gimplify.c (gimplify_scan_omp_clauses): No special handling for OMP_CLAUSE_TILE.

2017-02-09  Nathan Sidwell  <nathan@codesourcery.com>
	    Cesar Philippidis  <cesar@codesourcery.com>
	    Joseph Myers  <joseph@codesourcery.com>
	    Chung-Lin Tang  <cltang@codesourcery.com>

	gcc/
	* gimplify.c (gimplify_scan_omp_clauses): No special handling for
	OMP_CLAUSE_TILE.
	(gimplify_adjust_omp_clauses): Don't delete TILE.
	(gimplify_omp_for): Deal with TILE.
	* internal-fn.c (expand_GOACC_TILE): New function.
	* internal-fn.def (GOACC_DIM_POS): Comment may be overly conservative.
	(GOACC_TILE): New.
	* omp-expand.c (struct oacc_collapse): Add tile and outer fields.
	(expand_oacc_collapse_init): Add LOC paramter.  Initialize tile
	element fields.
	(expand_oacc_collapse_vars): Add INNER parm, adjust for tiling,
	avoid DIV for outermost collapse var.
	(expand_oacc_for): Insert tile element loop as needed.  Adjust.
	Remove out of date comments, fix whitespace.
	* omp-general.c (omp_extract_for_data): Deal with tiling.
	* omp-general.h (enum oacc_loop_flags): Add OLF_TILE flag,
	adjust OLF_DIM_BASE value.
	(struct omp_for_data): Add tiling field.
	* omp-low.c (scan_sharing_clauses): Allow OMP_CLAUSE_TILE.
	(lower_oacc_head_mark): Add OLF_TILE as appropriate.  Ensure 2 levels
	for auto loops.  Remove default auto determining, moved to
	oacc_loop_fixed_partitions.
	* omp-offload.c (struct oacc_loop): Change 'ifns' to vector of call
	stmts, add e_mask field.
	(oacc_dim_call): New function, abstracted out from oacc_thread_numbers.
	(oacc_thread_numbers): Use oacc_dim_call.
	(oacc_xform_tile): New.
	(new_oacc_loop_raw): Initialize e_mask, adjust for ifns vector.
	(finish_oacc_loop): Adjust for ifns vector.
	(oacc_loop_discover_walk): Append loop abstraction sites to list,
	add case for GOACC_TILE fns.
	(oacc_loop_xform_loop): Delete.
	(oacc_loop_process): Iterate over call list directly, and add
	handling for GOACC_TILE fns.
	(oacc_loop_fixed_partitions): Determine default auto, deal with TILE,
	dump partitioning.
	(oacc_loop_auto_partitions): Add outer_assign parm. Assign all but
	vector partitioning to outer loops.  Assign 2 partitions to loops
	when available. Add TILE handling.
	(oacc_loop_partition): Adjust oacc_loop_auto_partitions call.
	(execite_oacc_device_lower): Process GOACC_TILE fns, ignore unknown specs.
	* tree-nested.c (convert_nonlocal_omp_clauses): Allow OMP_CLAUSE_TILE.
	* tree.c (omp_clause_num_ops): Adjust TILE ops.
	* tree.h (OMP_CLAUSE_TILE_ITERVAR, OMP_CLAUSE_TILE_COUNT): New.

	gcc/c/
	* c-parser.c (c_parser_omp_clause_collapse): Disallow tile.
	(c_parser_oacc_clause_tile): Disallow collapse. Fix parsing and
	semantic checking.
	* c-parser.c (c_parser_omp_for_loop): Accept tiling constructs.

	gcc/cp/
	* parser.c (cp_parser_oacc_clause_tile): Disallow collapse.  Fix
	parsing.  Parse constant expression. Remove semantic checking.
	(cp_parser_omp_clause_collapse): Disallow tile.
	(cp_parser_omp_for_loop): Deal with tile clause.  Don't emit a parse
	error about missing for after already emitting one.  Use more
	conventional for idiom for unbounded loop.
	* pt.c (tsubst_omp_clauses): Handle OMP_CLAUSE_TILE.
	* semantics.c (finish_omp_clauses): Correct TILE semantic check.
	(finish_omp_for): Deal with tile clause.

	gcc/fortran/
	* openmp.c (resolve_omp_clauses): Error on directives
	containing both tile and collapse clauses.
	(resolve_oacc_loop_blocks): Represent '*' tile arguments as zero.
	* trans-openmp.c (gfc_trans_omp_do): Lower tiled loops like
	collapsed loops.

