[4/5] gomp-3_0-branch merge to trunk - OpenMP specific libgomp runtime changes
Jakub Jelinek
jakub@redhat.com
Thu Jun 5 16:06:00 GMT 2008
Hi!
This patch contains the OpenMP specific changes to the libgomp
runtime library.
If [1/5] and [2/5] patches are committed, I'll commit this to the
trunk, unless I hear objections.
2008-06-05 Jakub Jelinek <jakub@redhat.com>
Richard Henderson <rth@redhat.com>
Ulrich Drepper <drepper@redhat.com>
Jakob Blomer <jakob.blomer@ira.uka.de>
* configure.ac (LIBGOMP_GNU_SYMBOL_VERSIONING): New AC_DEFINE.
Substitute also OMP_*LOCK_25*.
* configure: Regenerated.
* config.h.in: Regenerated.
* Makefile.am (libgomp_la_SOURCES): Add loop_ull.c, iter_ull.c,
ptrlock.c and task.c.
* Makefile.in: Regenerated.
* testsuite/Makefile.in: Regenerated.
* task.c: New file.
* loop_ull.c: New file.
* iter_ull.c: New file.
* libgomp.h: Include ptrlock.h.
(enum gomp_task_kind): New type.
(struct gomp_team): Add task_lock, task_queue, task_count,
task_running_count, single_count fields. Add
work_share_list_free_lock ifndef HAVE_SYNC_BUILTINS.
Remove work_share_lock, generation_mask,
oldest_live_gen, num_live_gen and init_work_shares fields, add
work work_share_list_alloc, work_share_list_free and work_share_chunk
fields. Change work_shares from pointer to pointers into an array.
Change ordered_release field into gomp_sem_t ** from flexible array
member. Add implicit_task and initial_work_shares fields.
Move close to the end of the struct.
(struct gomp_team_state): Add single_count, last_work_share,
active_level and level fields, remove work_share_generation.
(gomp_barrier_handle_tasks): New prototype.
(gomp_finish_task): New inline function.
(struct gomp_work_share): Move chunk_size, end, incr into
transparent union/struct, add chunk_size_ull, end_ll, incr_ll and
next_ll fields. Reshuffle fields. Add next_alloc,
next_ws, next_free and inline_ordered_team_ids fields, change
ordered_team_ids into pointer from flexible array member.
Add mode field. Put lock and next into a different cache line
from most of the write-once fields.
(gomp_iter_ull_static_next, gomp_iter_ull_dynamic_next_locked,
gomp_iter_ull_guided_next_locked, gomp_iter_ull_dynamic_next,
gomp_iter_ull_guided_next): New prototypes.
(gomp_new_icv): New prototype.
(struct gomp_thread): Add thread_pool and task fields.
(struct gomp_thread_pool): New type.
(gomp_new_team): New prototype.
(gomp_team_start): Change type of last argument.
(gomp_new_work_share): Removed.
(gomp_init_work_share, gomp_fini_work_share): New prototypes.
(gomp_work_share_init_done): New static inline.
(gomp_throttled_spin_count_var, gomp_available_cpus,
gomp_managed_threads): New extern decls.
(gomp_init_task): New prototype.
(gomp_spin_count_var): New extern var decl.
(LIBGOMP_GNU_SYMBOL_VERSIONING): Undef if no visibility
or no alias support, or if not PIC.
(gomp_init_lock_30, gomp_destroy_lock_30, gomp_set_lock_30,
gomp_unset_lock_30, gomp_test_lock_30, gomp_init_nest_lock_30,
gomp_destroy_nest_lock_30, gomp_set_nest_lock_30,
gomp_unset_nest_lock_30, gomp_test_nest_lock_30, gomp_init_lock_25,
gomp_destroy_lock_25, gomp_set_lock_25, gomp_unset_lock_25,
gomp_test_lock_25, gomp_init_nest_lock_25, gomp_destroy_nest_lock_25,
gomp_set_nest_lock_25, gomp_unset_nest_lock_25,
gomp_test_nest_lock_25): New prototypes.
(omp_lock_symver, strong_alias): Define.
(gomp_remaining_threads_count, gomp_remaining_threads_lock): New
decls.
(gomp_end_task): New.
(struct gomp_task_icv, gomp_global_icv): New.
(gomp_thread_limit_var, gomp_max_active_levels_var): New.
(struct gomp_task): New.
(gomp_nthreads_var, gomp_dyn_var, gomp_nest_var,
gomp_run_sched_var, gomp_run_sched_chunk): Remove.
(gomp_icv): New.
(gomp_schedule_type): Reorder enum to match
omp_sched_t.
* team.c (struct gomp_thread_start_data): Add thread_pool and task
fields.
(gomp_thread_start): Add gomp_team_barrier_wait call.
