f014c65363
* gimplify.c (omp_notice_variable): If n is non-NULL
and no flags change in ORT_TARGET region, don't jump to
do_outer.
(struct gimplify_adjust_omp_clauses_data): New type.
(gimplify_adjust_omp_clauses_1): Adjust for data being
a struct gimplify_adjust_omp_clauses_data pointer instead
of tree *. Pass pre_p as a new argument to
lang_hooks.decls.omp_finish_clause hook.
(gimplify_adjust_omp_clauses): Add pre_p argument, adjust
splay_tree_foreach to pass both list_p and pre_p.
(gimplify_omp_parallel, gimplify_omp_task, gimplify_omp_for,
gimplify_omp_workshare, gimplify_omp_target_update): Adjust
gimplify_adjust_omp_clauses callers.
* langhooks.c (lhd_omp_finish_clause): New function.
* langhooks-def.h (lhd_omp_finish_clause): New prototype.
(LANG_HOOKS_OMP_FINISH_CLAUSE): Define to lhd_omp_finish_clause.
* langhooks.h (struct lang_hooks_for_decls): Add a new
gimple_seq * argument to omp_finish_clause hook.
* omp-low.c (scan_sharing_clauses): Call scan_omp_op on
non-DECL_P OMP_CLAUSE_DECL if ctx->outer.
(scan_omp_parallel, lower_omp_for): When adding
_LOOPTEMP_ clause var, add it to outer ctx's decl_map
as identity.
* tree-core.h (OMP_CLAUSE_MAP_TO_PSET): New map kind.
* tree-nested.c (convert_nonlocal_omp_clauses,
convert_local_omp_clauses): Handle various OpenMP 4.0 clauses.
* tree-pretty-print.c (dump_omp_clause): Handle
OMP_CLAUSE_MAP_TO_PSET.
gcc/cp/
* cp-gimplify.c (cxx_omp_finish_clause): Add a gimple_seq *
argument.
* cp-tree.h (cxx_omp_finish_clause): Adjust prototype.
gcc/fortran/
* cpp.c (cpp_define_builtins): Change _OPENMP macro to
201307.
* dump-parse-tree.c (show_omp_namelist): Add list_type
argument. Adjust for rop being u.reduction_op now,
handle depend_op or map_op.
(show_omp_node): Adjust callers. Print some new
OpenMP 4.0 clauses, adjust for OMP_LIST_DEPEND_{IN,OUT}
becoming a single OMP_LIST_DEPEND.
* f95-lang.c (gfc_handle_omp_declare_target_attribute): New
function.
(gfc_attribute_table): New variable.
(LANG_HOOKS_OMP_FINISH_CLAUSE, LANG_HOOKS_ATTRIBUTE_TABLE): Redefine.
* frontend-passes.c (gfc_code_walker): Handle new OpenMP target
EXEC_OMP_* codes and new clauses.
* gfortran.h (gfc_statement): Add ST_OMP_TARGET, ST_OMP_END_TARGET,
ST_OMP_TARGET_DATA, ST_OMP_END_TARGET_DATA, ST_OMP_TARGET_UPDATE,
ST_OMP_DECLARE_TARGET, ST_OMP_TEAMS, ST_OMP_END_TEAMS,
ST_OMP_DISTRIBUTE, ST_OMP_END_DISTRIBUTE, ST_OMP_DISTRIBUTE_SIMD,
ST_OMP_END_DISTRIBUTE_SIMD, ST_OMP_DISTRIBUTE_PARALLEL_DO,
ST_OMP_END_DISTRIBUTE_PARALLEL_DO, ST_OMP_DISTRIBUTE_PARALLEL_DO_SIMD,
ST_OMP_END_DISTRIBUTE_PARALLEL_DO_SIMD, ST_OMP_TARGET_TEAMS,
ST_OMP_END_TARGET_TEAMS, ST_OMP_TEAMS_DISTRIBUTE,
ST_OMP_END_TEAMS_DISTRIBUTE, ST_OMP_TEAMS_DISTRIBUTE_SIMD,
ST_OMP_END_TEAMS_DISTRIBUTE_SIMD, ST_OMP_TARGET_TEAMS_DISTRIBUTE,
ST_OMP_END_TARGET_TEAMS_DISTRIBUTE,
ST_OMP_TARGET_TEAMS_DISTRIBUTE_SIMD,
ST_OMP_END_TARGET_TEAMS_DISTRIBUTE_SIMD,
ST_OMP_TEAMS_DISTRIBUTE_PARALLEL_DO,
ST_OMP_END_TEAMS_DISTRIBUTE_PARALLEL_DO,
ST_OMP_TARGET_TEAMS_DISTRIBUTE_PARALLEL_DO,
ST_OMP_END_TARGET_TEAMS_DISTRIBUTE_PARALLEL_DO,
ST_OMP_TEAMS_DISTRIBUTE_PARALLEL_DO_SIMD,
ST_OMP_END_TEAMS_DISTRIBUTE_PARALLEL_DO_SIMD,
ST_OMP_TARGET_TEAMS_DISTRIBUTE_PARALLEL_DO_SIMD and
ST_OMP_END_TARGET_TEAMS_DISTRIBUTE_PARALLEL_DO_SIMD.
