mutex_lock (&acc_device_lock);
+  acc_dev = resolve_device (d, false);
+  gomp_mutex_unlock (&acc_device_lock);
+
+  if (!acc_dev)
+    return 0;
+
+  n = acc_dev->get_num_devices_func ();
+  if (n < 0)
+    n = 0;
+
+  return n;
+}
+
+ialias (acc_get_num_devices)
+
+/* Set the device type for the current thread only (using the current global
+   default device number), initialising that device if necessary.  Also set the
+   default device type for new threads to D.  */
+
+void
+acc_set_device_type (acc_device_t d)
+{
+  struct gomp_device_descr *base_dev, *acc_dev;
+  struct goacc_thread *thr = goacc_thread ();
+
+  gomp_mutex_lock (&acc_device_lock);
+
+  cached_base_dev = base_dev = resolve_device (d, true);
+  acc_dev = &base_dev[goacc_device_num];
+
+  gomp_mutex_lock (&acc_dev->lock);
+  if (acc_dev->state == GOMP_DEVICE_UNINITIALIZED)
+    gomp_init_device (acc_dev);
+  gomp_mutex_unlock (&acc_dev->lock);
+
+  gomp_mutex_unlock (&acc_device_lock);
+
+  /* We're changing device type: invalidate the current thread's dev and
+     base_dev pointers.  */
+  if (thr && thr->base_dev != base_dev)
+    {
+      thr->base_dev = thr->dev = NULL;
+    }
+
+  goacc_attach_host_thread_to_device (-1);
+}
+
+ialias (acc_set_device_type)
+
+acc_device_t
+acc_get_device_type (void)
+{
+  acc_device_t res = acc_device_none;
+  struct gomp_device_descr *dev;
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (thr && thr->base_dev)
+    res = acc_device_type (thr->base_dev->type);
+  else
+    {
+      gomp_mutex_lock (&acc_device_lock);
+      dev = resolve_device (acc_device_default, true);
+      gomp_mutex_unlock (&acc_device_lock);
+      res = acc_device_type (dev->type);
+    }
+
+  assert (res != acc_device_default
+	  && res != acc_device_not_host);
+
+  return res;
+}
+
+ialias (acc_get_device_type)
+
+int
+acc_get_device_num (acc_device_t d)
+{
+  const struct gomp_device_descr *dev;
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (d >= _ACC_device_hwm)
+    gomp_fatal ("unknown device type %u", (unsigned) d);
+
+  gomp_mutex_lock (&acc_device_lock);
+  dev = resolve_device (d, true);
+  gomp_mutex_unlock (&acc_device_lock);
+
+  if (thr && thr->base_dev == dev && thr->dev)
+    return thr->dev->target_id;
+
+  return goacc_device_num;
+}
+
+ialias (acc_get_device_num)
+
+void
+acc_set_device_num (int ord, acc_device_t d)
+{
+  struct gomp_device_descr *base_dev, *acc_dev;
+  int num_devices;
+
+  if (ord < 0)
+    ord = goacc_device_num;
+
+  if ((int) d == 0)
+    /* Set whatever device is being used by the current host thread to use
+       device instance ORD.  It's unclear if this is supposed to affect other
+       host threads too (OpenACC 2.0 (3.2.4) acc_set_device_num).  */
+    goacc_attach_host_thread_to_device (ord);
+  else
+    {
+      gomp_mutex_lock (&acc_device_lock);
+
+      cached_base_dev = base_dev = resolve_device (d, true);
+
+      num_devices = base_dev->get_num_devices_func ();
+
+      if (num_devices <= 0 || ord >= num_devices)
+        acc_dev_num_out_of_range (d, ord, num_devices);
+
+      acc_dev = &base_dev[ord];
+
+      gomp_mutex_lock (&acc_dev->lock);
+      if (acc_dev->state == GOMP_DEVICE_UNINITIALIZED)
+        gomp_init_device (acc_dev);
+      gomp_mutex_unlock (&acc_dev->lock);
+
+      gomp_mutex_unlock (&acc_device_lock);
+
+      goacc_attach_host_thread_to_device (ord);
+    }
+
+  goacc_device_num = ord;
+}
+
+ialias (acc_set_device_num)
+
+int
+acc_on_device (acc_device_t dev)
+{
+  return dev == acc_device_host || dev == acc_device_none;
+}
+
+ialias (acc_on_device)
+
+attribute_hidden void
+goacc_runtime_initialize (void)
+{
+  gomp_mutex_init (&acc_device_lock);
+
+#if !(defined HAVE_TLS || defined USE_EMUTLS)
+  pthread_key_create (&goacc_tls_key, NULL);
+#endif
+
+  pthread_key_create (&goacc_cleanup_key, goacc_destroy_thread);
+
+  cached_base_dev = NULL;
+
+  goacc_threads = NULL;
+  gomp_mutex_init (&goacc_thread_lock);
+
+  /* Initialize and register the 'host' device type.  */
+  goacc_host_init ();
+}
+
+/* Compiler helper functions */
+
+attribute_hidden void
+goacc_save_and_set_bind (acc_device_t d)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  assert (!thr->saved_bound_dev);
+
+  thr->saved_bound_dev = thr->dev;
+  thr->dev = dispatchers[d];
+}
+
+attribute_hidden void
+goacc_restore_bind (void)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  thr->dev = thr->saved_bound_dev;
+  thr->saved_bound_dev = NULL;
+}
+
+/* This is called from any OpenACC support function that may need to implicitly
+   initialize the libgomp runtime, either globally or from a new host thread.
+   On exit "goacc_thread" will return a valid & populated thread block.  */
+
+attribute_hidden void
+goacc_lazy_initialize (void)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (thr && thr->dev)
+    return;
+
+  if (!cached_base_dev)
+    acc_init (acc_device_default);
+  else
+    goacc_attach_host_thread_to_device (-1);
+}
--- libgomp/oacc-int.h.jj	2016-07-13 16:57:04.400535794 +0200
+++ libgomp/oacc-int.h	2016-07-13 16:57:04.400535794 +0200
@@ -0,0 +1,106 @@
+/* OpenACC Runtime - internal declarations
+
+   Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* This file contains data types and function declarations that are not
+   part of the official OpenACC user interface.  There are declarations
+   in here that are part of the GNU OpenACC ABI, in that the compiler is
+   required to know about them and use them.
+
+   The convention is that the all caps prefix "GOACC" is used group items
+   that are part of the external ABI, and the lower case prefix "goacc"
+   is used group items that are completely private to the library.  */
+
+#ifndef OACC_INT_H
+#define OACC_INT_H 1
+
+#include "openacc.h"
+#include "config.h"
+#include <stddef.h>
+#include <stdbool.h>
+#include <stdarg.h>
+
+#ifdef HAVE_ATTRIBUTE_VISIBILITY
+# pragma GCC visibility push(hidden)
+#endif
+
+static inline enum acc_device_t
+acc_device_type (enum offload_target_type type)
+{
+  return (enum acc_device_t) type;
+}
+
+struct goacc_thread
+{
+  /* The base device for the current thread.  */
+  struct gomp_device_descr *base_dev;
+
+  /* The device for the current thread.  */
+  struct gomp_device_descr *dev;
+
+  struct gomp_device_descr *saved_bound_dev;
+
+  /* This is a linked list of data mapped by the "acc data" pragma, following
+     strictly push/pop semantics according to lexical scope.  */
+  struct target_mem_desc *mapped_data;
+
+  /* These structures form a list: this is the next thread in that list.  */
+  struct goacc_thread *next;
+
+  /* Target-specific data (used by plugin).  */
+  void *target_tls;
+};
+
+#if defined HAVE_TLS || defined USE_EMUTLS
+extern __thread struct goacc_thread *goacc_tls_data;
+static inline struct goacc_thread *
+goacc_thread (void)
+{
+  return goacc_tls_data;
+}
+#else
+extern pthread_key_t goacc_tls_key;
+static inline struct goacc_thread *
+goacc_thread (void)
+{
+  return pthread_getspecific (goacc_tls_key);
+}
+#endif
+
+void goacc_register (struct gomp_device_descr *) __GOACC_NOTHROW;
+void goacc_attach_host_thread_to_device (int);
+void goacc_runtime_initialize (void);
+void goacc_save_and_set_bind (acc_device_t);
+void goacc_restore_bind (void);
+void goacc_lazy_initialize (void);
+void goacc_host_init (void);
+
+#ifdef HAVE_ATTRIBUTE_VISIBILITY
+# pragma GCC visibility pop
+#endif
+
+#endif
--- libgomp/oacc-host.c.jj	2016-07-13 16:57:13.489423096 +0200
+++ libgomp/oacc-host.c	2016-07-13 16:57:13.489423096 +0200
@@ -0,0 +1,266 @@
+/* OpenACC Runtime Library: acc_device_host.
+
+   Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include "libgomp.h"
+#include "oacc-int.h"
+#include "gomp-constants.h"
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+
+static struct gomp_device_descr host_dispatch;
+
+static const char *
+host_get_name (void)
+{
+  return host_dispatch.name;
+}
+
+static unsigned int
+host_get_caps (void)
+{
+  return host_dispatch.capabilities;
+}
+
+static int
+host_get_type (void)
+{
+  return host_dispatch.type;
+}
+
+static int
+host_get_num_devices (void)
+{
+  return 1;
+}
+
+static bool
+host_init_device (int n __attribute__ ((unused)))
+{
+  return true;
+}
+
+static bool
+host_fini_device (int n __attribute__ ((unused)))
+{
+  return true;
+}
+
+static unsigned
+host_version (void)
+{
+  return GOMP_VERSION;
+}
+
+static int
+host_load_image (int n __attribute__ ((unused)),
+		 unsigned v __attribute__ ((unused)),
+		 const void *t __attribute__ ((unused)),
+		 struct addr_pair **r __attribute__ ((unused)))
+{
+  return 0;
+}
+
+static bool
+host_unload_image (int n __attribute__ ((unused)),
+		   unsigned v __attribute__ ((unused)),
+		   const void *t __attribute__ ((unused)))
+{
+  return true;
+}
+
+static void *
+host_alloc (int n __attribute__ ((unused)), size_t s)
+{
+  return gomp_malloc (s);
+}
+
+static bool
+host_free (int n __attribute__ ((unused)), void *p)
+{
+  free (p);
+  return true;
+}
+
+static bool
+host_dev2host (int n __attribute__ ((unused)),
+	       void *h __attribute__ ((unused)),
+	       const void *d __attribute__ ((unused)),
+	       size_t s __attribute__ ((unused)))
+{
+  return true;
+}
+
+static bool
+host_host2dev (int n __attribute__ ((unused)),
+	       void *d __attribute__ ((unused)),
+	       const void *h __attribute__ ((unused)),
+	       size_t s __attribute__ ((unused)))
+{
+  return true;
+}
+
+static void
+host_run (int n __attribute__ ((unused)), void *fn_ptr, void *vars,
+	  void **args __attribute__((unused)))
+{
+  void (*fn)(void *) = (void (*)(void *)) fn_ptr;
+
+  fn (vars);
+}
+
+static void
+host_openacc_exec (void (*fn) (void *),
+		   size_t mapnum __attribute__ ((unused)),
+		   void **hostaddrs,
+		   void **devaddrs __attribute__ ((unused)),
+		   int async __attribute__ ((unused)),
+		   unsigned *dims __attribute ((unused)),
+		   void *targ_mem_desc __attribute__ ((unused)))
+{
+  fn (hostaddrs);
+}
+
+static void
+host_openacc_register_async_cleanup (void *targ_mem_desc __attribute__ ((unused)),
+				     int async __attribute__ ((unused)))
+{
+}
+
+static int
+host_openacc_async_test (int async __attribute__ ((unused)))
+{
+  return 1;
+}
+
+static int
+host_openacc_async_test_all (void)
+{
+  return 1;
+}
+
+static void
+host_openacc_async_wait (int async __attribute__ ((unused)))
+{
+}
+
+static void
+host_openacc_async_wait_async (int async1 __attribute__ ((unused)),
+			       int async2 __attribute__ ((unused)))
+{
+}
+
+static void
+host_openacc_async_wait_all (void)
+{
+}
+
+static void
+host_openacc_async_wait_all_async (int async __attribute__ ((unused)))
+{
+}
+
+static void
+host_openacc_async_set_async (int async __attribute__ ((unused)))
+{
+}
+
+static void *
+host_openacc_create_thread_data (int ord __attribute__ ((unused)))
+{
+  return NULL;
+}
+
+static void
+host_openacc_destroy_thread_data (void *tls_data __attribute__ ((unused)))
+{
+}
+
+static struct gomp_device_descr host_dispatch =
+  {
+    .name = "host",
+    .capabilities = (GOMP_OFFLOAD_CAP_SHARED_MEM
+		     | GOMP_OFFLOAD_CAP_NATIVE_EXEC
+		     | GOMP_OFFLOAD_CAP_OPENACC_200),
+    .target_id = 0,
+    .type = OFFLOAD_TARGET_TYPE_HOST,
+
+    .get_name_func = host_get_name,
+    .get_caps_func = host_get_caps,
+    .get_type_func = host_get_type,
+    .get_num_devices_func = host_get_num_devices,
+    .init_device_func = host_init_device,
+    .fini_device_func = host_fini_device,
+    .version_func = host_version,
+    .load_image_func = host_load_image,
+    .unload_image_func = host_unload_image,
+    .alloc_func = host_alloc,
+    .free_func = host_free,
+    .dev2host_func = host_dev2host,
+    .host2dev_func = host_host2dev,
+    .run_func = host_run,
+
+    .mem_map = { NULL },
+    /* .lock initilized in goacc_host_init.  */
+    .state = GOMP_DEVICE_UNINITIALIZED,
+
+    .openacc = {
+      .data_environ = NULL,
+
+      .exec_func = host_openacc_exec,
+
+      .register_async_cleanup_func = host_openacc_register_async_cleanup,
+
+      .async_test_func = host_openacc_async_test,
+      .async_test_all_func = host_openacc_async_test_all,
+      .async_wait_func = host_openacc_async_wait,
+      .async_wait_async_func = host_openacc_async_wait_async,
+      .async_wait_all_func = host_openacc_async_wait_all,
+      .async_wait_all_async_func = host_openacc_async_wait_all_async,
+      .async_set_async_func = host_openacc_async_set_async,
+
+      .create_thread_data_func = host_openacc_create_thread_data,
+      .destroy_thread_data_func = host_openacc_destroy_thread_data,
+
+      .cuda = {
+	.get_current_device_func = NULL,
+	.get_current_context_func = NULL,
+	.get_stream_func = NULL,
+	.set_stream_func = NULL,
+      }
+    }
+  };
+
+/* Initialize and register this device type.  */
+void
+goacc_host_init (void)
+{
+  gomp_mutex_init (&host_dispatch.lock);
+  goacc_register (&host_dispatch);
+}
--- libgomp/oacc-parallel.c.jj	2016-07-13 16:57:04.399535807 +0200
+++ libgomp/oacc-parallel.c	2016-07-14 18:53:06.694996381 +0200
@@ -0,0 +1,241 @@
+/* Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* This file handles OpenACC constructs.  */
+
+#include "openacc.h"
+#include "libgomp.h"
+#include "libgomp_g.h"
+#include "gomp-constants.h"
+#include "oacc-int.h"
+#ifdef HAVE_INTTYPES_H
+# include <inttypes.h>  /* For PRIu64.  */
+#endif
+#include <string.h>
+#include <stdarg.h>
+#include <assert.h>
+
+static void goacc_wait (int async, int num_waits, va_list *ap);
+
+
+/* Launch a possibly offloaded function on DEVICE.  FN is the host fn
+   address.  MAPNUM, HOSTADDRS, SIZES & KINDS  describe the memory
+   blocks to be copied to/from the device.  Varadic arguments are
+   keyed optional parameters terminated with a zero.  */
+
+void
+GOACC_parallel_keyed (int device, void (*fn) (void *),
+		      size_t mapnum, void **hostaddrs, size_t *sizes,
+		      unsigned short *kinds, ...)
