Patchwork [gomp4] #pragma omp distribute expansion

login
register
mail settings
Submitter Jakub Jelinek
Date May 28, 2013, 1:46 p.m.
Message ID <20130528134612.GO1377@tucnak.redhat.com>
Download mbox | patch
Permalink /patch/246866/
State New
Headers show

Comments

Jakub Jelinek - May 28, 2013, 1:46 p.m.
Hi!

This is on top of the earlier patch from today.
This handles expansion of #pragma omp distribute, so far without collapse
clause (for collapse > 1 we'll need expand_omp_for_static_*chunk to handle
collapsed loops, because we can't easily use library function for that,
that implicitly looks at thread num / num threads, while for distribute
we want to look at teams).
The nesting restriction for distribute was wrong, while teams construct
must be inside of target region (i.e. not orphaned), distribute construct
must be closely nested inside of teams region (i.e. can be orphaned).

2013-05-28  Jakub Jelinek  <jakub@redhat.com>

	* omp-builtins.def (BUILT_IN_OMP_GET_TEAM_NUM,
	BUILT_IN_OMP_GET_NUM_TEAMS): New built-ins.
	* omp-low.c (extract_omp_for_data, expand_omp_for_static_nochunk,
	expand_omp_for_static_chunk): Handle #pragma omp distribute.
	(expand_omp_for): Add assertion for non-finished distribute collapse
	> 1 support.
	(check_omp_nesting_restrictions): Allow orphaned distribute construct.


	Jakub

Patch

--- gcc/omp-builtins.def.jj	2013-04-10 19:11:23.000000000 +0200
+++ gcc/omp-builtins.def	2013-05-28 14:46:11.077263307 +0200
@@ -28,6 +28,10 @@  DEF_GOMP_BUILTIN (BUILT_IN_OMP_GET_THREA
 		  BT_FN_INT, ATTR_CONST_NOTHROW_LEAF_LIST)
 DEF_GOMP_BUILTIN (BUILT_IN_OMP_GET_NUM_THREADS, "omp_get_num_threads",
 		  BT_FN_INT, ATTR_CONST_NOTHROW_LEAF_LIST)
+DEF_GOMP_BUILTIN (BUILT_IN_OMP_GET_TEAM_NUM, "omp_get_team_num",
+		  BT_FN_INT, ATTR_CONST_NOTHROW_LEAF_LIST)
+DEF_GOMP_BUILTIN (BUILT_IN_OMP_GET_NUM_TEAMS, "omp_get_num_teams",
+		  BT_FN_INT, ATTR_CONST_NOTHROW_LEAF_LIST)
 
 DEF_GOMP_BUILTIN (BUILT_IN_GOMP_ATOMIC_START, "GOMP_atomic_start",
 		  BT_FN_VOID, ATTR_NOTHROW_LEAF_LIST)
--- gcc/omp-low.c.jj	2013-05-27 19:14:16.000000000 +0200
+++ gcc/omp-low.c	2013-05-28 15:35:31.660017839 +0200
@@ -224,6 +224,8 @@  extract_omp_for_data (gimple for_stmt, s
   struct omp_for_data_loop dummy_loop;
   location_t loc = gimple_location (for_stmt);
   bool non_ws = gimple_omp_for_kind (for_stmt) == GF_OMP_FOR_KIND_SIMD;
+  bool distribute = gimple_omp_for_kind (for_stmt)
+		    == GF_OMP_FOR_KIND_DISTRIBUTE;
 
   fd->for_stmt = for_stmt;
   fd->pre = NULL;
@@ -233,7 +235,8 @@  extract_omp_for_data (gimple for_stmt, s
   else
     fd->loops = &fd->loop;
 