(symbol_attribute): Add omp_declare_target field.
(gfc_omp_depend_op, gfc_omp_map_op): New enums.
(gfc_omp_namelist): Replace rop field with union
containing reduction_op, depend_op and map_op.
(OMP_LIST_DEPEND_IN, OMP_LIST_DEPEND_OUT): Remove.
(OMP_LIST_DEPEND, OMP_LIST_MAP, OMP_LIST_TO, OMP_LIST_FROM): New.
(gfc_omp_clauses): Add num_teams, device, thread_limit,
dist_sched_kind, dist_chunk_size fields.
(gfc_common_head): Add omp_declare_target field.
(gfc_exec_op): Add EXEC_OMP_TARGET, EXEC_OMP_TARGET_DATA,
EXEC_OMP_TEAMS, EXEC_OMP_DISTRIBUTE, EXEC_OMP_DISTRIBUTE_SIMD,
EXEC_OMP_DISTRIBUTE_PARALLEL_DO, EXEC_OMP_DISTRIBUTE_PARALLEL_DO_SIMD,
EXEC_OMP_TARGET_TEAMS, EXEC_OMP_TEAMS_DISTRIBUTE,
EXEC_OMP_TEAMS_DISTRIBUTE_SIMD, EXEC_OMP_TARGET_TEAMS_DISTRIBUTE,
EXEC_OMP_TARGET_TEAMS_DISTRIBUTE_SIMD,
EXEC_OMP_TEAMS_DISTRIBUTE_PARALLEL_DO,
EXEC_OMP_TARGET_TEAMS_DISTRIBUTE_PARALLEL_DO,
EXEC_OMP_TEAMS_DISTRIBUTE_PARALLEL_DO_SIMD,
EXEC_OMP_TARGET_TEAMS_DISTRIBUTE_PARALLEL_DO_SIMD and
EXEC_OMP_TARGET_UPDATE.
(gfc_add_omp_declare_target): New prototype.
* match.h (gfc_match_omp_declare_target, gfc_match_omp_distribute,
gfc_match_omp_distribute_parallel_do,
gfc_match_omp_distribute_parallel_do_simd,
gfc_match_omp_distribute_simd, gfc_match_omp_target,
gfc_match_omp_target_data, gfc_match_omp_target_teams,
gfc_match_omp_target_teams_distribute,
gfc_match_omp_target_teams_distribute_parallel_do,
gfc_match_omp_target_teams_distribute_parallel_do_simd,
gfc_match_omp_target_teams_distribute_simd,
gfc_match_omp_target_update, gfc_match_omp_teams,
gfc_match_omp_teams_distribute,
gfc_match_omp_teams_distribute_parallel_do,
gfc_match_omp_teams_distribute_parallel_do_simd,
gfc_match_omp_teams_distribute_simd): New prototypes.
* module.c (ab_attribute): Add AB_OMP_DECLARE_TARGET.
(attr_bits): Likewise.
(mio_symbol_attribute): Handle omp_declare_target attribute.
(gfc_free_omp_clauses): Free num_teams, device, thread_limit
and dist_chunk_size expressions.
(OMP_CLAUSE_PRIVATE, OMP_CLAUSE_FIRSTPRIVATE, OMP_CLAUSE_LASTPRIVATE,
OMP_CLAUSE_COPYPRIVATE, OMP_CLAUSE_SHARED, OMP_CLAUSE_COPYIN,
OMP_CLAUSE_REDUCTION, OMP_CLAUSE_IF, OMP_CLAUSE_NUM_THREADS,
OMP_CLAUSE_SCHEDULE, OMP_CLAUSE_DEFAULT, OMP_CLAUSE_ORDERED,
OMP_CLAUSE_COLLAPSE, OMP_CLAUSE_UNTIED, OMP_CLAUSE_FINAL,
OMP_CLAUSE_MERGEABLE, OMP_CLAUSE_ALIGNED, OMP_CLAUSE_DEPEND,
OMP_CLAUSE_INBRANCH, OMP_CLAUSE_LINEAR, OMP_CLAUSE_NOTINBRANCH,
OMP_CLAUSE_PROC_BIND, OMP_CLAUSE_SAFELEN, OMP_CLAUSE_SIMDLEN,
OMP_CLAUSE_UNIFORM): Use 1U instead of 1.