+{
+  bool host_fallback = device == GOMP_DEVICE_HOST_FALLBACK;
+  struct goacc_thread *thr;
+  struct gomp_device_descr *acc_dev;
+
+#ifdef HAVE_INTTYPES_H
+  gomp_debug (0, "%s: mapnum=%"PRIu64", hostaddrs=%p, size=%p, kinds=%p\n",
+	      __FUNCTION__, (uint64_t) mapnum, hostaddrs, sizes, kinds);
+#else
+  gomp_debug (0, "%s: mapnum=%lu, hostaddrs=%p, sizes=%p, kinds=%p\n",
+	      __FUNCTION__, (unsigned long) mapnum, hostaddrs, sizes, kinds);
+#endif
+  goacc_lazy_initialize ();
+
+  thr = goacc_thread ();
+  acc_dev = thr->dev;
+
+  /* Host fallback if "if" clause is false or if the current device is set to
+     the host.  */
+  if (host_fallback)
+    {
+      goacc_save_and_set_bind (acc_device_host);
+      fn (hostaddrs);
+      goacc_restore_bind ();
+      return;
+    }
+  else if (acc_device_type (acc_dev->type) == acc_device_host)
+    {
+      fn (hostaddrs);
+      return;
+    }
+
+  /* acc_device_host is the only supported device type.  */
+}
+
+/* Legacy entry point, only provide host execution.  */
+
+void
+GOACC_parallel (int device, void (*fn) (void *),
+		size_t mapnum, void **hostaddrs, size_t *sizes,
+		unsigned short *kinds,
+		int num_gangs, int num_workers, int vector_length,
+		int async, int num_waits, ...)
+{
+  goacc_save_and_set_bind (acc_device_host);
+  fn (hostaddrs);
+  goacc_restore_bind ();
+}
+
+void
+GOACC_data_start (int device, size_t mapnum,
+		  void **hostaddrs, size_t *sizes, unsigned short *kinds)
+{
+  goacc_lazy_initialize ();
+}
+
+void
+GOACC_data_end (void)
+{
+  gomp_debug (0, "  %s: restore mappings\n", __FUNCTION__);
+  gomp_debug (0, "  %s: mappings restored\n", __FUNCTION__);
+}
+
+void
+GOACC_enter_exit_data (int device, size_t mapnum,
+		       void **hostaddrs, size_t *sizes, unsigned short *kinds,
+		       int async, int num_waits, ...)
+{
+  goacc_lazy_initialize ();
+}
+
+static void
+goacc_wait (int async, int num_waits, va_list *ap)
+{
+  struct goacc_thread *thr = goacc_thread ();
+  struct gomp_device_descr *acc_dev = thr->dev;
+
+  while (num_waits--)
+    {
+      int qid = va_arg (*ap, int);
+
+      if (acc_async_test (qid))
+	continue;
+
+      if (async == acc_async_sync)
+	acc_wait (qid);
+      else if (qid == async)
+	;/* If we're waiting on the same asynchronous queue as we're
+	    launching on, the queue itself will order work as
+	    required, so there's no need to wait explicitly.  */
+      else
+	acc_dev->openacc.async_wait_async_func (qid, async);
+    }
+}
+
+void
+GOACC_update (int device, size_t mapnum,
+	      void **hostaddrs, size_t *sizes, unsigned short *kinds,
+	      int async, int num_waits, ...)
+{
+  goacc_lazy_initialize ();
+}
+
+void
+GOACC_wait (int async, int num_waits, ...)
+{
+  if (num_waits)
+    {
+      va_list ap;
+
+      va_start (ap, num_waits);
+      goacc_wait (async, num_waits, &ap);
+      va_end (ap);
+    }
+  else if (async == acc_async_sync)
+    acc_wait_all ();
+  else if (async == acc_async_noval)
+    goacc_thread ()->dev->openacc.async_wait_all_async_func (acc_async_noval);
+}
+
+int
+GOACC_get_num_threads (void)
+{
+  return 1;
+}
+
+int
+GOACC_get_thread_num (void)
+{
+  return 0;
+}
+
+void
+GOACC_declare (int device, size_t mapnum,
+	       void **hostaddrs, size_t *sizes, unsigned short *kinds)
+{
+  int i;
+
+  for (i = 0; i < mapnum; i++)
+    {
+      unsigned char kind = kinds[i] & 0xff;
+
+      if (kind == GOMP_MAP_POINTER || kind == GOMP_MAP_TO_PSET)
+	continue;
+
+      switch (kind)
+	{
+	  case GOMP_MAP_FORCE_ALLOC:
+	  case GOMP_MAP_FORCE_FROM:
+	  case GOMP_MAP_FORCE_TO:
+	  case GOMP_MAP_POINTER:
+	  case GOMP_MAP_DELETE:
+	    GOACC_enter_exit_data (device, 1, &hostaddrs[i], &sizes[i],
+				   &kinds[i], 0, 0);
+	    break;
+
+	  case GOMP_MAP_FORCE_DEVICEPTR:
+	    break;
+
+	  case GOMP_MAP_ALLOC:
+	    if (!acc_is_present (hostaddrs[i], sizes[i]))
+	      GOACC_enter_exit_data (device, 1, &hostaddrs[i], &sizes[i],
+				     &kinds[i], 0, 0);
+	    break;
+
+	  case GOMP_MAP_TO:
+	    GOACC_enter_exit_data (device, 1, &hostaddrs[i], &sizes[i],
+				   &kinds[i], 0, 0);
+
+	    break;
+
+	  case GOMP_MAP_FROM:
+	    kinds[i] = GOMP_MAP_FORCE_FROM;
+	    GOACC_enter_exit_data (device, 1, &hostaddrs[i], &sizes[i],
+				   &kinds[i], 0, 0);
+	    break;
+
+	  case GOMP_MAP_FORCE_PRESENT:
+	    if (!acc_is_present (hostaddrs[i], sizes[i]))
+	      gomp_fatal ("[%p,%ld] is not mapped", hostaddrs[i],
+			  (unsigned long) sizes[i]);
+	    break;
+
+	  default:
+	    assert (0);
+	    break;
+	}
+    }
+}
--- libgomp/oacc-cuda.c.jj	2016-07-13 16:57:04.432535397 +0200
+++ libgomp/oacc-cuda.c	2016-07-13 16:57:04.432535397 +0200
@@ -0,0 +1,86 @@
+/* OpenACC Runtime Library: CUDA support glue.
+
+   Copyright (C) 2014-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include "openacc.h"
+#include "config.h"
+#include "libgomp.h"
+#include "oacc-int.h"
+
+void *
+acc_get_current_cuda_device (void)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (thr && thr->dev && thr->dev->openacc.cuda.get_current_device_func)
+    return thr->dev->openacc.cuda.get_current_device_func ();
+
+  return NULL;
+}
+
+void *
+acc_get_current_cuda_context (void)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (thr && thr->dev && thr->dev->openacc.cuda.get_current_context_func)
+    return thr->dev->openacc.cuda.get_current_context_func ();
+
+  return NULL;
+}
+
+void *
+acc_get_cuda_stream (int async)
+{
+  struct goacc_thread *thr = goacc_thread ();
+
+  if (async < 0)
+    return NULL;
+
+  if (thr && thr->dev && thr->dev->openacc.cuda.get_stream_func)
+    return thr->dev->openacc.cuda.get_stream_func (async);
+
+  return NULL;
+}
+
+int
+acc_set_cuda_stream (int async, void *stream)
+{
+  struct goacc_thread *thr;
+
+  if (async < 0 || stream == NULL)
+    return 0;
+
+  goacc_lazy_initialize ();
+
+  thr = goacc_thread ();
+
+  if (thr && thr->dev && thr->dev->openacc.cuda.set_stream_func)
+    return thr->dev->openacc.cuda.set_stream_func (async, stream);
+
+  return -1;
+}
--- libgomp/openacc_lib.h.jj	2016-07-13 16:57:13.486423134 +0200
+++ libgomp/openacc_lib.h	2016-07-13 16:57:13.486423134 +0200
@@ -0,0 +1,382 @@
+!  OpenACC Runtime Library Definitions.			-*- mode: fortran -*-
+
+!  Copyright (C) 2014-2016 Free Software Foundation, Inc.
+
+!  Contributed by Tobias Burnus <burnus@net-b.de>
+!              and Mentor Embedded.
+
+!  This file is part of the GNU Offloading and Multi Processing Library
+!  (libgomp).
+
+!  Libgomp is free software; you can redistribute it and/or modify it
+!  under the terms of the GNU General Public License as published by
+!  the Free Software Foundation; either version 3, or (at your option)
+!  any later version.
+
+!  Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+!  WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+!  FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+!  more details.
+
+!  Under Section 7 of GPL version 3, you are granted additional
+!  permissions described in the GCC Runtime Library Exception, version
+!  3.1, as published by the Free Software Foundation.
+
+!  You should have received a copy of the GNU General Public License and
+!  a copy of the GCC Runtime Library Exception along with this program;
+!  see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+!  <http://www.gnu.org/licenses/>.
+
+! NOTE: Due to the use of dimension (..), the code only works when compiled
+! with -std=f2008ts/gnu/legacy but not with other standard settings.
+! Alternatively, the user can use the module version, which permits
+! compilation with -std=f95.
+
+      integer, parameter :: acc_device_kind = 4
+
+!     Keep in sync with include/gomp-constants.h.
+      integer (acc_device_kind), parameter :: acc_device_none = 0
+      integer (acc_device_kind), parameter :: acc_device_default = 1
+      integer (acc_device_kind), parameter :: acc_device_host = 2
+!     integer (acc_device_kind), parameter :: acc_device_host_nonshm = 3
+!     removed.
+      integer (acc_device_kind), parameter :: acc_device_not_host = 4
+      integer (acc_device_kind), parameter :: acc_device_nvidia = 5
+
+      integer, parameter :: acc_handle_kind = 4
+
+!     Keep in sync with include/gomp-constants.h.
