re PR target/80090 (Incorrect assembler - output_addr_const may generate visibility...
[gcc.git] / libgomp / loop_ull.c
index d5d1eb8cc52662441bedf0e8b635b14a685f69da..5cb3db5fa469242e54d8d375796a86f7a6af930c 100644 (file)
@@ -1,29 +1,27 @@
-/* Copyright (C) 2005, 2008 Free Software Foundation, Inc.
+/* Copyright (C) 2005-2017 Free Software Foundation, Inc.
    Contributed by Richard Henderson <rth@redhat.com>.
 
-   This file is part of the GNU OpenMP Library (libgomp).
+   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 Lesser General Public License as published by
-   the Free Software Foundation; either version 2.1 of the License, or
-   (at your option) any later version.
+   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 Lesser General Public License for
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
    more details.
 
-   You should have received a copy of the GNU Lesser General Public License
-   along with libgomp; see the file COPYING.LIB.  If not, write to the
-   Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
-   MA 02110-1301, USA.  */
+   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.
 
-/* As a special exception, if you link this library with other files, some
-   of which are compiled with GCC, to produce an executable, this library
-   does not by itself cause the resulting executable to be covered by the
-   GNU General Public License.  This exception does not however invalidate
-   any other reasons why the executable file might be covered by the GNU
-   General Public License.  */
+   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 LOOP (FOR/DO) construct.  */
 
@@ -177,15 +175,15 @@ GOMP_loop_ull_runtime_start (bool up, gomp_ull start, gomp_ull end,
     {
     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
@@ -281,15 +279,15 @@ GOMP_loop_ull_ordered_runtime_start (bool up, gomp_ull start, gomp_ull end,
     {
     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
@@ -301,6 +299,114 @@ GOMP_loop_ull_ordered_runtime_start (bool up, gomp_ull start, gomp_ull end,
     }
 }
 
+/* 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
@@ -460,6 +566,10 @@ extern __typeof(gomp_loop_ull_dynamic_start) GOMP_loop_ull_dynamic_start
        __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")));
@@ -468,12 +578,23 @@ extern __typeof(gomp_loop_ull_ordered_dynamic_start) GOMP_loop_ull_ordered_dynam
 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")));
@@ -509,6 +630,25 @@ GOMP_loop_ull_guided_start (bool up, gomp_ull start, gomp_ull end,
                                     iend);
 }
 
+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,
@@ -536,6 +676,33 @@ GOMP_loop_ull_ordered_guided_start (bool up, gomp_ull start, gomp_ull end,
                                             istart, iend);
 }
 
+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)
 {
@@ -554,6 +721,18 @@ GOMP_loop_ull_guided_next (gomp_ull *istart, gomp_ull *iend)
   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)
 {