(OMP_CLAUSE_DEVICE, OMP_CLAUSE_MAP, OMP_CLAUSE_TO, OMP_CLAUSE_FROM,
OMP_CLAUSE_NUM_TEAMS, OMP_CLAUSE_THREAD_LIMIT,
OMP_CLAUSE_DIST_SCHEDULE): Define.
(gfc_match_omp_clauses): Change mask parameter to unsigned int.
Adjust for rop becoming u.reduction_op. Disallow inbranch with
notinbranch. For depend clause, always create OMP_LIST_DEPEND
and fill in u.depend_op. Handle num_teams, device, map,
to, from, thread_limit and dist_schedule clauses.
(OMP_DECLARE_SIMD_CLAUSES): Or in OMP_CLAUSE_INBRANCH and
OMP_CLAUSE_NOTINBRANCH.
(OMP_TARGET_CLAUSES, OMP_TARGET_DATA_CLAUSES,
OMP_TARGET_UPDATE_CLAUSES, OMP_TEAMS_CLAUSES,
OMP_DISTRIBUTE_CLAUSES): Define.
(match_omp): New function.
(gfc_match_omp_do, gfc_match_omp_do_simd, gfc_match_omp_parallel,
gfc_match_omp_parallel_do, gfc_match_omp_parallel_do_simd,
gfc_match_omp_parallel_sections, gfc_match_omp_parallel_workshare,
gfc_match_omp_sections, gfc_match_omp_simd, gfc_match_omp_single,
gfc_match_omp_task): Rewritten using match_omp.
(gfc_match_omp_threadprivate, gfc_match_omp_declare_reduction):
Diagnose if the directives are followed by unexpected junk.
(gfc_match_omp_distribute, gfc_match_omp_distribute_parallel_do,
gfc_match_omp_distribute_parallel_do_simd,
gfc_match_omp_distrbute_simd, gfc_match_omp_declare_target,
gfc_match_omp_target, gfc_match_omp_target_data,
gfc_match_omp_target_teams, gfc_match_omp_target_teams_distribute,
gfc_match_omp_target_teams_distribute_parallel_do,
gfc_match_omp_target_teams_distribute_parallel_do_simd,
gfc_match_omp_target_teams_distrbute_simd, gfc_match_omp_target_update,
gfc_match_omp_teams, gfc_match_omp_teams_distribute,
gfc_match_omp_teams_distribute_parallel_do,
gfc_match_omp_teams_distribute_parallel_do_simd,
gfc_match_omp_teams_distrbute_simd): New functions.
* openmp.c (resolve_omp_clauses): Adjust for
OMP_LIST_DEPEND_{IN,OUT} being changed to OMP_LIST_DEPEND. Handle
OMP_LIST_MAP, OMP_LIST_FROM, OMP_LIST_TO, num_teams, device,
dist_chunk_size and thread_limit.
(gfc_resolve_omp_parallel_blocks): Only put sharing clauses into
ctx.sharing_clauses. Call gfc_resolve_omp_do_blocks for various
new EXEC_OMP_* codes.
(resolve_omp_do): Handle various new EXEC_OMP_* codes.
(gfc_resolve_omp_directive): Likewise.
(gfc_resolve_omp_declare_simd): Add missing space to diagnostics.
* parse.c (decode_omp_directive): Handle parsing of OpenMP 4.0
offloading related directives.
(case_executable): Add ST_OMP_TARGET_UPDATE.
(case_exec_markers): Add ST_OMP_TARGET*, ST_OMP_TEAMS*,
ST_OMP_DISTRIBUTE*.
(case_decl): Add ST_OMP_DECLARE_TARGET.
(gfc_ascii_statement): Handle new ST_OMP_* codes.
(parse_omp_do): Handle various new ST_OMP_* codes.
(parse_executable): Likewise.
* resolve.c (gfc_resolve_blocks): Handle various new EXEC_OMP_*
codes.
(resolve_code): Likewise.
(resolve_symbol): Change that !$OMP DECLARE TARGET variables
are saved.
* st.c (gfc_free_statement): Handle various new EXEC_OMP_* codes.
* symbol.c (check_conflict): Check omp_declare_target conflicts.
(gfc_add_omp_declare_target): New function.
(gfc_copy_attr): Copy omp_declare_target.
* trans.c (trans_code): Handle various new EXEC_OMP_* codes.
* trans-common.c (build_common_decl): Add "omp declare target"
attribute if needed.
* trans-decl.c (add_attributes_to_decl): Likewise.
* trans.h (gfc_omp_finish_clause): New prototype.
* trans-openmp.c (gfc_omp_finish_clause): New function.
(gfc_trans_omp_reduction_list): Adjust for rop being renamed
to u.reduction_op.
(gfc_trans_omp_clauses): Adjust for OMP_LIST_DEPEND_{IN,OUT}
change to OMP_LIST_DEPEND and fix up depend handling.