+      integer (acc_handle_kind), parameter :: acc_async_noval = -1
+      integer (acc_handle_kind), parameter :: acc_async_sync = -2
+
+      integer, parameter :: openacc_version = 201306
+
+      interface acc_get_num_devices
+        function acc_get_num_devices_h (d)
+          import acc_device_kind
+          integer acc_get_num_devices_h
+          integer (acc_device_kind) d
+        end function
+      end interface
+
+      interface acc_set_device_type
+        subroutine acc_set_device_type_h (d)
+          import acc_device_kind
+          integer (acc_device_kind) d
+        end subroutine
+      end interface
+
+      interface acc_get_device_type
+        function acc_get_device_type_h ()
+          import acc_device_kind
+          integer (acc_device_kind) acc_get_device_type_h
+        end function
+      end interface
+
+      interface acc_set_device_num
+        subroutine acc_set_device_num_h (n, d)
+          import acc_device_kind
+          integer n
+          integer (acc_device_kind) d
+        end subroutine
+      end interface
+
+      interface acc_get_device_num
+        function acc_get_device_num_h (d)
+          import acc_device_kind
+          integer acc_get_device_num_h
+          integer (acc_device_kind) d
+        end function
+      end interface
+
+      interface acc_async_test
+        function acc_async_test_h (a)
+          logical acc_async_test_h
+          integer a
+        end function
+      end interface
+
+      interface acc_async_test_all
+        function acc_async_test_all_h ()
+          logical acc_async_test_all_h
+        end function
+      end interface
+
+      interface acc_wait
+        subroutine acc_wait_h (a)
+          integer a
+        end subroutine
+      end interface
+
+      interface acc_wait_async
+        subroutine acc_wait_async_h (a1, a2)
+          integer a1, a2
+        end subroutine
+      end interface
+
+      interface acc_wait_all
+        subroutine acc_wait_all_h ()
+        end subroutine
+      end interface
+
+      interface acc_wait_all_async
+        subroutine acc_wait_all_async_h (a)
+          integer a
+        end subroutine
+      end interface
+
+      interface acc_init
+        subroutine acc_init_h (devicetype)
+          import acc_device_kind
+          integer (acc_device_kind) devicetype
+        end subroutine
+      end interface
+
+      interface acc_shutdown
+        subroutine acc_shutdown_h (devicetype)
+          import acc_device_kind
+          integer (acc_device_kind) devicetype
+        end subroutine
+      end interface
+
+      interface acc_on_device
+        function acc_on_device_h (devicetype)
+          import acc_device_kind
+          logical acc_on_device_h
+          integer (acc_device_kind) devicetype
+        end function
+      end interface
+
+      ! acc_malloc: Only available in C/C++
+      ! acc_free: Only available in C/C++
+
+      interface acc_copyin
+        subroutine acc_copyin_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_copyin_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_copyin_array_h (a)
+          type (*), dimension (..), contiguous :: a
+          end subroutine
+      end interface
+
+      interface acc_present_or_copyin
+        subroutine acc_present_or_copyin_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_present_or_copyin_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_present_or_copyin_array_h (a)
+          type (*), dimension (..), contiguous :: a
+          end subroutine
+      end interface
+
+      interface acc_pcopyin
+        subroutine acc_pcopyin_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_pcopyin_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_pcopyin_array_h (a)
+          type (*), dimension (..), contiguous :: a
+          end subroutine
+      end interface
+
+      interface acc_create
+        subroutine acc_create_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_create_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_create_array_h (a)
+          type (*), dimension (..), contiguous :: a
+          end subroutine
+      end interface
+
+      interface acc_present_or_create
+        subroutine acc_present_or_create_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_present_or_create_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_present_or_create_array_h (a)
+          type (*), dimension (..), contiguous :: a
+          end subroutine
+      end interface
+
+      interface acc_pcreate
+        subroutine acc_pcreate_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_pcreate_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_pcreate_array_h (a)
+          type (*), dimension (..), contiguous :: a
+          end subroutine
+      end interface
+
+      interface acc_copyout
+        subroutine acc_copyout_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_copyout_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_copyout_array_h (a)
+          type (*), dimension (..), contiguous :: a
+        end subroutine
+      end interface
+
+      interface acc_delete
+        subroutine acc_delete_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_delete_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_delete_array_h (a)
+          type (*), dimension (..), contiguous :: a
+        end subroutine
+      end interface
+
+      interface acc_update_device
+        subroutine acc_update_device_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_update_device_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_update_device_array_h (a)
+          type (*), dimension (..), contiguous :: a
+        end subroutine
+      end interface
+
+      interface acc_update_self
+        subroutine acc_update_self_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end subroutine
+
+        subroutine acc_update_self_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end subroutine
+
+        subroutine acc_update_self_array_h (a)
+          type (*), dimension (..), contiguous :: a
+        end subroutine
+      end interface
+
+      ! acc_map_data: Only available in C/C++
+      ! acc_unmap_data: Only available in C/C++
+      ! acc_deviceptr: Only available in C/C++
+      ! acc_ostptr: Only available in C/C++
+
+      interface acc_is_present
+        function acc_is_present_32_h (a, len)
+          use iso_c_binding, only: c_int32_t
+          logical acc_is_present_32_h
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int32_t) len
+        end function
+
+        function acc_is_present_64_h (a, len)
+          use iso_c_binding, only: c_int64_t
+          logical acc_is_present_64_h
+          !GCC$ ATTRIBUTES NO_ARG_CHECK :: a
+          type (*), dimension (*) :: a
+          integer (c_int64_t) len
+        end function
+
+        function acc_is_present_array_h (a)
+          logical acc_is_present_array_h
+          type (*), dimension (..), contiguous :: a
+        end function
+      end interface
+
+      ! acc_memcpy_to_device: Only available in C/C++
+      ! acc_memcpy_from_device: Only available in C/C++
--- libgomp/gomp-constants.h.jj	2016-07-14 16:02:47.212545826 +0200
+++ libgomp/gomp-constants.h	2016-05-26 21:04:40.000000000 +0200
@@ -0,0 +1,259 @@
+/* Communication between GCC and libgomp.
+
+   Copyright (C) 2014-2015 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef GOMP_CONSTANTS_H
+#define GOMP_CONSTANTS_H 1
+
+/* Memory mapping types.  */
+
+/* One byte.  */
+#define GOMP_MAP_LAST			(1 << 8)
+
+#define GOMP_MAP_FLAG_TO		(1 << 0)
+#define GOMP_MAP_FLAG_FROM		(1 << 1)
+/* Special map kinds, enumerated starting here.  */
+#define GOMP_MAP_FLAG_SPECIAL_0		(1 << 2)
+#define GOMP_MAP_FLAG_SPECIAL_1		(1 << 3)
+#define GOMP_MAP_FLAG_SPECIAL_2		(1 << 4)
+#define GOMP_MAP_FLAG_SPECIAL		(GOMP_MAP_FLAG_SPECIAL_1 \
+					 | GOMP_MAP_FLAG_SPECIAL_0)
+/* Flag to force a specific behavior (or else, trigger a run-time error).  */
+#define GOMP_MAP_FLAG_FORCE		(1 << 7)
+
+enum gomp_map_kind
+  {
+    /* If not already present, allocate.  */
+    GOMP_MAP_ALLOC =			0,
+    /* ..., and copy to device.  */
+    GOMP_MAP_TO =			(GOMP_MAP_ALLOC | GOMP_MAP_FLAG_TO),
+    /* ..., and copy from device.  */
+    GOMP_MAP_FROM =			(GOMP_MAP_ALLOC | GOMP_MAP_FLAG_FROM),
+    /* ..., and copy to and from device.  */
+    GOMP_MAP_TOFROM =			(GOMP_MAP_TO | GOMP_MAP_FROM),
+    /* The following kind is an internal only map kind, used for pointer based
+       array sections.  OMP_CLAUSE_SIZE for these is not the pointer size,
+       which is implicitly POINTER_SIZE_UNITS, but the bias.  */
+    GOMP_MAP_POINTER =			(GOMP_MAP_FLAG_SPECIAL_0 | 0),
+    /* Also internal, behaves like GOMP_MAP_TO, but additionally any
+       GOMP_MAP_POINTER records consecutive after it which have addresses
+       falling into that range will not be ignored if GOMP_MAP_TO_PSET wasn't
+       mapped already.  */
+    GOMP_MAP_TO_PSET =			(GOMP_MAP_FLAG_SPECIAL_0 | 1),
+    /* Must already be present.  */
+    GOMP_MAP_FORCE_PRESENT =		(GOMP_MAP_FLAG_SPECIAL_0 | 2),
+    /* Deallocate a mapping, without copying from device.  */
+    GOMP_MAP_DELETE =			(GOMP_MAP_FLAG_SPECIAL_0 | 3),
+    /* Is a device pointer.  OMP_CLAUSE_SIZE for these is unused; is implicitly
+       POINTER_SIZE_UNITS.  */
+    GOMP_MAP_FORCE_DEVICEPTR =		(GOMP_MAP_FLAG_SPECIAL_1 | 0),
+    /* Do not map, copy bits for firstprivate instead.  */
+    /* OpenACC device_resident.  */
+    GOMP_MAP_DEVICE_RESIDENT =		(GOMP_MAP_FLAG_SPECIAL_1 | 1),
+    /* OpenACC link.  */
+    GOMP_MAP_LINK =			(GOMP_MAP_FLAG_SPECIAL_1 | 2),
+    /* Allocate.  */
+    GOMP_MAP_FIRSTPRIVATE =		(GOMP_MAP_FLAG_SPECIAL | 0),
+    /* Similarly, but store the value in the pointer rather than
+       pointed by the pointer.  */
+    GOMP_MAP_FIRSTPRIVATE_INT =		(GOMP_MAP_FLAG_SPECIAL | 1),
+    /* Pointer translate host address into device address and copy that
+       back to host.  */
+    GOMP_MAP_USE_DEVICE_PTR =		(GOMP_MAP_FLAG_SPECIAL | 2),
+    /* Allocate a zero length array section.  Prefer next non-zero length
+       mapping over previous non-zero length mapping over zero length mapping
+       at the address.  If not already mapped, do nothing (and pointer translate
+       to NULL).  */
+    GOMP_MAP_ZERO_LEN_ARRAY_SECTION = 	(GOMP_MAP_FLAG_SPECIAL | 3),
+    /* Allocate.  */
+    GOMP_MAP_FORCE_ALLOC =		(GOMP_MAP_FLAG_FORCE | GOMP_MAP_ALLOC),
+    /* ..., and copy to device.  */
+    GOMP_MAP_FORCE_TO =			(GOMP_MAP_FLAG_FORCE | GOMP_MAP_TO),
+    /* ..., and copy from device.  */
+    GOMP_MAP_FORCE_FROM =		(GOMP_MAP_FLAG_FORCE | GOMP_MAP_FROM),
+    /* ..., and copy to and from device.  */
+    GOMP_MAP_FORCE_TOFROM =		(GOMP_MAP_FLAG_FORCE | GOMP_MAP_TOFROM),
+    /* If not already present, allocate.  And unconditionally copy to
+       device.  */
+    GOMP_MAP_ALWAYS_TO =		(GOMP_MAP_FLAG_SPECIAL_2 | GOMP_MAP_TO),
+    /* If not already present, allocate.  And unconditionally copy from
+       device.  */
+    GOMP_MAP_ALWAYS_FROM =		(GOMP_MAP_FLAG_SPECIAL_2
+					 | GOMP_MAP_FROM),
+    /* If not already present, allocate.  And unconditionally copy to and from
+       device.  */
+    GOMP_MAP_ALWAYS_TOFROM =		(GOMP_MAP_FLAG_SPECIAL_2
+					 | GOMP_MAP_TOFROM),
+    /* Map a sparse struct; the address is the base of the structure, alignment
+       it's required alignment, and size is the number of adjacent entries
+       that belong to the struct.  The adjacent entries should be sorted by
+       increasing address, so it is easy to determine lowest needed address
+       (address of the first adjacent entry) and highest needed address
+       (address of the last adjacent entry plus its size).  */
+    GOMP_MAP_STRUCT =			(GOMP_MAP_FLAG_SPECIAL_2
+					 | GOMP_MAP_FLAG_SPECIAL | 0),
+    /* On a location of a pointer/reference that is assumed to be already mapped
+       earlier, store the translated address of the preceeding mapping.