-  fd->have_nowait = fd->have_ordered = false;
+  fd->have_nowait = distribute;
+  fd->have_ordered = false;
   fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC;
   fd->chunk_size = NULL_TREE;
   collapse_iter = NULL;
@@ -249,9 +252,14 @@  extract_omp_for_data (gimple for_stmt, s
 	fd->have_ordered = true;
 	break;
       case OMP_CLAUSE_SCHEDULE:
+	gcc_assert (!distribute);
 	fd->sched_kind = OMP_CLAUSE_SCHEDULE_KIND (t);
 	fd->chunk_size = OMP_CLAUSE_SCHEDULE_CHUNK_EXPR (t);
 	break;
+      case OMP_CLAUSE_DIST_SCHEDULE:
+	gcc_assert (distribute);
+	fd->chunk_size = OMP_CLAUSE_DIST_SCHEDULE_CHUNK_EXPR (t);
+	break;
       case OMP_CLAUSE_COLLAPSE:
 	if (fd->collapse > 1)
 	  {
@@ -1903,7 +1911,7 @@  check_omp_nesting_restrictions (gimple s
 	return true;
       if (gimple_omp_for_kind (stmt) == GF_OMP_FOR_KIND_DISTRIBUTE)
 	{
-	  if (ctx == NULL || gimple_code (ctx->stmt) != GIMPLE_OMP_TEAMS)
+	  if (ctx != NULL && gimple_code (ctx->stmt) != GIMPLE_OMP_TEAMS)
 	    {
 	      error_at (gimple_location (stmt),
 			"distribute construct must be closely nested inside "
@@ -4527,6 +4535,8 @@  expand_omp_for_static_nochunk (struct om
   gimple_stmt_iterator gsi;
   gimple stmt;
   edge ep;
+  enum built_in_function get_num_threads = BUILT_IN_OMP_GET_NUM_THREADS;
+  enum built_in_function get_thread_num = BUILT_IN_OMP_GET_THREAD_NUM;
 
   itype = type = TREE_TYPE (fd->loop.v);
   if (POINTER_TYPE_P (type))
@@ -4547,6 +4557,12 @@  expand_omp_for_static_nochunk (struct om
   gsi = gsi_last_bb (entry_bb);
   gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR);
 
+  if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE)
+    {
+      get_num_threads = BUILT_IN_OMP_GET_NUM_TEAMS;
+      get_thread_num = BUILT_IN_OMP_GET_TEAM_NUM;
+    }
+
   t = fold_binary (fd->loop.cond_code, boolean_type_node,
 		   fold_convert (type, fd->loop.n1),
 		   fold_convert (type, fd->loop.n2));
@@ -4591,12 +4607,12 @@  expand_omp_for_static_nochunk (struct om
       gsi = gsi_last_bb (entry_bb);
     }
 
-  t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS), 0);
+  t = build_call_expr (builtin_decl_explicit (get_num_threads), 0);
   t = fold_convert (itype, t);
   nthreads = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
 				       true, GSI_SAME_STMT);
 
-  t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM), 0);
+  t = build_call_expr (builtin_decl_explicit (get_thread_num), 0);
   t = fold_convert (itype, t);
   threadid = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
 				       true, GSI_SAME_STMT);
@@ -4798,6 +4814,8 @@  expand_omp_for_static_chunk (struct omp_
   gimple_stmt_iterator si;
   gimple stmt;
   edge se;
+  enum built_in_function get_num_threads = BUILT_IN_OMP_GET_NUM_THREADS;
+  enum built_in_function get_thread_num = BUILT_IN_OMP_GET_THREAD_NUM;
 
   itype = type = TREE_TYPE (fd->loop.v);
   if (POINTER_TYPE_P (type))
@@ -4823,6 +4841,12 @@  expand_omp_for_static_chunk (struct omp_
   si = gsi_last_bb (entry_bb);
   gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_FOR);
 
+  if (gimple_omp_for_kind (fd->for_stmt) == GF_OMP_FOR_KIND_DISTRIBUTE)
+    {
+      get_num_threads = BUILT_IN_OMP_GET_NUM_TEAMS;
+      get_thread_num = BUILT_IN_OMP_GET_TEAM_NUM;
+    }
+
   t = fold_binary (fd->loop.cond_code, boolean_type_node,
 		   fold_convert (type, fd->loop.n1),
 		   fold_convert (type, fd->loop.n2));
@@ -4867,12 +4891,12 @@  expand_omp_for_static_chunk (struct omp_
       si = gsi_last_bb (entry_bb);
     }
 
-  t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS), 0);
+  t = build_call_expr (builtin_decl_explicit (get_num_threads), 0);
   t = fold_convert (itype, t);
   nthreads = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
 				       true, GSI_SAME_STMT);
 
-  t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM), 0);
+  t = build_call_expr (builtin_decl_explicit (get_thread_num), 0);
   t = fold_convert (itype, t);
   threadid = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
 				       true, GSI_SAME_STMT);
@@ -5488,6 +5512,10 @@  expand_omp_for (struct omp_region *regio
     {
       int fn_index, start_ix, next_ix;
 
+      /* FIXME: expand_omp_for_static_*chunk needs to handle
+	 collapse > 1 for distribute.  */
+      gcc_assert (gimple_omp_for_kind (fd.for_stmt)
+		  != GF_OMP_FOR_KIND_DISTRIBUTE);
       if (fd.chunk_size == NULL
 	  && fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC)
 	fd.chunk_size = integer_zero_node;