OpenACC 2.7: Implement reductions for arrays and records

This patch is a merge of:
https://gcc.gnu.org/pipermail/gcc-patches/2025-February/675222.html

This patch implements reductions for arrays, array sections, and
struct/record types as according to the OpenACC 2.7 specification.

2025-02-23  Chung-Lin Tang  <cltang@baylibre.com>

gcc/c/ChangeLog:

	* c-parser.cc (c_parser_omp_variable_list): Adjust parsing of opening
	square bracket.
	(c_parser_omp_clause_reduction): Adjustments for
	OpenACC-specific cases.
	* c-typeck.cc (c_oacc_reduction_defined_type_p): New function.
	(c_oacc_reduction_code_name): Likewise.
	(c_finish_omp_clauses): Handle OpenACC cases using new functions.

gcc/cp/ChangeLog:

	* parser.cc (cp_parser_omp_var_list_no_open): Adjust parsing of opening
	square bracket.
	(cp_parser_omp_clause_reduction): Adjustments for
	OpenACC-specific cases.
	* semantics.cc (cp_oacc_reduction_defined_type_p): New function.
	(cp_oacc_reduction_code_name): Likewise.
	(finish_omp_reduction_clause): Handle OpenACC cases using new
	functions.

gcc/fortran/ChangeLog:

	* openmp.cc (oacc_reduction_defined_type_p): New function.
	(resolve_omp_clauses): Adjust OpenACC array reduction error case.
	Adjust OMP_LIST_REDUCTION case. Use oacc_reduction_defined_type_p for
	OpenACC.
	* trans-openmp.cc (gfc_trans_omp_array_reduction_or_udr):
	Add 'stmtblock_t *block', and 'bool openacc' parameters. Add array and
	array section handling for openacc case. Adjust part of function to be
	!openacc only.
	(gfc_trans_omp_reduction_list):
	Add 'stmtblock_t *block', and 'bool openacc' parameters, pass to calls
	to gfc_trans_omp_array_reduction_or_udr.
	(gfc_trans_omp_array_section): Adjust setting of OMP_CLAUSE_SIZE to only
	OMP_CLAUSE_MAP clauses. Adjust calculations of decls and bias to use
	temporary variables instead of tree expression inside clauses.
	(gfc_trans_omp_clauses): Add 'block' and 'openacc' arguments to calls to
	gfc_trans_omp_reduction_list.
	(gfc_trans_omp_do): Pass 'op == EXEC_OACC_LOOP' as 'bool openacc'
	parameter in call to gfc_trans_omp_clauses.

gcc/ChangeLog:

	* config/gcn/gcn-tree.cc (#include "omp-offload.h"): Add include.
	(#include "memmodel.h"): Add include.
	(gcn_array_reduction_buffers): New vec<tree>
	for holding DECLs for reduction buffer pointer variables.
	(gcn_lockfull_update): Add pointer type fold_converts.
	(gcn_reduction_update): Additions for handling ARRAY_TYPE, pointer to
	ARRAY_TYPE, and RECORD_TYPE reductions.
	(gcn_goacc_get_worker_red_decl): Adjust parameters to handle
	non-constant offset case.
	(gcn_goacc_get_worker_array_reduction_buffer): New function.
	(gcn_create_if_else_seq): New function.
	(gcn_create_do_while_loop_seq): New function.
	(gcn_goacc_reduction_setup): Adjustments to handle arrays and records.
	(gcn_goacc_reduction_init): Likewise.
	(gcn_goacc_reduction_fini): Likewise.
	(gcn_goacc_reduction_teardown): Likewise.
	* config/nvptx/nvptx.cc (nvptx_gen_shuffle): Properly generate
	V2SI shuffle using vec_extract op.
	(nvptx_expand_shared_addr): Adjustments to handle non-constant size.
	(nvptx_get_shared_red_addr): Adjust type/alignment calculations to
	use TYPE_SIZE/ALIGN_UNIT instead of machine mode based.
	(nvptx_get_shared_red_addr): New function with array_max_idx parameter.
	(nvptx_reduction_update): Additions for handling ARRAY_TYPE, pointer to
	ARRA_TYPE, and RECORD_TYPE reductions.
	(nvptx_goacc_reduction_setup): Likewise.
	(nvptx_goacc_reduction_init): Likewise.
	(nvptx_goacc_reduction_fini): Likewise.
	(nvptx_goacc_reduction_teardown): Likewise.
	* gimplify.cc (gimplify_scan_omp_clauses): Gimplify inside COMPONENT_REF
	and convert codes for OMP_CLAUSE_REDUCTION cases. Add DECL_P check for
	do_add/do_add_decl goto case.
	(gimplify_adjust_omp_clauses): Avoid GOMP_MAP_POINTER OMP_CLAUSE_SIZE
	handling for OpenACC kernels. Call omp_add_variable for ARRAY_REF index.
	Peel away array MEM_REF for decl lookup.
	* omp-low.cc (struct omp_context):
	Add 'hash_map<tree, tree> *block_vars_map' field.
	(omp_copy_decl_2): Create/lookup using ctx->block_vars_map first. Add
	new copy into ctx->block_vars_map.
	(install_var_field): Add 'bool field_may_exist = false' parameter.
	Adjust lookup assertions.
	(delete_omp_context): Add delete of ctx->block_vars_map.
	(scan_sharing_clauses): Adjust calls to install_var_field. Adjust
	ARRAY_REF pointer type building to use decl type, rather than generic
	ptr_type_node. For ARRAY_REFs on offloaded constructs, also add base
	expression as field lookup key.
	(omp_reduction_init_op): Add ARRAY_TYPE and RECORD_TYPE init op
	construction.
	(oacc_array_reduction_bias): New function.
	(lower_oacc_reductions): Add array reduction handling code. Arrays use
	a different mode of IFN parameters, using additional 'array_addr' and
	'array_max_idx' arguments. The LHS var is a simple integer for
	dependency ordering.
	(lower_omp_target): Adjust 'offload' condition for GOMP_MAP_POINTER
	case. Generate BUILT_IN_ALLOCA_WITH_ALIGN to create private copy
	for reductions of non-constant size types.
	* omp-oacc-neuter-broadcast.cc (worker_single_copy):
	Add 'hash_set<tree> *array_reduction_base_vars' parameter. Avoid
	propagation for SSA_NAMEs used for array reduction accesses.
	(neuter_worker_single): Add 'hash_set<tree> *array_reduction_base_vars'
	parameter. Adjust recursive calls to self and worker_single_copy.
	(oacc_do_neutering): Add 'hash_set<tree> *array_reduction_base_vars'
	parameter. Adjust call to neuter_worker_single.
	(execute_omp_oacc_neuter_broadcast): Add local
	'hash_set<tree> array_reduction_base_vars' declaration. Collect MEM_REF
	base-pointer SSA_NAMEs of arrays into array_reduction_base_vars. Add
	'&array_reduction_base_vars' argument to call of oacc_do_neutering.
	* omp-offload.cc (#include "cfghooks.h"): Add include.
	(oacc_build_array_copy): New function.
	(oacc_build_array_copy_loop): New function.
	(oacc_build_indexed_ssa_loop): New function.
	(default_goacc_reduction): Adjustments to handle arrays.
	* omp-offload.h (oacc_build_array_copy): New declaration.
	(oacc_build_array_copy_loop): New declaration.
	(oacc_build_indexed_ssa_loop): New declaration.
	* tree-loop-distribution.cc (generate_memset_builtin): Under OpenACC,
	when last stmt of pre-header block is a UNIQUE(OACC_FORK) internal-fn,
	split a new basic block to serve as place of insertion, otherwise
	may fail later checking because UNIQUE(OACC_FORK) counts as control
	flow stmt.
	(generate_memcpy_builtin): Likewise.

gcc/testsuite/ChangeLog:

	* c-c++-common/goacc/readonly-2.c: Adjust test.
	* c-c++-common/goacc/reduction-9.c: Adjust test.
	* c-c++-common/goacc/reduction-11.c: New test.
	* c-c++-common/goacc/reduction-12.c: New test.
	* c-c++-common/goacc/reduction-13.c: New test.
	* c-c++-common/goacc/reduction-14.c: New test.
	* c-c++-common/goacc/reduction-15.c: New test.
	* c-c++-common/goacc/reduction-16.c: New test.

	* g++.dg/goacc/reductions-1.C: Adjust test.

	* gfortran.dg/goacc/array-reduction.f90: Adjust test.
	* gfortran.dg/goacc/enter-exit-data-2.f90: Adjust test.
	* gfortran.dg/goacc/finalize-1.f: Adjust test.
	* gfortran.dg/goacc/kernels-decompose-1.f95: Adjust test.
	* gfortran.dg/goacc/pr70828.f90: Adjust test.
	* gfortran.dg/goacc/reduction.f95: Adjust test.
	* gfortran.dg/gomp/target-enter-exit-data.f90: Adjust test.

libgomp/ChangeLog:

	* testsuite/libgomp.oacc-c-c++-common/reduction.h
	(check_reduction_array_xx): New macro.
	(operator_apply): Likewise.
	(check_reduction_array_op): Likewise.
	(check_reduction_arraysec_op): Likewise.
	(function_apply): Likewise.
	(check_reduction_array_macro): Likewise.
	(check_reduction_arraysec_macro): Likewise.
	(check_reduction_xxx_xx_all): Likewise.
	* testsuite/libgomp.oacc-c-c++-common/reduction-arrays-1.c: New test.
	* testsuite/libgomp.oacc-c-c++-common/reduction-arrays-2.c: New test.
	* testsuite/libgomp.oacc-c-c++-common/reduction-arrays-3.c: New test.
	* testsuite/libgomp.oacc-c-c++-common/reduction-arrays-4.c: New test.
	* testsuite/libgomp.oacc-c-c++-common/reduction-arrays-5.c: New test.
	* testsuite/libgomp.oacc-c-c++-common/reduction-structs-1.c: New test.
	* testsuite/libgomp.oacc-fortran/reduction-10.f90: New test.
	* testsuite/libgomp.oacc-fortran/reduction-11.f90: New test.
	* testsuite/libgomp.oacc-fortran/reduction-12.f90: New test.
	* testsuite/libgomp.oacc-fortran/reduction-13.f90: New test.
	* testsuite/libgomp.oacc-fortran/reduction-14.f90: New test.
	* testsuite/libgomp.oacc-fortran/reduction-15.f90: New test.
	* testsuite/libgomp.oacc-fortran/reduction-16.f90: New test.

1 files changed, 17 insertions, 6 deletions

diff --git a/gcc/c/c-parser.cc b/gcc/c/c-parser.cc
index 6d607e9..fbaa917 100644
--- a/gcc/c/c-parser.cc
+++ b/gcc/c/c-parser.cc
@@ -16945,8 +16945,7 @@ c_parser_omp_variable_list (c_parser *parser,
 	    case OMP_CLAUSE_HAS_DEVICE_ADDR:
 	      array_section_p = false;
 	      dims.truncate (0);
-	      while ((ort != C_ORT_ACC || kind != OMP_CLAUSE_REDUCTION)
-		     && c_parser_next_token_is (parser, CPP_OPEN_SQUARE))
+	      while (c_parser_next_token_is (parser, CPP_OPEN_SQUARE))
 		{
 		  location_t loc = UNKNOWN_LOCATION;
 		  tree low_bound = NULL_TREE, length = NULL_TREE;
@@ -18867,13 +18866,21 @@ c_parser_omp_clause_reduction (c_parser *parser, enum omp_clause_code kind,
 		code = MAX_EXPR;
 		break;
 	      }
+	    if (ort == C_ORT_ACC)
+	      goto name_error;
 	    reduc_id = c_parser_peek_token (parser)->value;
 	    break;
 	  }
 	default:
-	  c_parser_error (parser,
-			  "expected %<+%>, %<*%>, %<-%>, %<&%>, "
-			  "%<^%>, %<|%>, %<&&%>, %<||%> or identifier");
+	name_error:
+	  if (ort == C_ORT_OMP)
+	    c_parser_error (parser,
+			    "expected %<+%>, %<*%>, %<-%>, %<&%>, "
+			    "%<^%>, %<|%>, %<&&%>, %<||%> or identifier");
+	  else
+	    c_parser_error (parser,
+			    "expected %<+%>, %<*%>, %<-%>, %<&%>, "
+			    "%<^%>, %<|%>, %<&&%>, %<||%>, %<min%> or %<max%>");
 	  c_parser_skip_until_found (parser, CPP_CLOSE_PAREN, 0);
 	  return list;
 	}
@@ -18887,6 +18894,11 @@ c_parser_omp_clause_reduction (c_parser *parser, enum omp_clause_code kind,
 
 	  for (c = nl; c != list; c = OMP_CLAUSE_CHAIN (c))
 	    {
+	      OMP_CLAUSE_REDUCTION_CODE (c) = code;
+	      /* OpenACC does not require anything below.  */
+	      if (ort == C_ORT_ACC)
+		continue;
+
 	      tree d = OMP_CLAUSE_DECL (c), type;
 	      if (TREE_CODE (d) != OMP_ARRAY_SECTION)
 		type = TREE_TYPE (d);
@@ -18910,7 +18922,6 @@ c_parser_omp_clause_reduction (c_parser *parser, enum omp_clause_code kind,
 		}
 	      while (TREE_CODE (type) == ARRAY_TYPE)
 		type = TREE_TYPE (type);
-	      OMP_CLAUSE_REDUCTION_CODE (c) = code;
 	      if (task)
 		OMP_CLAUSE_REDUCTION_TASK (c) = 1;
 	      else if (inscan)