+       No refcount is bumped by this, and the store is done unconditionally.  */
+    GOMP_MAP_ALWAYS_POINTER =		(GOMP_MAP_FLAG_SPECIAL_2
+					 | GOMP_MAP_FLAG_SPECIAL | 1),
+    /* Forced deallocation of zero length array section.  */
+    GOMP_MAP_DELETE_ZERO_LEN_ARRAY_SECTION
+      =					(GOMP_MAP_FLAG_SPECIAL_2
+					 | GOMP_MAP_FLAG_SPECIAL | 3),
+    /* Decrement usage count and deallocate if zero.  */
+    GOMP_MAP_RELEASE =			(GOMP_MAP_FLAG_SPECIAL_2
+					 | GOMP_MAP_DELETE),
+
+    /* Internal to GCC, not used in libgomp.  */
+    /* Do not map, but pointer assign a pointer instead.  */
+    GOMP_MAP_FIRSTPRIVATE_POINTER =	(GOMP_MAP_LAST | 1),
+    /* Do not map, but pointer assign a reference instead.  */
+    GOMP_MAP_FIRSTPRIVATE_REFERENCE =	(GOMP_MAP_LAST | 2)
+  };
+
+#define GOMP_MAP_COPY_TO_P(X) \
+  (!((X) & GOMP_MAP_FLAG_SPECIAL) \
+   && ((X) & GOMP_MAP_FLAG_TO))
+
+#define GOMP_MAP_COPY_FROM_P(X) \
+  (!((X) & GOMP_MAP_FLAG_SPECIAL) \
+   && ((X) & GOMP_MAP_FLAG_FROM))
+
+#define GOMP_MAP_POINTER_P(X) \
+  ((X) == GOMP_MAP_POINTER)
+
+#define GOMP_MAP_ALWAYS_TO_P(X) \
+  (((X) == GOMP_MAP_ALWAYS_TO) || ((X) == GOMP_MAP_ALWAYS_TOFROM))
+
+#define GOMP_MAP_ALWAYS_FROM_P(X) \
+  (((X) == GOMP_MAP_ALWAYS_FROM) || ((X) == GOMP_MAP_ALWAYS_TOFROM))
+
+#define GOMP_MAP_ALWAYS_P(X) \
+  (GOMP_MAP_ALWAYS_TO_P (X) || ((X) == GOMP_MAP_ALWAYS_FROM))
+
+
+/* Asynchronous behavior.  Keep in sync with
+   libgomp/{openacc.h,openacc.f90,openacc_lib.h}:acc_async_t.  */
+
+#define GOMP_ASYNC_NOVAL		-1
+#define GOMP_ASYNC_SYNC			-2
+
+
+/* Device codes.  Keep in sync with
+   libgomp/{openacc.h,openacc.f90,openacc_lib.h}:acc_device_t as well as
+   libgomp/libgomp-plugin.h.  */
+#define GOMP_DEVICE_NONE		0
+#define GOMP_DEVICE_DEFAULT		1
+#define GOMP_DEVICE_HOST		2
+/* #define GOMP_DEVICE_HOST_NONSHM	3 removed.  */
+#define GOMP_DEVICE_NOT_HOST		4
+#define GOMP_DEVICE_NVIDIA_PTX		5
+#define GOMP_DEVICE_INTEL_MIC		6
+#define GOMP_DEVICE_HSA			7
+
+#define GOMP_DEVICE_ICV			-1
+#define GOMP_DEVICE_HOST_FALLBACK	-2
+
+/* GOMP_task/GOMP_taskloop* flags argument.  */
+#define GOMP_TASK_FLAG_UNTIED		(1 << 0)
+#define GOMP_TASK_FLAG_FINAL		(1 << 1)
+#define GOMP_TASK_FLAG_MERGEABLE	(1 << 2)
+#define GOMP_TASK_FLAG_DEPEND		(1 << 3)
+#define GOMP_TASK_FLAG_PRIORITY		(1 << 4)
+#define GOMP_TASK_FLAG_UP		(1 << 8)
+#define GOMP_TASK_FLAG_GRAINSIZE	(1 << 9)
+#define GOMP_TASK_FLAG_IF		(1 << 10)
+#define GOMP_TASK_FLAG_NOGROUP		(1 << 11)
+
+/* GOMP_target{_ext,update_ext,enter_exit_data} flags argument.  */
+#define GOMP_TARGET_FLAG_NOWAIT		(1 << 0)
+#define GOMP_TARGET_FLAG_EXIT_DATA	(1 << 1)
+/* Internal to libgomp.  */
+#define GOMP_TARGET_FLAG_UPDATE		(1U << 31)
+
+/* Versions of libgomp and device-specific plugins.  GOMP_VERSION
+   should be incremented whenever an ABI-incompatible change is introduced
+   to the plugin interface defined in libgomp/libgomp.h.  */
+#define GOMP_VERSION	1
+#define GOMP_VERSION_NVIDIA_PTX 1
+#define GOMP_VERSION_INTEL_MIC 0
+#define GOMP_VERSION_HSA 0
+
+#define GOMP_VERSION_PACK(LIB, DEV) (((LIB) << 16) | (DEV))
+#define GOMP_VERSION_LIB(PACK) (((PACK) >> 16) & 0xffff)
+#define GOMP_VERSION_DEV(PACK) ((PACK) & 0xffff)
+
+#define GOMP_DIM_GANG	0
+#define GOMP_DIM_WORKER	1
+#define GOMP_DIM_VECTOR	2
+#define GOMP_DIM_MAX	3
+#define GOMP_DIM_MASK(X) (1u << (X))
+
+/* Varadic launch arguments.  End of list is marked by a zero.  */
+#define GOMP_LAUNCH_DIM		1  /* Launch dimensions, op = mask */
+#define GOMP_LAUNCH_ASYNC	2  /* Async, op = cst val if not MAX  */
+#define GOMP_LAUNCH_WAIT	3  /* Waits, op = num waits.  */
+#define GOMP_LAUNCH_CODE_SHIFT	28
+#define GOMP_LAUNCH_DEVICE_SHIFT 16
+#define GOMP_LAUNCH_OP_SHIFT 0
+#define GOMP_LAUNCH_PACK(CODE,DEVICE,OP)	\
+  (((CODE) << GOMP_LAUNCH_CODE_SHIFT)		\
+   | ((DEVICE) << GOMP_LAUNCH_DEVICE_SHIFT)	\
+   | ((OP) << GOMP_LAUNCH_OP_SHIFT))
+#define GOMP_LAUNCH_CODE(X) (((X) >> GOMP_LAUNCH_CODE_SHIFT) & 0xf)
+#define GOMP_LAUNCH_DEVICE(X) (((X) >> GOMP_LAUNCH_DEVICE_SHIFT) & 0xfff)
+#define GOMP_LAUNCH_OP(X) (((X) >> GOMP_LAUNCH_OP_SHIFT) & 0xffff)
+#define GOMP_LAUNCH_OP_MAX 0xffff
+
+/* Bitmask to apply in order to find out the intended device of a target
+   argument.  */
+#define GOMP_TARGET_ARG_DEVICE_MASK		((1 << 7) - 1)
+/* The target argument is significant for all devices.  */
+#define GOMP_TARGET_ARG_DEVICE_ALL		0
+
+/* Flag set when the subsequent element in the device-specific argument
+   values.  */
+#define GOMP_TARGET_ARG_SUBSEQUENT_PARAM	(1 << 7)
+
+/* Bitmask to apply to a target argument to find out the value identifier.  */
+#define GOMP_TARGET_ARG_ID_MASK			(((1 << 8) - 1) << 8)
+/* Target argument index of NUM_TEAMS.  */
+#define GOMP_TARGET_ARG_NUM_TEAMS		(1 << 8)
+/* Target argument index of THREAD_LIMIT.  */
+#define GOMP_TARGET_ARG_THREAD_LIMIT		(2 << 8)
+
+/* If the value is directly embeded in target argument, it should be a 16-bit
+   at most and shifted by this many bits.  */
+#define GOMP_TARGET_ARG_VALUE_SHIFT		16
+
+/* HSA specific data structures.  */
+
+/* Identifiers of device-specific target arguments.  */
+#define GOMP_TARGET_ARG_HSA_KERNEL_ATTRIBUTES	(1 << 8)
+
+#endif
--- libgomp/oacc-mem.c.jj	2016-07-13 16:57:04.433535385 +0200
+++ libgomp/oacc-mem.c	2016-07-14 15:39:44.644631308 +0200
@@ -0,0 +1,204 @@
+/* OpenACC Runtime initialization routines
+
+   Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include "openacc.h"
+#include "config.h"
+#include "libgomp.h"
+#include "gomp-constants.h"
+#include "oacc-int.h"
+#include <stdint.h>
+#include <string.h>
+#include <assert.h>
+
+/* OpenACC is silent on how memory exhaustion is indicated.  We return
+   NULL.  */
+
+void *
+acc_malloc (size_t s)
+{
+  if (!s)
+    return NULL;
+
+  goacc_lazy_initialize ();
+  return malloc (s);
+}
+
+/* OpenACC 2.0a (3.2.16) doesn't specify what to do in the event
+   the device address is mapped. We choose to check if it mapped,
+   and if it is, to unmap it. */
+void
+acc_free (void *d)
+{
+  return free (d);
+}
+
+void
+acc_memcpy_to_device (void *d, void *h, size_t s)
+{
+  memmove (d, h, s);
+}
+
+void
+acc_memcpy_from_device (void *h, void *d, size_t s)
+{
+  memmove (h, d, s);
+}
+
+/* Return the device pointer that corresponds to host data H.  Or NULL
+   if no mapping.  */
+
+void *
+acc_deviceptr (void *h)
+{
+  goacc_lazy_initialize ();
+  return h;
+}
+
+/* Return the host pointer that corresponds to device data D.  Or NULL
+   if no mapping.  */
+
+void *
+acc_hostptr (void *d)
+{
+  goacc_lazy_initialize ();
+  return d;
+}
+
+/* Return 1 if host data [H,+S] is present on the device.  */
+
+int
+acc_is_present (void *h, size_t s)
+{
+  if (!s || !h)
+    return 0;
+
+  goacc_lazy_initialize ();
+  return h != NULL;
+}
+
+/* Create a mapping for host [H,+S] -> device [D,+S] */
+
+void
+acc_map_data (void *h, void *d, size_t s)
+{
+  goacc_lazy_initialize ();
+
+  if (d != h)
+    gomp_fatal ("cannot map data on shared-memory system");
+}
+
+void
+acc_unmap_data (void *h)
+{
+}
+
+#define FLAG_PRESENT (1 << 0)
+#define FLAG_CREATE (1 << 1)
+#define FLAG_COPY (1 << 2)
+
+static void *
+present_create_copy (unsigned f, void *h, size_t s)
+{
+  if (!h || !s)
+    gomp_fatal ("[%p,+%d] is a bad range", (void *)h, (int)s);
+
+  goacc_lazy_initialize ();
+  return h;
+}
+
+void *
+acc_create (void *h, size_t s)
+{
+  return present_create_copy (FLAG_CREATE, h, s);
+}
+
+void *
+acc_copyin (void *h, size_t s)
+{
+  return present_create_copy (FLAG_CREATE | FLAG_COPY, h, s);
+}
+
+void *
+acc_present_or_create (void *h, size_t s)
+{
+  return present_create_copy (FLAG_PRESENT | FLAG_CREATE, h, s);
+}
+
+void *
+acc_present_or_copyin (void *h, size_t s)
+{
+  return present_create_copy (FLAG_PRESENT | FLAG_CREATE | FLAG_COPY, h, s);
+}
+
+#define FLAG_COPYOUT (1 << 0)
+
+static void
+delete_copyout (unsigned f, void *h, size_t s, const char *libfnname)
+{
+}
+
+void
+acc_delete (void *h , size_t s)
+{
+  delete_copyout (0, h, s, __FUNCTION__);
+}
+
+void
+acc_copyout (void *h, size_t s)
+{
+  delete_copyout (FLAG_COPYOUT, h, s, __FUNCTION__);
+}
+
+static void
+update_dev_host (int is_dev, void *h, size_t s)
+{
+  goacc_lazy_initialize ();
+}
+
+void
+acc_update_device (void *h, size_t s)
+{
+  update_dev_host (1, h, s);
+}
+
+void
+acc_update_self (void *h, size_t s)
+{
+  update_dev_host (0, h, s);
+}
+
+void
+gomp_acc_insert_pointer (size_t mapnum, void **hostaddrs, size_t *sizes,
+			 void *kinds)
+{
+}
+
+void
+gomp_acc_remove_pointer (void *h, bool force_copyfrom, int async, int mapnum)
+{
+}
--- libgomp/oacc-plugin.h.jj	2016-07-13 16:57:13.487423121 +0200
+++ libgomp/oacc-plugin.h	2016-07-13 16:57:13.487423121 +0200
@@ -0,0 +1,33 @@
+/* Copyright (C) 2014-2016 Free Software Foundation, Inc.
+
+   Contributed by Mentor Embedded.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef OACC_PLUGIN_H
+#define OACC_PLUGIN_H 1
+
+extern void GOMP_PLUGIN_async_unmap_vars (void *, int);
+extern void *GOMP_PLUGIN_acc_thread (void);
+
+#endif
--- libgomp/taskloop.c.jj	2016-07-13 16:57:18.935355570 +0200
+++ libgomp/taskloop.c	2016-07-13 16:57:18.935355570 +0200
@@ -0,0 +1,340 @@
+/* Copyright (C) 2015-2016 Free Software Foundation, Inc.