	gcc/testsuite/
	* c-c++-common/goacc/combined-directives.c: Remove xfail.
	* c-c++-common/goacc/loop-auto-1.c: Adjust and add additional case.
	* c-c++-common/goacc/loop-auto-2.c: New.
	* c-c++-common/goacc/tile.c: Include stdbool, fix expected errors.
	* c-c++-common/goacc/tile-2.c: New.
	* g++.dg/goacc/template.C: Test tile subst.  Adjust erroneous uses.
	* g++.dg/goacc/tile-1.C: New, check tile subst.
	* gcc.dg/goacc/loop-processing-1.c: Adjust dg-final pattern.
	* gfortran.dg/goacc/combined-directives.f90: Remove xfail.
	* gfortran.dg/goacc/tile-1.f90: New test.
	* gfortran.dg/goacc/tile-2.f90: New test.
	* gfortran.dg/goacc/tile-lowering.f95: New test.

	libgomp/
	* testsuite/libgomp.oacc-c-c++-common/tile-1.c: New.
	* testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c: Adjust and
	add additional case.
	* testsuite/libgomp.oacc-c-c++-common/vprop.c: XFAIL under
	"openacc_nvidia_accel_selected".
	* libgomp.oacc-fortran/nested-function-1.f90 (test2):
	Add num_workers(8) clause.

From-SVN: r245300

1 files changed, 19 insertions, 9 deletions

diff --git a/gcc/omp-low.c b/gcc/omp-low.c
index ff0f447..35df02c 100644
--- a/gcc/omp-low.c
+++ b/gcc/omp-low.c
@@ -1330,6 +1330,7 @@ scan_sharing_clauses (tree clauses, omp_context *ctx,
 	case OMP_CLAUSE_INDEPENDENT:
 	case OMP_CLAUSE_AUTO:
 	case OMP_CLAUSE_SEQ:
+	case OMP_CLAUSE_TILE:
 	case OMP_CLAUSE__SIMT_:
 	  break;
 
@@ -1340,7 +1341,6 @@ scan_sharing_clauses (tree clauses, omp_context *ctx,
 	    install_var_local (decl, ctx);
 	  break;
 
-	case OMP_CLAUSE_TILE:
 	case OMP_CLAUSE__CACHE_:
 	default:
 	  gcc_unreachable ();
@@ -1501,11 +1501,11 @@ scan_sharing_clauses (tree clauses, omp_context *ctx,
 	case OMP_CLAUSE_INDEPENDENT:
 	case OMP_CLAUSE_AUTO:
 	case OMP_CLAUSE_SEQ:
+	case OMP_CLAUSE_TILE:
 	case OMP_CLAUSE__GRIDDIM_:
 	case OMP_CLAUSE__SIMT_:
 	  break;
 
-	case OMP_CLAUSE_TILE:
 	case OMP_CLAUSE__CACHE_:
 	default:
 	  gcc_unreachable ();
@@ -5610,6 +5610,10 @@ lower_oacc_head_mark (location_t loc, tree ddvar, tree clauses,
 	  tag |= OLF_INDEPENDENT;
 	  break;
 
+	case OMP_CLAUSE_TILE:
+	  tag |= OLF_TILE;
+	  break;
+
 	default:
 	  continue;
 	}
@@ -5627,14 +5631,20 @@ lower_oacc_head_mark (location_t loc, tree ddvar, tree clauses,
   if (!tgt || is_oacc_parallel (tgt))
     tag |= OLF_INDEPENDENT;
 
-  /* A loop lacking SEQ, GANG, WORKER and/or VECTOR is implicitly AUTO.  */
-  if (!(tag & (((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1) << OLF_DIM_BASE)
-	       | OLF_SEQ)))
-      tag |= OLF_AUTO;
+  if (tag & OLF_TILE)
+    /* Tiling could use all 3 levels.  */ 
+    levels = 3;
+  else
+    {
+      /* A loop lacking SEQ, GANG, WORKER and/or VECTOR could be AUTO.
+	 Ensure at least one level, or 2 for possible auto
+	 partitioning */
+      bool maybe_auto = !(tag & (((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1)
+				  << OLF_DIM_BASE) | OLF_SEQ));
 
-  /* Ensure at least one level.  */
-  if (!levels)
-    levels++;
+      if (levels < 1u + maybe_auto)
+	levels = 1u + maybe_auto;
+    }
 
   args.quick_push (build_int_cst (integer_type_node, levels));
   args.quick_push (build_int_cst (integer_type_node, tag));