Hi, a previous patch of Cesar's has made the middle-end omp-lowering automatically create and insert a tofrom (i.e. present_or_copy) map for parallel reductions. This allowed the user to not need explicit clauses to copy out the reduction result, but because reduction arguments are not marked addressable, async does not work as expected, i.e. the asynchronous copy-out results are not used in the compiler generated code.
This patch fixes this in the front-ends, I've tested this patch without new regressions, and fixes some C++ OpenACC tests that regressed after my last OpenACC async patch. Is this okay for trunk? Thanks, Chung-Lin 2016-05-30 Chung-Lin Tang <clt...@codesourcery.com> c/ * c-typeck.c (c_finish_omp_clauses): Mark OpenACC reduction arguments as addressable. cp/ * semantics.c (finish_omp_clauses): Mark OpenACC reduction arguments as addressable. fortran/ * trans-openmp.c (gfc_trans_oacc_construct): Mark OpenACC reduction arguments as addressable. (gfc_trans_oacc_combined_directive): Likewise.
Index: c/c-typeck.c =================================================================== --- c/c-typeck.c (revision 236845) +++ c/c-typeck.c (working copy) @@ -12575,6 +12575,8 @@ c_finish_omp_clauses (tree clauses, enum c_omp_reg remove = true; break; } + if (ort & C_ORT_ACC) + c_mark_addressable (t); type = TREE_TYPE (t); if (TREE_CODE (t) == MEM_REF) type = TREE_TYPE (type); Index: cp/semantics.c =================================================================== --- cp/semantics.c (revision 236845) +++ cp/semantics.c (working copy) @@ -5827,6 +5827,8 @@ finish_omp_clauses (tree clauses, enum c_omp_regio t = n; goto check_dup_generic_t; } + if (ort & C_ORT_ACC) + cxx_mark_addressable (t); goto check_dup_generic; case OMP_CLAUSE_COPYPRIVATE: copyprivate_seen = true; Index: fortran/trans-openmp.c =================================================================== --- fortran/trans-openmp.c (revision 236845) +++ fortran/trans-openmp.c (working copy) @@ -2704,6 +2704,10 @@ gfc_trans_oacc_construct (gfc_code *code) gfc_start_block (&block); oacc_clauses = gfc_trans_omp_clauses (&block, code->ext.omp_clauses, code->loc); + for (tree c = oacc_clauses; c; c = OMP_CLAUSE_CHAIN (c)) + if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION + && DECL_P (OMP_CLAUSE_DECL (c))) + TREE_ADDRESSABLE (OMP_CLAUSE_DECL (c)) = 1; stmt = gfc_trans_omp_code (code->block->next, true); stmt = build2_loc (input_location, construct_code, void_type_node, stmt, oacc_clauses); @@ -3501,6 +3505,10 @@ gfc_trans_oacc_combined_directive (gfc_code *code) construct_clauses.lists[OMP_LIST_REDUCTION] = NULL; oacc_clauses = gfc_trans_omp_clauses (&block, &construct_clauses, code->loc); + for (tree c = oacc_clauses; c; c = OMP_CLAUSE_CHAIN (c)) + if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION + && DECL_P (OMP_CLAUSE_DECL (c))) + TREE_ADDRESSABLE (OMP_CLAUSE_DECL (c)) = 1; } if (!loop_clauses.seq) pblock = █