+   Contributed by Jakub Jelinek <jakub@redhat.com>.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* This file handles the taskloop construct.  It is included twice, once
+   for the long and once for unsigned long long variant.  */
+
+/* Called when encountering an explicit task directive.  If IF_CLAUSE is
+   false, then we must not delay in executing the task.  If UNTIED is true,
+   then the task may be executed by any member of the team.  */
+
+void
+GOMP_taskloop (void (*fn) (void *), void *data, void (*cpyfn) (void *, void *),
+	       long arg_size, long arg_align, unsigned flags,
+	       unsigned long num_tasks, int priority,
+	       TYPE start, TYPE end, TYPE step)
+{
+  struct gomp_thread *thr = gomp_thread ();
+  struct gomp_team *team = thr->ts.team;
+
+#ifdef HAVE_BROKEN_POSIX_SEMAPHORES
+  /* If pthread_mutex_* is used for omp_*lock*, then each task must be
+     tied to one thread all the time.  This means UNTIED tasks must be
+     tied and if CPYFN is non-NULL IF(0) must be forced, as CPYFN
+     might be running on different thread than FN.  */
+  if (cpyfn)
+    flags &= ~GOMP_TASK_FLAG_IF;
+  flags &= ~GOMP_TASK_FLAG_UNTIED;
+#endif
+
+  /* If parallel or taskgroup has been cancelled, don't start new tasks.  */
+  if (team && gomp_team_barrier_cancelled (&team->barrier))
+    return;
+
+#ifdef TYPE_is_long
+  TYPE s = step;
+  if (step > 0)
+    {
+      if (start >= end)
+	return;
+      s--;
+    }
+  else
+    {
+      if (start <= end)
+	return;
+      s++;
+    }
+  UTYPE n = (end - start + s) / step;
+#else
+  UTYPE n;
+  if (flags & GOMP_TASK_FLAG_UP)
+    {
+      if (start >= end)
+	return;
+      n = (end - start + step - 1) / step;
+    }
+  else
+    {
+      if (start <= end)
+	return;
+      n = (start - end - step - 1) / -step;
+    }
+#endif
+
+  TYPE task_step = step;
+  unsigned long nfirst = n;
+  if (flags & GOMP_TASK_FLAG_GRAINSIZE)
+    {
+      unsigned long grainsize = num_tasks;
+#ifdef TYPE_is_long
+      num_tasks = n / grainsize;
+#else
+      UTYPE ndiv = n / grainsize;
+      num_tasks = ndiv;
+      if (num_tasks != ndiv)
+	num_tasks = ~0UL;
+#endif
+      if (num_tasks <= 1)
+	{
+	  num_tasks = 1;
+	  task_step = end - start;
+	}
+      else if (num_tasks >= grainsize
+#ifndef TYPE_is_long
+	       && num_tasks != ~0UL
+#endif
+	      )
+	{
+	  UTYPE mul = num_tasks * grainsize;
+	  task_step = (TYPE) grainsize * step;
+	  if (mul != n)
+	    {
+	      task_step += step;
+	      nfirst = n - mul - 1;
+	    }
+	}
+      else
+	{
+	  UTYPE div = n / num_tasks;
+	  UTYPE mod = n % num_tasks;
+	  task_step = (TYPE) div * step;
+	  if (mod)
+	    {
+	      task_step += step;
+	      nfirst = mod - 1;
+	    }
+	}
+    }
+  else
+    {
+      if (num_tasks == 0)
+	num_tasks = team ? team->nthreads : 1;
+      if (num_tasks >= n)
+	num_tasks = n;
+      else
+	{
+	  UTYPE div = n / num_tasks;
+	  UTYPE mod = n % num_tasks;
+	  task_step = (TYPE) div * step;
+	  if (mod)
+	    {
+	      task_step += step;
+	      nfirst = mod - 1;
+	    }
+	}
+    }
+
+  if (flags & GOMP_TASK_FLAG_NOGROUP)
+    {
+      if (thr->task && thr->task->taskgroup && thr->task->taskgroup->cancelled)
+	return;
+    }
+  else
+    ialias_call (GOMP_taskgroup_start) ();
+
+  if (priority > gomp_max_task_priority_var)
+    priority = gomp_max_task_priority_var;
+
+  if ((flags & GOMP_TASK_FLAG_IF) == 0 || team == NULL
+      || (thr->task && thr->task->final_task)
+      || team->task_count + num_tasks > 64 * team->nthreads)
+    {
+      unsigned long i;
+      if (__builtin_expect (cpyfn != NULL, 0))
+	{
+	  struct gomp_task task[num_tasks];
+	  struct gomp_task *parent = thr->task;
+	  arg_size = (arg_size + arg_align - 1) & ~(arg_align - 1);
+	  char buf[num_tasks * arg_size + arg_align - 1];
+	  char *arg = (char *) (((uintptr_t) buf + arg_align - 1)
+				& ~(uintptr_t) (arg_align - 1));
+	  char *orig_arg = arg;
+	  for (i = 0; i < num_tasks; i++)
+	    {
+	      gomp_init_task (&task[i], parent, gomp_icv (false));
+	      task[i].priority = priority;
+	      task[i].kind = GOMP_TASK_UNDEFERRED;
+	      task[i].final_task = (thr->task && thr->task->final_task)
+				   || (flags & GOMP_TASK_FLAG_FINAL);
+	      if (thr->task)
+		{
+		  task[i].in_tied_task = thr->task->in_tied_task;
+		  task[i].taskgroup = thr->task->taskgroup;
+		}
+	      thr->task = &task[i];
+	      cpyfn (arg, data);
+	      arg += arg_size;
+	    }
+	  arg = orig_arg;
+	  for (i = 0; i < num_tasks; i++)
+	    {
+	      thr->task = &task[i];
+	      ((TYPE *)arg)[0] = start;
+	      start += task_step;
+	      ((TYPE *)arg)[1] = start;
+	      if (i == nfirst)
+		task_step -= step;
+	      fn (arg);
+	      arg += arg_size;
+	      if (!priority_queue_empty_p (&task[i].children_queue,
+					   MEMMODEL_RELAXED))
+		{
+		  gomp_mutex_lock (&team->task_lock);
+		  gomp_clear_parent (&task[i].children_queue);
+		  gomp_mutex_unlock (&team->task_lock);
+		}
+	      gomp_end_task ();
+	    }
+	}
+      else
+	for (i = 0; i < num_tasks; i++)
+	  {
+	    struct gomp_task task;
+
+	    gomp_init_task (&task, thr->task, gomp_icv (false));
+	    task.priority = priority;
+	    task.kind = GOMP_TASK_UNDEFERRED;
+	    task.final_task = (thr->task && thr->task->final_task)
+			      || (flags & GOMP_TASK_FLAG_FINAL);
+	    if (thr->task)
+	      {
+		task.in_tied_task = thr->task->in_tied_task;
+		task.taskgroup = thr->task->taskgroup;
+	      }
+	    thr->task = &task;
+	    ((TYPE *)data)[0] = start;
+	    start += task_step;
+	    ((TYPE *)data)[1] = start;
+	    if (i == nfirst)
+	      task_step -= step;
+	    fn (data);
+	    if (!priority_queue_empty_p (&task.children_queue,
+					 MEMMODEL_RELAXED))
+	      {
+		gomp_mutex_lock (&team->task_lock);
+		gomp_clear_parent (&task.children_queue);
+		gomp_mutex_unlock (&team->task_lock);
+	      }
+	    gomp_end_task ();
+	  }
+    }
+  else
+    {
+      struct gomp_task *tasks[num_tasks];
+      struct gomp_task *parent = thr->task;
+      struct gomp_taskgroup *taskgroup = parent->taskgroup;
+      char *arg;
+      int do_wake;
+      unsigned long i;
+
+      for (i = 0; i < num_tasks; i++)
+	{
+	  struct gomp_task *task
+	    = gomp_malloc (sizeof (*task) + arg_size + arg_align - 1);
+	  tasks[i] = task;
+	  arg = (char *) (((uintptr_t) (task + 1) + arg_align - 1)
+			  & ~(uintptr_t) (arg_align - 1));
+	  gomp_init_task (task, parent, gomp_icv (false));
+	  task->priority = priority;
+	  task->kind = GOMP_TASK_UNDEFERRED;
+	  task->in_tied_task = parent->in_tied_task;
+	  task->taskgroup = taskgroup;
+	  thr->task = task;
+	  if (cpyfn)
+	    {
+	      cpyfn (arg, data);
+	      task->copy_ctors_done = true;
+	    }
+	  else
+	    memcpy (arg, data, arg_size);
+	  ((TYPE *)arg)[0] = start;
+	  start += task_step;
+	  ((TYPE *)arg)[1] = start;
+	  if (i == nfirst)
+	    task_step -= step;
+	  thr->task = parent;
+	  task->kind = GOMP_TASK_WAITING;
+	  task->fn = fn;
+	  task->fn_data = arg;
+	  task->final_task = (flags & GOMP_TASK_FLAG_FINAL) >> 1;
+	}
+      gomp_mutex_lock (&team->task_lock);
+      /* If parallel or taskgroup has been cancelled, don't start new
+	 tasks.  */
+      if (__builtin_expect ((gomp_team_barrier_cancelled (&team->barrier)
+			     || (taskgroup && taskgroup->cancelled))
+			    && cpyfn == NULL, 0))
+	{
+	  gomp_mutex_unlock (&team->task_lock);
+	  for (i = 0; i < num_tasks; i++)
+	    {
+	      gomp_finish_task (tasks[i]);
+	      free (tasks[i]);
+	    }
+	  if ((flags & GOMP_TASK_FLAG_NOGROUP) == 0)
+	    ialias_call (GOMP_taskgroup_end) ();
+	  return;
+	}
+      if (taskgroup)
+	taskgroup->num_children += num_tasks;
+      for (i = 0; i < num_tasks; i++)
+	{
+	  struct gomp_task *task = tasks[i];
+	  priority_queue_insert (PQ_CHILDREN, &parent->children_queue,
+				 task, priority,
+				 PRIORITY_INSERT_BEGIN,
+				 /*last_parent_depends_on=*/false,
+				 task->parent_depends_on);
+	  if (taskgroup)
+	    priority_queue_insert (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
+				   task, priority, PRIORITY_INSERT_BEGIN,
+				   /*last_parent_depends_on=*/false,
+				   task->parent_depends_on);
+	  priority_queue_insert (PQ_TEAM, &team->task_queue, task, priority,
+				 PRIORITY_INSERT_END,
+				 /*last_parent_depends_on=*/false,
+				 task->parent_depends_on);
+	  ++team->task_count;
+	  ++team->task_queued_count;
+	}
+      gomp_team_barrier_set_task_pending (&team->barrier);
+      if (team->task_running_count + !parent->in_tied_task
+	  < team->nthreads)
+	{
+	  do_wake = team->nthreads - team->task_running_count
+		    - !parent->in_tied_task;
+	  if ((unsigned long) do_wake > num_tasks)
+	    do_wake = num_tasks;
+	}
+      else
+	do_wake = 0;
+      gomp_mutex_unlock (&team->task_lock);
+      if (do_wake)
+	gomp_team_barrier_wake (&team->barrier, do_wake);
+    }
+  if ((flags & GOMP_TASK_FLAG_NOGROUP) == 0)
+    ialias_call (GOMP_taskgroup_end) ();
+}
--- libgomp/priority_queue.h.jj	2016-07-13 16:57:04.438535323 +0200
+++ libgomp/priority_queue.h	2016-07-13 16:57:04.438535323 +0200
@@ -0,0 +1,485 @@
+/* Copyright (C) 2015-2016 Free Software Foundation, Inc.
+   Contributed by Aldy Hernandez <aldyh@redhat.com>.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* Header file for a priority queue of GOMP tasks.  */
+
+/* ?? Perhaps all the priority_tree_* functions are complex and rare
+   enough to go out-of-line and be moved to priority_queue.c.  ??  */
+
+#ifndef _PRIORITY_QUEUE_H_
+#define _PRIORITY_QUEUE_H_
+
+/* One task.  */
+
+struct priority_node
+{
+  /* Next and previous chains in a circular doubly linked list for
+     tasks within this task's priority.  */
+  struct priority_node *next, *prev;
+};
+
+/* All tasks within the same priority.  */
+
+struct priority_list
+{
+  /* Priority of the tasks in this set.  */
+  int priority;
+
+  /* Tasks.  */
+  struct priority_node *tasks;
+
+  /* This points to the last of the higher priority WAITING tasks.
+     Remember that for the children queue, we have:
+
+	parent_depends_on WAITING tasks.
+	!parent_depends_on WAITING tasks.
+	TIED tasks.
+
+     This is a pointer to the last of the parent_depends_on WAITING
+     tasks which are essentially, higher priority items within their
+     priority.  */
+  struct priority_node *last_parent_depends_on;
+};
+
+/* Another splay tree instantiation, for priority_list's.  */
+typedef struct prio_splay_tree_node_s *prio_splay_tree_node;
+typedef struct prio_splay_tree_s *prio_splay_tree;
+typedef struct prio_splay_tree_key_s *prio_splay_tree_key;
+struct prio_splay_tree_key_s {
+  /* This structure must only containing a priority_list, as we cast
+     prio_splay_tree_key to priority_list throughout.  */
+  struct priority_list l;
+};
+#define splay_tree_prefix prio
+#include "splay-tree.h"
+
+/* The entry point into a priority queue of tasks.
+
+   There are two alternate implementations with which to store tasks:
+   as a balanced tree of sorts, or as a simple list of tasks.  If
+   there are only priority-0 items (ROOT is NULL), we use the simple
+   list, otherwise (ROOT is non-NULL) we use the tree.  */
+
+struct priority_queue
+{
+  /* If t.root != NULL, this is a splay tree of priority_lists to hold
+     all tasks.  This is only used if multiple priorities are in play,
+     otherwise we use the priority_list `l' below to hold all
+     (priority-0) tasks.  */
+  struct prio_splay_tree_s t;
+
+  /* If T above is NULL, only priority-0 items exist, so keep them
+     in a simple list.  */
+  struct priority_list l;
+};
+
+enum priority_insert_type {
+  /* Insert at the beginning of a priority list.  */
+  PRIORITY_INSERT_BEGIN,
+  /* Insert at the end of a priority list.  */
+  PRIORITY_INSERT_END
+};
+
+/* Used to determine in which queue a given priority node belongs in.
+   See pnode field of gomp_task.  */
+
+enum priority_queue_type
+{
+  PQ_TEAM,	    /* Node belongs in gomp_team's task_queue.  */
+  PQ_CHILDREN,	    /* Node belongs in parent's children_queue.  */
+  PQ_TASKGROUP,	    /* Node belongs in taskgroup->taskgroup_queue.  */
+  PQ_IGNORED = 999
+};
+
+/* Priority queue implementation prototypes.  */
+
+extern bool priority_queue_task_in_queue_p (enum priority_queue_type,
+					    struct priority_queue *,
+					    struct gomp_task *);
+extern void priority_queue_dump (enum priority_queue_type,
+				 struct priority_queue *);
+extern void priority_queue_verify (enum priority_queue_type,
+				   struct priority_queue *, bool);
+extern void priority_tree_remove (enum priority_queue_type,
+				  struct priority_queue *,
+				  struct priority_node *);
+extern struct gomp_task *priority_tree_next_task (enum priority_queue_type,
+						  struct priority_queue *,
+						  enum priority_queue_type,
+						  struct priority_queue *,
+						  bool *);
+
+/* Return TRUE if there is more than one priority in HEAD.  This is
+   used throughout to to choose between the fast path (priority 0 only
+   items) and a world with multiple priorities.  */
+
+static inline bool
+priority_queue_multi_p (struct priority_queue *head)
+{
+  return __builtin_expect (head->t.root != NULL, 0);
+}
+
+/* Initialize a priority queue.  */
+
+static inline void
+priority_queue_init (struct priority_queue *head)
+{
+  head->t.root = NULL;
+  /* To save a few microseconds, we don't initialize head->l.priority
+     to 0 here.  It is implied that priority will be 0 if head->t.root
+     == NULL.
+
+     priority_tree_insert() will fix this when we encounter multiple
+     priorities.  */
+  head->l.tasks = NULL;
+  head->l.last_parent_depends_on = NULL;
+}
+
+static inline void
+priority_queue_free (struct priority_queue *head)
+{
+  /* There's nothing to do, as tasks were freed as they were removed
+     in priority_queue_remove.  */
+}
+
+/* Forward declarations.  */
+static inline size_t priority_queue_offset (enum priority_queue_type);
+static inline struct gomp_task *priority_node_to_task
+				(enum priority_queue_type,
+				 struct priority_node *);
+static inline struct priority_node *task_to_priority_node
+				    (enum priority_queue_type,
+				     struct gomp_task *);
+
+/* Return TRUE if priority queue HEAD is empty.
+
+   MODEL IS MEMMODEL_ACQUIRE if we should use an acquire atomic to
+   read from the root of the queue, otherwise MEMMODEL_RELAXED if we
+   should use a plain load.  */
+
+static inline _Bool
+priority_queue_empty_p (struct priority_queue *head, enum memmodel model)
+{
+  /* Note: The acquire barriers on the loads here synchronize with
+     the write of a NULL in gomp_task_run_post_remove_parent.  It is
+     not necessary that we synchronize with other non-NULL writes at
+     this point, but we must ensure that all writes to memory by a
+     child thread task work function are seen before we exit from
+     GOMP_taskwait.  */
+  if (priority_queue_multi_p (head))
+    {
+      if (model == MEMMODEL_ACQUIRE)
+	return __atomic_load_n (&head->t.root, MEMMODEL_ACQUIRE) == NULL;
+      return head->t.root == NULL;
+    }
+  if (model == MEMMODEL_ACQUIRE)
+    return __atomic_load_n (&head->l.tasks, MEMMODEL_ACQUIRE) == NULL;
+  return head->l.tasks == NULL;
+}
+
+/* Look for a given PRIORITY in HEAD.  Return it if found, otherwise
+   return NULL.  This only applies to the tree variant in HEAD.  There
+   is no point in searching for priorities in HEAD->L.  */
+
+static inline struct priority_list *
+priority_queue_lookup_priority (struct priority_queue *head, int priority)
+{
+  if (head->t.root == NULL)
+    return NULL;
+  struct prio_splay_tree_key_s k;
+  k.l.priority = priority;
+  return (struct priority_list *)
+    prio_splay_tree_lookup (&head->t, &k);
+}
+
+/* Insert task in DATA, with PRIORITY, in the priority list in LIST.