Handle OMP_LIST_MAP, OMP_LIST_TO, OMP_LIST_FROM, num_teams,
thread_limit, device, dist_chunk_size and dist_sched_kind.
(gfc_trans_omp_do): Handle EXEC_OMP_DISTRIBUTE.
(GFC_OMP_SPLIT_DISTRIBUTE, GFC_OMP_SPLIT_TEAMS,
GFC_OMP_SPLIT_TARGET, GFC_OMP_SPLIT_NUM, GFC_OMP_MASK_DISTRIBUTE,
GFC_OMP_MASK_TEAMS, GFC_OMP_MASK_TARGET, GFC_OMP_MASK_NUM): New.
(gfc_split_omp_clauses): Handle splitting of clauses for new
EXEC_OMP_* codes.
(gfc_trans_omp_do_simd): Add pblock argument, adjust for being
callable for combined constructs.
(gfc_trans_omp_parallel_do, gfc_trans_omp_parallel_do_simd): Likewise.
(gfc_trans_omp_distribute, gfc_trans_omp_teams,
gfc_trans_omp_target, gfc_trans_omp_target_data,
gfc_trans_omp_target_update): New functions.
(gfc_trans_omp_directive): Adjust gfc_trans_omp_* callers, handle
new EXEC_OMP_* codes.
gcc/testsuite/
* gfortran.dg/gomp/declare-simd-1.f90: New test.
* gfortran.dg/gomp/depend-1.f90: New test.
* gfortran.dg/gomp/target1.f90: New test.
* gfortran.dg/gomp/target2.f90: New test.
* gfortran.dg/gomp/target3.f90: New test.
* gfortran.dg/gomp/udr4.f90: Adjust expected diagnostics.
* gfortran.dg/openmp-define-3.f90: Expect _OPENMP 201307 instead of
201107.
libgomp/
* omp_lib.f90.in (openmp_version): Set to 201307.
* omp_lib.h.in (openmp_version): Likewise.
* testsuite/libgomp.c/target-8.c: New test.
* testsuite/libgomp.fortran/declare-simd-1.f90: Add notinbranch
and inbranch clauses.
* testsuite/libgomp.fortran/depend-3.f90: New test.
* testsuite/libgomp.fortran/openmp_version-1.f: Adjust for new
openmp_version.
* testsuite/libgomp.fortran/openmp_version-2.f90: Likewise.
* testsuite/libgomp.fortran/target1.f90: New test.
* testsuite/libgomp.fortran/target2.f90: New test.
* testsuite/libgomp.fortran/target3.f90: New test.
* testsuite/libgomp.fortran/target4.f90: New test.
* testsuite/libgomp.fortran/target5.f90: New test.
* testsuite/libgomp.fortran/target6.f90: New test.
* testsuite/libgomp.fortran/target7.f90: New test.
From-SVN: r211768
51 lines
1.2 KiB
Fortran
51 lines
1.2 KiB
Fortran
! { dg-do run }
|
|
|
|
module target6
|
|
contains
|
|
subroutine foo (p, v, w, n)
|
|
double precision, pointer :: p(:), v(:), w(:)
|
|
double precision :: q(n)
|
|
integer :: i, n
|
|
!$omp target data if (n > 256) map (to: v(1:n), w(:n)) map (from: p(1:n), q)
|
|
!$omp target if (n > 256)
|
|
!$omp parallel do simd
|
|
do i = 1, n
|
|
p(i) = v(i) * w(i)
|
|
q(i) = p(i)
|
|
end do
|
|
!$omp end target
|
|
!$omp target update if (n > 256) from (p)
|
|
do i = 1, n
|
|
if (p(i) /= i * iand (i, 63)) call abort
|
|
v(i) = v(i) + 1
|
|
end do
|
|
!$omp target update if (n > 256) to (v(1:n))
|
|
!$omp target if (n > 256)
|
|
!$omp parallel do simd
|
|
do i = 1, n
|
|
p(i) = v(i) * w(i)
|
|
end do
|
|
!$omp end target
|
|
!$omp end target data
|
|
do i = 1, n
|
|
if (q(i) /= (v(i) - 1) * w(i)) call abort
|
|
if (p(i) /= q(i) + w(i)) call abort
|
|
end do
|
|
end subroutine
|
|
end module target6
|
|
use target6, only : foo
|
|
integer :: n, i
|
|
double precision, pointer :: p(:), v(:), w(:)
|
|
n = 10000
|
|
allocate (p(n), v(n), w(n))
|
|
do i = 1, n
|
|
v(i) = i
|
|
w(i) = iand (i, 63)
|
|
end do
|
|
call foo (p, v, w, n)
|
|
do i = 1, n
|
|
if (p(i) /= (i + 1) * iand (i, 63)) call abort
|
|
end do
|
|
deallocate (p, v, w)
|
|
end
|