+   LIST contains items of type TYPE.
+
+   If POS is PRIORITY_INSERT_BEGIN, the new task is inserted at the
+   top of its respective priority.  If POS is PRIORITY_INSERT_END, the
+   task is inserted at the end of its priority.
+
+   If ADJUST_PARENT_DEPENDS_ON is TRUE, LIST is a children queue, and
+   we must keep track of higher and lower priority WAITING tasks by
+   keeping the queue's last_parent_depends_on field accurate.  This
+   only applies to the children queue, and the caller must ensure LIST
+   is a children queue in this case.
+
+   If ADJUST_PARENT_DEPENDS_ON is TRUE, TASK_IS_PARENT_DEPENDS_ON is
+   set to the task's parent_depends_on field.  If
+   ADJUST_PARENT_DEPENDS_ON is FALSE, this field is irrelevant.
+
+   Return the new priority_node.  */
+
+static inline void
+priority_list_insert (enum priority_queue_type type,
+		      struct priority_list *list,
+		      struct gomp_task *task,
+		      int priority,
+		      enum priority_insert_type pos,
+		      bool adjust_parent_depends_on,
+		      bool task_is_parent_depends_on)
+{
+  struct priority_node *node = task_to_priority_node (type, task);
+  if (list->tasks)
+    {
+      /* If we are keeping track of higher/lower priority items,
+	 but this is a lower priority WAITING task
+	 (parent_depends_on != NULL), put it after all ready to
+	 run tasks.  See the comment in
+	 priority_queue_upgrade_task for a visual on how tasks
+	 should be organized.  */
+      if (adjust_parent_depends_on
+	  && pos == PRIORITY_INSERT_BEGIN
+	  && list->last_parent_depends_on
+	  && !task_is_parent_depends_on)
+	{
+	  struct priority_node *last_parent_depends_on
+	    = list->last_parent_depends_on;
+	  node->next = last_parent_depends_on->next;
+	  node->prev = last_parent_depends_on;
+	}
+      /* Otherwise, put it at the top/bottom of the queue.  */
+      else
+	{
+	  node->next = list->tasks;
+	  node->prev = list->tasks->prev;
+	  if (pos == PRIORITY_INSERT_BEGIN)
+	    list->tasks = node;
+	}
+      node->next->prev = node;
+      node->prev->next = node;
+    }
+  else
+    {
+      node->next = node;
+      node->prev = node;
+      list->tasks = node;
+    }
+  if (adjust_parent_depends_on
+      && list->last_parent_depends_on == NULL
+      && task_is_parent_depends_on)
+    list->last_parent_depends_on = node;
+}
+
+/* Tree version of priority_list_insert.  */
+
+static inline void
+priority_tree_insert (enum priority_queue_type type,
+		      struct priority_queue *head,
+		      struct gomp_task *task,
+		      int priority,
+		      enum priority_insert_type pos,
+		      bool adjust_parent_depends_on,
+		      bool task_is_parent_depends_on)
+{
+  if (__builtin_expect (head->t.root == NULL, 0))
+    {
+      /* The first time around, transfer any priority 0 items to the
+	 tree.  */
+      if (head->l.tasks != NULL)
+	{
+	  prio_splay_tree_node k = gomp_malloc (sizeof (*k));
+	  k->left = NULL;
+	  k->right = NULL;
+	  k->key.l.priority = 0;
+	  k->key.l.tasks = head->l.tasks;
+	  k->key.l.last_parent_depends_on = head->l.last_parent_depends_on;
+	  prio_splay_tree_insert (&head->t, k);
+	  head->l.tasks = NULL;
+	}
+    }
+  struct priority_list *list
+    = priority_queue_lookup_priority (head, priority);
+  if (!list)
+    {
+      prio_splay_tree_node k = gomp_malloc (sizeof (*k));
+      k->left = NULL;
+      k->right = NULL;
+      k->key.l.priority = priority;
+      k->key.l.tasks = NULL;
+      k->key.l.last_parent_depends_on = NULL;
+      prio_splay_tree_insert (&head->t, k);
+      list = &k->key.l;
+    }
+  priority_list_insert (type, list, task, priority, pos,
+			adjust_parent_depends_on,
+			task_is_parent_depends_on);
+}
+
+/* Generic version of priority_*_insert.  */
+
+static inline void
+priority_queue_insert (enum priority_queue_type type,
+		       struct priority_queue *head,
+		       struct gomp_task *task,
+		       int priority,
+		       enum priority_insert_type pos,
+		       bool adjust_parent_depends_on,
+		       bool task_is_parent_depends_on)
+{
+#if _LIBGOMP_CHECKING_
+  if (priority_queue_task_in_queue_p (type, head, task))
+    gomp_fatal ("Attempt to insert existing task %p", task);
+#endif
+  if (priority_queue_multi_p (head) || __builtin_expect (priority > 0, 0))
+    priority_tree_insert (type, head, task, priority, pos,
+			  adjust_parent_depends_on,
+			  task_is_parent_depends_on);
+  else
+    priority_list_insert (type, &head->l, task, priority, pos,
+			  adjust_parent_depends_on,
+			  task_is_parent_depends_on);
+}
+
+/* If multiple priorities are in play, return the highest priority
+   task from within Q1 and Q2, while giving preference to tasks from
+   Q1.  If the returned task is chosen from Q1, *Q1_CHOSEN_P is set to
+   TRUE, otherwise it is set to FALSE.
+
+   If multiple priorities are not in play (only 0 priorities are
+   available), the next task is chosen exclusively from Q1.
+
+   As a special case, Q2 can be NULL, in which case, we just choose
+   the highest priority WAITING task in Q1.  This is an optimization
+   to speed up looking through only one queue.
+
+   We assume Q1 has at least one item.  */
+
+static inline struct gomp_task *
+priority_queue_next_task (enum priority_queue_type t1,
+			  struct priority_queue *q1,
+			  enum priority_queue_type t2,
+			  struct priority_queue *q2,
+			  bool *q1_chosen_p)
+{
+#if _LIBGOMP_CHECKING_
+  if (priority_queue_empty_p (q1, MEMMODEL_RELAXED))
+    gomp_fatal ("priority_queue_next_task: Q1 is empty");
+#endif
+  if (priority_queue_multi_p (q1))
+    {
+      struct gomp_task *t
+	= priority_tree_next_task (t1, q1, t2, q2, q1_chosen_p);
+      /* If T is NULL, there are no WAITING tasks in Q1.  In which
+	 case, return any old (non-waiting) task which will cause the
+	 caller to do the right thing when checking T->KIND ==
+	 GOMP_TASK_WAITING.  */
+      if (!t)
+	{
+#if _LIBGOMP_CHECKING_
+	  if (*q1_chosen_p == false)
+	    gomp_fatal ("priority_queue_next_task inconsistency");
+#endif
+	  return priority_node_to_task (t1, q1->t.root->key.l.tasks);
+	}
+      return t;
+    }
+  else
+    {
+      *q1_chosen_p = true;
+      return priority_node_to_task (t1, q1->l.tasks);
+    }
+}
+
+/* Remove NODE from LIST.
+
+   If we are removing the one and only item in the list, and MODEL is
+   MEMMODEL_RELEASE, use an atomic release to clear the list.
+
+   If the list becomes empty after the remove, return TRUE.  */
+
+static inline bool
+priority_list_remove (struct priority_list *list,
+		      struct priority_node *node,
+		      enum memmodel model)
+{
+  bool empty = false;
+  node->prev->next = node->next;
+  node->next->prev = node->prev;
+  if (list->tasks == node)
+    {
+      if (node->next != node)
+	list->tasks = node->next;
+      else
+	{
+	  /* We access task->children in GOMP_taskwait outside of
+	     the task lock mutex region, so need a release barrier
+	     here to ensure memory written by child_task->fn above
+	     is flushed before the NULL is written.  */
+	  if (model == MEMMODEL_RELEASE)
+	    __atomic_store_n (&list->tasks, NULL, MEMMODEL_RELEASE);
+	  else
+	    list->tasks = NULL;
+	  empty = true;
+	  goto remove_out;
+	}
+    }
+remove_out:
+#if _LIBGOMP_CHECKING_
+  memset (node, 0xaf, sizeof (*node));
+#endif
+  return empty;
+}
+
+/* This is the generic version of priority_list_remove.
+
+   Remove NODE from priority queue HEAD.  HEAD contains tasks of type TYPE.
+
+   If we are removing the one and only item in the priority queue and
+   MODEL is MEMMODEL_RELEASE, use an atomic release to clear the queue.
+
+   If the queue becomes empty after the remove, return TRUE.  */
+
+static inline bool
+priority_queue_remove (enum priority_queue_type type,
+		       struct priority_queue *head,
+		       struct gomp_task *task,
+		       enum memmodel model)
+{
+#if _LIBGOMP_CHECKING_
+  if (!priority_queue_task_in_queue_p (type, head, task))
+    gomp_fatal ("Attempt to remove missing task %p", task);
+#endif
+  if (priority_queue_multi_p (head))
+    {
+      priority_tree_remove (type, head, task_to_priority_node (type, task));
+      if (head->t.root == NULL)
+	{
+	  if (model == MEMMODEL_RELEASE)
+	    /* Errr, we store NULL twice, the alternative would be to
+	       use an atomic release directly in the splay tree
+	       routines.  Worth it?  */
+	    __atomic_store_n (&head->t.root, NULL, MEMMODEL_RELEASE);
+	  return true;
+	}
+      return false;
+    }
+  else
+    return priority_list_remove (&head->l,
+				 task_to_priority_node (type, task), model);
+}
+
+#endif /* _PRIORITY_QUEUE_H_ */
--- libgomp/priority_queue.c.jj	2016-07-13 16:57:04.435535360 +0200
+++ libgomp/priority_queue.c	2016-07-13 16:57:04.435535360 +0200
@@ -0,0 +1,300 @@
+/* Copyright (C) 2015-2016 Free Software Foundation, Inc.
+   Contributed by Aldy Hernandez <aldyh@redhat.com>.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* Priority queue implementation of GOMP tasks.  */
+
+#include "libgomp.h"
+
+#if _LIBGOMP_CHECKING_
+#include <stdio.h>
+
+/* Sanity check to verify whether a TASK is in LIST.  Return TRUE if
+   found, FALSE otherwise.
+
+   TYPE is the type of priority queue this task resides in.  */
+
+static inline bool
+priority_queue_task_in_list_p (enum priority_queue_type type,
+			       struct priority_list *list,
+			       struct gomp_task *task)
+{
+  struct priority_node *p = list->tasks;
+  do
+    {
+      if (priority_node_to_task (type, p) == task)
+	return true;
+      p = p->next;
+    }
+  while (p != list->tasks);
+  return false;
+}
+
+/* Tree version of priority_queue_task_in_list_p.  */
+
+static inline bool
+priority_queue_task_in_tree_p (enum priority_queue_type type,
+			       struct priority_queue *head,
+			       struct gomp_task *task)
+{
+  struct priority_list *list
+    = priority_queue_lookup_priority (head, task->priority);
+  if (!list)
+    return false;
+  return priority_queue_task_in_list_p (type, list, task);
+}
+
+/* Generic version of priority_queue_task_in_list_p that works for
+   trees or lists.  */
+
+bool
+priority_queue_task_in_queue_p (enum priority_queue_type type,
+				struct priority_queue *head,
+				struct gomp_task *task)
+{
+  if (priority_queue_empty_p (head, MEMMODEL_RELAXED))
+    return false;
+  if (priority_queue_multi_p (head))
+    return priority_queue_task_in_tree_p (type, head, task);
+  else
+    return priority_queue_task_in_list_p (type, &head->l, task);
+}
+
+/* Sanity check LIST to make sure the tasks therein are in the right
+   order.  LIST is a priority list of type TYPE.
+
+   The expected order is that GOMP_TASK_WAITING tasks come before
+   GOMP_TASK_TIED/GOMP_TASK_ASYNC_RUNNING ones.
+
+   If CHECK_DEPS is TRUE, we also check that parent_depends_on WAITING
+   tasks come before !parent_depends_on WAITING tasks.  This is only
+   applicable to the children queue, and the caller is expected to
+   ensure that we are verifying the children queue.  */
+
+static void
+priority_list_verify (enum priority_queue_type type,
+		      struct priority_list *list, bool check_deps)
+{
+  bool seen_tied = false;
+  bool seen_plain_waiting = false;
+  struct priority_node *p = list->tasks;
+  while (1)
+    {
+      struct gomp_task *t = priority_node_to_task (type, p);
+      if (seen_tied && t->kind == GOMP_TASK_WAITING)
+	gomp_fatal ("priority_queue_verify: WAITING task after TIED");
+      if (t->kind >= GOMP_TASK_TIED)
+	seen_tied = true;
+      else if (check_deps && t->kind == GOMP_TASK_WAITING)
+	{
+	  if (t->parent_depends_on)
+	    {
+	      if (seen_plain_waiting)
+		gomp_fatal ("priority_queue_verify: "
+			    "parent_depends_on after !parent_depends_on");
+	    }
+	  else
+	    seen_plain_waiting = true;
+	}
+      p = p->next;
+      if (p == list->tasks)
+	break;
+    }
+}
+
+/* Callback type for priority_tree_verify_callback.  */
+struct cbtype
+{
+  enum priority_queue_type type;
+  bool check_deps;
+};
+
+/* Verify every task in NODE.
+
+   Callback for splay_tree_foreach.  */
+
+static void
+priority_tree_verify_callback (prio_splay_tree_key key, void *data)
+{
+  struct cbtype *cb = (struct cbtype *) data;
+  priority_list_verify (cb->type, &key->l, cb->check_deps);
+}
+
+/* Generic version of priority_list_verify.
+
+   Sanity check HEAD to make sure the tasks therein are in the right
+   order.  The priority_queue holds tasks of type TYPE.
+
+   If CHECK_DEPS is TRUE, we also check that parent_depends_on WAITING
+   tasks come before !parent_depends_on WAITING tasks.  This is only
+   applicable to the children queue, and the caller is expected to
+   ensure that we are verifying the children queue.  */
+
+void
+priority_queue_verify (enum priority_queue_type type,
+		       struct priority_queue *head, bool check_deps)
+{
+  if (priority_queue_empty_p (head, MEMMODEL_RELAXED))
+    return;
+  if (priority_queue_multi_p (head))
+    {
+      struct cbtype cb = { type, check_deps };
+      prio_splay_tree_foreach (&head->t,
+			       priority_tree_verify_callback, &cb);
+    }
+  else
+    priority_list_verify (type, &head->l, check_deps);
+}
+#endif /* _LIBGOMP_CHECKING_ */
+
+/* Remove NODE from priority queue HEAD, wherever it may be inside the
+   tree.  HEAD contains tasks of type TYPE.  */
+
+void
+priority_tree_remove (enum priority_queue_type type,
+		      struct priority_queue *head,
+		      struct priority_node *node)
+{
+  /* ?? The only reason this function is not inlined is because we
+     need to find the priority within gomp_task (which has not been
+     completely defined in the header file).  If the lack of inlining
+     is a concern, we could pass the priority number as a
+     parameter, or we could move this to libgomp.h.  */
+  int priority = priority_node_to_task (type, node)->priority;
+
+  /* ?? We could avoid this lookup by keeping a pointer to the key in
+     the priority_node.  */
+  struct priority_list *list
+    = priority_queue_lookup_priority (head, priority);
+#if _LIBGOMP_CHECKING_
+  if (!list)
+    gomp_fatal ("Unable to find priority %d", priority);
+#endif
+  /* If NODE was the last in its priority, clean up the priority.  */
+  if (priority_list_remove (list, node, MEMMODEL_RELAXED))
+    {
+      prio_splay_tree_remove (&head->t, (prio_splay_tree_key) list);
+      list->tasks = NULL;
+#if _LIBGOMP_CHECKING_
+      memset (list, 0xaf, sizeof (*list));
+#endif
+      free (list);
+    }
+}
+
+/* Return the highest priority WAITING task in a splay tree NODE.  If
+   there are no WAITING tasks available, return NULL.
+
+   NODE is a priority list containing tasks of type TYPE.
+
+   The right most node in a tree contains the highest priority.
+   Recurse down to find such a node.  If the task at that max node is
+   not WAITING, bubble back up and look at the remaining tasks
+   in-order.  */
+
+static struct gomp_task *
+priority_tree_next_task_1 (enum priority_queue_type type,
+			   prio_splay_tree_node node)
+{
+ again:
+  if (!node)
+    return NULL;
+  struct gomp_task *ret = priority_tree_next_task_1 (type, node->right);
+  if (ret)
+    return ret;
+  ret = priority_node_to_task (type, node->key.l.tasks);
+  if (ret->kind == GOMP_TASK_WAITING)
+    return ret;
+  node = node->left;
+  goto again;
+}
+
+/* Return the highest priority WAITING task from within Q1 and Q2,
+   while giving preference to tasks from Q1.  Q1 is a queue containing
+   items of type TYPE1.  Q2 is a queue containing items of type TYPE2.
+
+   Since we are mostly interested in Q1, if there are no WAITING tasks
+   in Q1, we don't bother checking Q2, and just return NULL.
+
+   As a special case, Q2 can be NULL, in which case, we just choose
+   the highest priority WAITING task in Q1.  This is an optimization
+   to speed up looking through only one queue.
+
+   If the returned task is chosen from Q1, *Q1_CHOSEN_P is set to
+   TRUE, otherwise it is set to FALSE.  */
+
+struct gomp_task *
+priority_tree_next_task (enum priority_queue_type type1,
+			 struct priority_queue *q1,
+			 enum priority_queue_type type2,
+			 struct priority_queue *q2,
+			 bool *q1_chosen_p)
+{
+  struct gomp_task *t1 = priority_tree_next_task_1 (type1, q1->t.root);
+  if (!t1
+      /* Special optimization when only searching through one queue.  */
+      || !q2)
+    {
+      *q1_chosen_p = true;
+      return t1;
+    }
+  struct gomp_task *t2 = priority_tree_next_task_1 (type2, q2->t.root);
+  if (!t2 || t1->priority > t2->priority)
+    {
+      *q1_chosen_p = true;
+      return t1;
+    }
+  if (t2->priority > t1->priority)
+    {
+      *q1_chosen_p = false;
+      return t2;
+    }
+  /* If we get here, the priorities are the same, so we must look at
+     parent_depends_on to make our decision.  */
+#if _LIBGOMP_CHECKING_
+  if (t1 != t2)
+    gomp_fatal ("priority_tree_next_task: t1 != t2");
+#endif
+  if (t2->parent_depends_on && !t1->parent_depends_on)
+    {
+      *q1_chosen_p = false;
+      return t2;
+    }
+  *q1_chosen_p = true;
+  return t1;
+}
+
+/* Priority splay trees comparison function.  */
+static inline int
+prio_splay_compare (prio_splay_tree_key x, prio_splay_tree_key y)
+{
+  if (x->l.priority == y->l.priority)
+    return 0;
+  return x->l.priority < y->l.priority ? -1 : 1;
+}
+
+/* Define another splay tree instantiation, for priority_list's.  */
+#define splay_tree_prefix prio
+#define splay_tree_c
+#include "splay-tree.h"
--- libgomp/openacc.f90.jj	2016-07-13 16:57:04.434535373 +0200
+++ libgomp/openacc.f90	2016-07-14 19:01:54.901230875 +0200
@@ -0,0 +1,911 @@
+!  OpenACC Runtime Library Definitions.
+
+!  Copyright (C) 2014-2016 Free Software Foundation, Inc.
+
+!  Contributed by Tobias Burnus <burnus@net-b.de>
+!              and Mentor Embedded.
+
+!  This file is part of the GNU Offloading and Multi Processing Library
+!  (libgomp).
+
+!  Libgomp is free software; you can redistribute it and/or modify it
+!  under the terms of the GNU General Public License as published by
+!  the Free Software Foundation; either version 3, or (at your option)
+!  any later version.
+
+!  Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+!  WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+!  FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+!  more details.
+
+!  Under Section 7 of GPL version 3, you are granted additional
+!  permissions described in the GCC Runtime Library Exception, version
+!  3.1, as published by the Free Software Foundation.
+
+!  You should have received a copy of the GNU General Public License and
+!  a copy of the GCC Runtime Library Exception along with this program;
+!  see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+!  <http://www.gnu.org/licenses/>.
+
+module openacc_kinds
+  use iso_fortran_env, only: int32
+  implicit none
+
+  private :: int32
+  public :: acc_device_kind
+
+  integer, parameter :: acc_device_kind = int32
+
+  public :: acc_device_none, acc_device_default, acc_device_host
+  public :: acc_device_not_host, acc_device_nvidia
+
+  ! Keep in sync with include/gomp-constants.h.
+  integer (acc_device_kind), parameter :: acc_device_none = 0
+  integer (acc_device_kind), parameter :: acc_device_default = 1
+  integer (acc_device_kind), parameter :: acc_device_host = 2
+  ! integer (acc_device_kind), parameter :: acc_device_host_nonshm = 3 removed.
+  integer (acc_device_kind), parameter :: acc_device_not_host = 4
+  integer (acc_device_kind), parameter :: acc_device_nvidia = 5
+
+  public :: acc_handle_kind
+
+  integer, parameter :: acc_handle_kind = int32
+
+  public :: acc_async_noval, acc_async_sync
+
+  ! Keep in sync with include/gomp-constants.h.
+  integer (acc_handle_kind), parameter :: acc_async_noval = -1
+  integer (acc_handle_kind), parameter :: acc_async_sync = -2
+
+end module
+
+module openacc_internal
+  use openacc_kinds
+  implicit none
+
+  interface
+    function acc_get_num_devices_h (d)
+      import
+      integer acc_get_num_devices_h
+      integer (acc_device_kind) d
+    end function
+
+    subroutine acc_set_device_type_h (d)
+      import
+      integer (acc_device_kind) d
+    end subroutine
+
+    function acc_get_device_type_h ()
+      import
+      integer (acc_device_kind) acc_get_device_type_h
+    end function
+
+    subroutine acc_set_device_num_h (n, d)
+      import
+      integer n
+      integer (acc_device_kind) d
+    end subroutine
+
+    function acc_get_device_num_h (d)
+      import
+      integer acc_get_device_num_h
+      integer (acc_device_kind) d
+    end function
+
+    function acc_async_test_h (a)
+      logical acc_async_test_h
+      integer a
+    end function
+
+    function acc_async_test_all_h ()
+      logical acc_async_test_all_h
+    end function
+
+    subroutine acc_wait_h (a)
+      integer a
+    end subroutine
+
+    subroutine acc_wait_async_h (a1, a2)
+      integer a1, a2
+    end subroutine
+
+    subroutine acc_wait_all_h ()
+    end subroutine
+
+    subroutine acc_wait_all_async_h (a)
+      integer a
+    end subroutine
+
+    subroutine acc_init_h (d)
+      import
+      integer (acc_device_kind) d
+    end subroutine
+
+    subroutine acc_shutdown_h (d)
+      import
+      integer (acc_device_kind) d
+    end subroutine
+
+    function acc_on_device_h (d)
+      import
+      integer (acc_device_kind) d
+      logical acc_on_device_h
+    end function
+
+    subroutine acc_copyin_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_copyin_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_copyin_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    subroutine acc_present_or_copyin_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_present_or_copyin_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_present_or_copyin_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    subroutine acc_create_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_create_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_create_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    subroutine acc_present_or_create_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_present_or_create_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_present_or_create_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    subroutine acc_copyout_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_copyout_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_copyout_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    subroutine acc_delete_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_delete_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_delete_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    subroutine acc_update_device_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_update_device_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_update_device_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    subroutine acc_update_self_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end subroutine
+
+    subroutine acc_update_self_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end subroutine
+
+    subroutine acc_update_self_array_h (a)
+      type (*), dimension (..), contiguous :: a
+    end subroutine
+
+    function acc_is_present_32_h (a, len)
+      use iso_c_binding, only: c_int32_t
+      logical acc_is_present_32_h
+      type (*), dimension (*) :: a
+      integer (c_int32_t) len
+    end function
+
+    function acc_is_present_64_h (a, len)
+      use iso_c_binding, only: c_int64_t
+      logical acc_is_present_64_h
+      type (*), dimension (*) :: a
+      integer (c_int64_t) len
+    end function
+
+    function acc_is_present_array_h (a)
+      logical acc_is_present_array_h
+      type (*), dimension (..), contiguous :: a
+    end function
+  end interface
+
+  interface
+    function acc_get_num_devices_l (d) &
+        bind (C, name = "acc_get_num_devices")
+      use iso_c_binding, only: c_int
+      integer (c_int) :: acc_get_num_devices_l
+      integer (c_int), value :: d
+    end function
+
+    subroutine acc_set_device_type_l (d) &
+        bind (C, name = "acc_set_device_type")
+      use iso_c_binding, only: c_int
+      integer (c_int), value :: d
+    end subroutine
+
+    function acc_get_device_type_l () &
+        bind (C, name = "acc_get_device_type")
+      use iso_c_binding, only: c_int
+      integer (c_int) :: acc_get_device_type_l
+    end function
+
+    subroutine acc_set_device_num_l (n, d) &
+        bind (C, name = "acc_set_device_num")
+      use iso_c_binding, only: c_int
+      integer (c_int), value :: n, d
+    end subroutine
+
+    function acc_get_device_num_l (d) &
+        bind (C, name = "acc_get_device_num")
+      use iso_c_binding, only: c_int
+      integer (c_int) :: acc_get_device_num_l
+      integer (c_int), value :: d
+    end function
+
+    function acc_async_test_l (a) &
+        bind (C, name = "acc_async_test")
+      use iso_c_binding, only: c_int
+      integer (c_int) :: acc_async_test_l
+      integer (c_int), value :: a
+    end function
+
+    function acc_async_test_all_l () &
+        bind (C, name = "acc_async_test_all")
+      use iso_c_binding, only: c_int
+      integer (c_int) :: acc_async_test_all_l
+    end function
+
+    subroutine acc_wait_l (a) &
+        bind (C, name = "acc_wait")
+      use iso_c_binding, only: c_int
+      integer (c_int), value :: a
+    end subroutine
+
+    subroutine acc_wait_async_l (a1, a2) &
+        bind (C, name = "acc_wait_async")
+      use iso_c_binding, only: c_int
+      integer (c_int), value :: a1, a2
+    end subroutine
+
+    subroutine acc_wait_all_l () &
+        bind (C, name = "acc_wait_all")
+      use iso_c_binding, only: c_int
+    end subroutine
+
+    subroutine acc_wait_all_async_l (a) &
+        bind (C, name = "acc_wait_all_async")
+      use iso_c_binding, only: c_int
+      integer (c_int), value :: a
+    end subroutine
+
+    subroutine acc_init_l (d) &
+        bind (C, name = "acc_init")
+      use iso_c_binding, only: c_int
+      integer (c_int), value :: d
+    end subroutine
+
+    subroutine acc_shutdown_l (d) &
+        bind (C, name = "acc_shutdown")
+      use iso_c_binding, only: c_int
+      integer (c_int), value :: d
+    end subroutine
+
+    function acc_on_device_l (d) &
+        bind (C, name = "acc_on_device")
+      use iso_c_binding, only: c_int
+      integer (c_int) :: acc_on_device_l
+      integer (c_int), value :: d
+    end function
+
+    subroutine acc_copyin_l (a, len) &
+        bind (C, name = "acc_copyin")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    subroutine acc_present_or_copyin_l (a, len) &
+        bind (C, name = "acc_present_or_copyin")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    subroutine acc_create_l (a, len) &
+        bind (C, name = "acc_create")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    subroutine acc_present_or_create_l (a, len) &
+        bind (C, name = "acc_present_or_create")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    subroutine acc_copyout_l (a, len) &
+        bind (C, name = "acc_copyout")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    subroutine acc_delete_l (a, len) &
+        bind (C, name = "acc_delete")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    subroutine acc_update_device_l (a, len) &
+        bind (C, name = "acc_update_device")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    subroutine acc_update_self_l (a, len) &
+        bind (C, name = "acc_update_self")
+      use iso_c_binding, only: c_size_t
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end subroutine
+
+    function acc_is_present_l (a, len) &
+        bind (C, name = "acc_is_present")
+      use iso_c_binding, only: c_int32_t, c_size_t
+      integer (c_int32_t) :: acc_is_present_l
+      type (*), dimension (*) :: a
+      integer (c_size_t), value :: len
+    end function
+  end interface
+end module
+
+module openacc
+  use openacc_kinds
+  use openacc_internal
+  implicit none
+
+  public :: openacc_version
+
+  public :: acc_get_num_devices, acc_set_device_type, acc_get_device_type
+  public :: acc_set_device_num, acc_get_device_num, acc_async_test
+  public :: acc_async_test_all, acc_wait, acc_wait_async, acc_wait_all
+  public :: acc_wait_all_async, acc_init, acc_shutdown, acc_on_device
+  public :: acc_copyin, acc_present_or_copyin, acc_pcopyin, acc_create
+  public :: acc_present_or_create, acc_pcreate, acc_copyout, acc_delete
+  public :: acc_update_device, acc_update_self, acc_is_present
+
+  integer, parameter :: openacc_version = 201306
+
+  interface acc_get_num_devices
+    procedure :: acc_get_num_devices_h
+  end interface
+
+  interface acc_set_device_type
+    procedure :: acc_set_device_type_h
+  end interface
+
+  interface acc_get_device_type
+    procedure :: acc_get_device_type_h
+  end interface
+
+  interface acc_set_device_num
+    procedure :: acc_set_device_num_h
+  end interface
+
+  interface acc_get_device_num
+    procedure :: acc_get_device_num_h
+  end interface
+
+  interface acc_async_test
+    procedure :: acc_async_test_h
+  end interface
+
+  interface acc_async_test_all
+    procedure :: acc_async_test_all_h
+  end interface
+
+  interface acc_wait
+    procedure :: acc_wait_h
+  end interface
+
+  interface acc_wait_async
+    procedure :: acc_wait_async_h
+  end interface
+
+  interface acc_wait_all
+    procedure :: acc_wait_all_h
+  end interface
+
+  interface acc_wait_all_async
+    procedure :: acc_wait_all_async_h
+  end interface
+
+  interface acc_init
+    procedure :: acc_init_h
+  end interface
+
+  interface acc_shutdown
+    procedure :: acc_shutdown_h
+  end interface
+
+  interface acc_on_device
+    procedure :: acc_on_device_h
+  end interface
+
+  ! acc_malloc: Only available in C/C++
+  ! acc_free: Only available in C/C++
+
+  ! As vendor extension, the following code supports both 32bit and 64bit
+  ! arguments for "size"; the OpenACC standard only permits default-kind
+  ! integers, which are of kind 4 (i.e. 32 bits).
+  ! Additionally, the two-argument version also takes arrays as argument.
+  ! and the one argument version also scalars. Note that the code assumes
+  ! that the arrays are contiguous.
+
+  interface acc_copyin
+    procedure :: acc_copyin_32_h
+    procedure :: acc_copyin_64_h
+    procedure :: acc_copyin_array_h
+  end interface
+
+  interface acc_present_or_copyin
+    procedure :: acc_present_or_copyin_32_h
+    procedure :: acc_present_or_copyin_64_h
+    procedure :: acc_present_or_copyin_array_h
+  end interface
+
+  interface acc_pcopyin
+    procedure :: acc_present_or_copyin_32_h
+    procedure :: acc_present_or_copyin_64_h
+    procedure :: acc_present_or_copyin_array_h
+  end interface
+
+  interface acc_create
+    procedure :: acc_create_32_h
+    procedure :: acc_create_64_h
+    procedure :: acc_create_array_h
+  end interface
+
+  interface acc_present_or_create
+    procedure :: acc_present_or_create_32_h
+    procedure :: acc_present_or_create_64_h
+    procedure :: acc_present_or_create_array_h
+  end interface
+
+  interface acc_pcreate
+    procedure :: acc_present_or_create_32_h
+    procedure :: acc_present_or_create_64_h
+    procedure :: acc_present_or_create_array_h
+  end interface
+
+  interface acc_copyout
+    procedure :: acc_copyout_32_h
+    procedure :: acc_copyout_64_h
+    procedure :: acc_copyout_array_h
+  end interface
+
+  interface acc_delete
+    procedure :: acc_delete_32_h
+    procedure :: acc_delete_64_h
+    procedure :: acc_delete_array_h
+  end interface
+
+  interface acc_update_device
+    procedure :: acc_update_device_32_h
+    procedure :: acc_update_device_64_h
+    procedure :: acc_update_device_array_h
+  end interface
+
+  interface acc_update_self
+    procedure :: acc_update_self_32_h
+    procedure :: acc_update_self_64_h
+    procedure :: acc_update_self_array_h
+  end interface
+
+  ! acc_map_data: Only available in C/C++
+  ! acc_unmap_data: Only available in C/C++
+  ! acc_deviceptr: Only available in C/C++
+  ! acc_hostptr: Only available in C/C++
+
+  interface acc_is_present
+    procedure :: acc_is_present_32_h
+    procedure :: acc_is_present_64_h
+    procedure :: acc_is_present_array_h
+  end interface
+
+  ! acc_memcpy_to_device: Only available in C/C++
+  ! acc_memcpy_from_device: Only available in C/C++
+
+end module
+
+function acc_get_num_devices_h (d)
+  use openacc_internal, only: acc_get_num_devices_l
+  use openacc_kinds
+  integer acc_get_num_devices_h
+  integer (acc_device_kind) d
+  acc_get_num_devices_h = acc_get_num_devices_l (d)
+end function
+
+subroutine acc_set_device_type_h (d)
+  use openacc_internal, only: acc_set_device_type_l
+  use openacc_kinds
+  integer (acc_device_kind) d
+  call acc_set_device_type_l (d)
+end subroutine
+
+function acc_get_device_type_h ()
+  use openacc_internal, only: acc_get_device_type_l
+  use openacc_kinds
+  integer (acc_device_kind) acc_get_device_type_h
+  acc_get_device_type_h = acc_get_device_type_l ()
+end function
+
+subroutine acc_set_device_num_h (n, d)
+  use openacc_internal, only: acc_set_device_num_l
+  use openacc_kinds
+  integer n
+  integer (acc_device_kind) d
+  call acc_set_device_num_l (n, d)
+end subroutine
+
+function acc_get_device_num_h (d)
+  use openacc_internal, only: acc_get_device_num_l
+  use openacc_kinds
+  integer acc_get_device_num_h
+  integer (acc_device_kind) d
+  acc_get_device_num_h = acc_get_device_num_l (d)
+end function
+
+function acc_async_test_h (a)
+  use openacc_internal, only: acc_async_test_l
+  logical acc_async_test_h
+  integer a
+  if (acc_async_test_l (a) .eq. 1) then
+    acc_async_test_h = .TRUE.
+  else
+    acc_async_test_h = .FALSE.
+  end if
+end function
+
+function acc_async_test_all_h ()
+  use openacc_internal, only: acc_async_test_all_l
+  logical acc_async_test_all_h
+  if (acc_async_test_all_l () .eq. 1) then
+    acc_async_test_all_h = .TRUE.
+  else
+    acc_async_test_all_h = .FALSE.
+  end if
+end function
+
+subroutine acc_wait_h (a)
+  use openacc_internal, only: acc_wait_l
+  integer a
+  call acc_wait_l (a)
+end subroutine
+
+subroutine acc_wait_async_h (a1, a2)
+  use openacc_internal, only: acc_wait_async_l
+  integer a1, a2
+  call acc_wait_async_l (a1, a2)
+end subroutine
+
+subroutine acc_wait_all_h ()
+  use openacc_internal, only: acc_wait_all_l
+  call acc_wait_all_l ()
+end subroutine
+
+subroutine acc_wait_all_async_h (a)
+  use openacc_internal, only: acc_wait_all_async_l
+  integer a
+  call acc_wait_all_async_l (a)
+end subroutine
+
+subroutine acc_init_h (d)
+  use openacc_internal, only: acc_init_l
+  use openacc_kinds
+  integer (acc_device_kind) d
+  call acc_init_l (d)
+end subroutine
+
+subroutine acc_shutdown_h (d)
+  use openacc_internal, only: acc_shutdown_l
+  use openacc_kinds
+  integer (acc_device_kind) d
+  call acc_shutdown_l (d)
+end subroutine
+
+function acc_on_device_h (d)
+  use openacc_internal, only: acc_on_device_l
+  use openacc_kinds
+  integer (acc_device_kind) d
+  logical acc_on_device_h
+  if (acc_on_device_l (d) .eq. 1) then
+    acc_on_device_h = .TRUE.
+  else
+    acc_on_device_h = .FALSE.
+  end if
+end function
+
+subroutine acc_copyin_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_copyin_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_copyin_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_copyin_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_copyin_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_copyin_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_copyin_array_h (a)
+  use openacc_internal, only: acc_copyin_l
+  type (*), dimension (..), contiguous :: a
+  call acc_copyin_l (a, sizeof (a))
+end subroutine
+
+subroutine acc_present_or_copyin_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_present_or_copyin_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_present_or_copyin_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_present_or_copyin_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_present_or_copyin_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_present_or_copyin_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_present_or_copyin_array_h (a)
+  use openacc_internal, only: acc_present_or_copyin_l
+  type (*), dimension (..), contiguous :: a
+  call acc_present_or_copyin_l (a, sizeof (a))
+end subroutine
+
+subroutine acc_create_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_create_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_create_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_create_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_create_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_create_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_create_array_h (a)
+  use openacc_internal, only: acc_create_l
+  type (*), dimension (..), contiguous :: a
+  call acc_create_l (a, sizeof (a))
+end subroutine
+
+subroutine acc_present_or_create_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_present_or_create_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_present_or_create_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_present_or_create_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_present_or_create_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_present_or_create_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_present_or_create_array_h (a)
+  use openacc_internal, only: acc_present_or_create_l
+  type (*), dimension (..), contiguous :: a
+  call acc_present_or_create_l (a, sizeof (a))
+end subroutine
+
+subroutine acc_copyout_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_copyout_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_copyout_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_copyout_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_copyout_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_copyout_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_copyout_array_h (a)
+  use openacc_internal, only: acc_copyout_l
+  type (*), dimension (..), contiguous :: a
+  call acc_copyout_l (a, sizeof (a))
+end subroutine
+
+subroutine acc_delete_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_delete_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_delete_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_delete_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_delete_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_delete_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_delete_array_h (a)
+  use openacc_internal, only: acc_delete_l
+  type (*), dimension (..), contiguous :: a
+  call acc_delete_l (a, sizeof (a))
+end subroutine
+
+subroutine acc_update_device_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_update_device_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_update_device_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_update_device_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_update_device_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_update_device_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_update_device_array_h (a)
+  use openacc_internal, only: acc_update_device_l
+  type (*), dimension (..), contiguous :: a
+  call acc_update_device_l (a, sizeof (a))
+end subroutine
+
+subroutine acc_update_self_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_update_self_l
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  call acc_update_self_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_update_self_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_update_self_l
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  call acc_update_self_l (a, int (len, kind = c_size_t))
+end subroutine
+
+subroutine acc_update_self_array_h (a)
+  use openacc_internal, only: acc_update_self_l
+  type (*), dimension (..), contiguous :: a
+  call acc_update_self_l (a, sizeof (a))
+end subroutine
+
+function acc_is_present_32_h (a, len)
+  use iso_c_binding, only: c_int32_t, c_size_t
+  use openacc_internal, only: acc_is_present_l
+  logical acc_is_present_32_h
+  type (*), dimension (*) :: a
+  integer (c_int32_t) len
+  if (acc_is_present_l (a, int (len, kind = c_size_t)) .eq. 1) then
+    acc_is_present_32_h = .TRUE.
+  else
+    acc_is_present_32_h = .FALSE.
+  end if
+end function
+
+function acc_is_present_64_h (a, len)
+  use iso_c_binding, only: c_int64_t, c_size_t
+  use openacc_internal, only: acc_is_present_l
+  logical acc_is_present_64_h
+  type (*), dimension (*) :: a
+  integer (c_int64_t) len
+  if (acc_is_present_l (a, int (len, kind = c_size_t)) .eq. 1) then
+    acc_is_present_64_h = .TRUE.
+  else
+    acc_is_present_64_h = .FALSE.
+  end if
+end function
+
+function acc_is_present_array_h (a)
+  use openacc_internal, only: acc_is_present_l
+  logical acc_is_present_array_h
+  type (*), dimension (..), contiguous :: a
+  acc_is_present_array_h = acc_is_present_l (a, sizeof (a)) == 1
+end function