For non-nested case remove clearing of docked thread thr fields.
Use pool fields instead of global gomp_* variables. Use
gomp_barrier_wait_last when needed. Initialize ts.active_level.
Create tasks for each member thread.
(free_team): Only destroy team barrier, task_lock here and free it.
(gomp_free_thread): Free last_team if non-NULL.
(gomp_team_end): Call gomp_team_barrier_wait instead of
gomp_barrier_wait. For nested case call one extra
gomp_barrier_wait. Move here some destruction from free_team.
Call free_team on pool->last_team if any, rather than freeing
current team. Destroy work_share_list_free_lock ifndef
HAVE_SYNC_BUILTINS.
(gomp_new_icv): New function.
(gomp_threads, gomp_threads_size, gomp_threads_used,
gomp_threads_dock): Removed.
(gomp_thread_destructor): New variable.
(gomp_new_thread_pool, gomp_free_pool_helper, gomp_free_thread): New
functions.
(gomp_team_start): Create new pool if current thread doesn't have
one. Use pool fields instead of global gomp_* variables.
Initialize thread_pool field for new threads. Clear single_count.
Change last argument from ws to team, don't create
new team, set ts.work_share to &team->work_shares[0] and clear
ts.last_work_share. Don't clear ts.work_share_generation.
If number of threads changed, adjust atomically gomp_managed_threads.
Use gomp_init_task instead of gomp_new_task,
set thr->task to the corresponding implicit_task array entry.
Create tasks for each member thread. Initialize ts.level.
(initialize_team): Call pthread_key_create on
gomp_thread_destructor.
(team_destructor): New function.
(new_team): Removed.
(gomp_new_team): New function.
(free_team): Free gomp_work_share blocks chained through next_alloc,
instead of freeing work_shares and destroying work_share_lock.
(gomp_team_end): Call gomp_fini_work_share. If number of threads
changed, adjust atomically gomp_managed_threads. Use gomp_end_task.
* barrier.c (GOMP_barrier): Call gomp_team_barrier_wait instead
of gomp_barrier_wait.
* single.c (GOMP_single_copy_start): Call gomp_team_barrier_wait
instead of gomp_barrier_wait. Call gomp_work_share_init_done
if gomp_work_share_start returned true. Don't unlock ws->lock.
(GOMP_single_copy_end): Call gomp_team_barrier_wait instead
of gomp_barrier_wait.
(GOMP_single_start): Rewritten if HAVE_SYNC_BUILTINS. Call
gomp_work_share_init_done if gomp_work_share_start returned true.
Don't unlock ws->lock.
* work.c: Include stddef.h.
(free_work_share): Use work_share_list_free_lock instead
of atomic chaining ifndef HAVE_SYNC_BUILTINS. Add team argument.
Call gomp_fini_work_share and then either free ws if orphaned, or
put it into work_share_list_free list of the current team.
(alloc_work_share, gomp_init_work_share, gomp_fini_work_share): New
functions.
(gomp_work_share_start, gomp_work_share_end,
gomp_work_share_end_nowait): Rewritten.
* omp_lib.f90.in Change some tabs to spaces to prevent warnings.
(openmp_version): Set to 200805.
(omp_sched_kind, omp_sched_static, omp_sched_dynamic,
omp_sched_guided, omp_sched_auto): New parameters.
(omp_set_schedule, omp_get_schedule, omp_get_thread_limit,
omp_set_max_active_levels, omp_get_max_active_levels,
omp_get_level, omp_get_ancestor_thread_num, omp_get_team_size,
omp_get_active_level): New interfaces.
* omp_lib.h.in (openmp_version): Set to 200805.
(omp_sched_kind, omp_sched_static, omp_sched_dynamic,
omp_sched_guided, omp_sched_auto): New parameters.
(omp_set_schedule, omp_get_schedule, omp_get_thread_limit,
omp_set_max_active_levels, omp_get_max_active_levels,
omp_get_level, omp_get_ancestor_thread_num, omp_get_team_size,
omp_get_active_level): New externals.
* loop.c: Include limits.h.
(GOMP_loop_runtime_next, GOMP_loop_ordered_runtime_next): Handle
GFS_AUTO.
(GOMP_loop_runtime_start, GOMP_loop_ordered_runtime_start):
Likewise. Use gomp_icv.
(gomp_loop_static_start, gomp_loop_dynamic_start): Clear
ts.static_trip here.
(gomp_loop_static_start, gomp_loop_ordered_static_start): Call
gomp_work_share_init_done after gomp_loop_init. Don't unlock ws->lock.
(gomp_loop_dynamic_start, gomp_loop_guided_start): Call
gomp_work_share_init_done after gomp_loop_init. If HAVE_SYNC_BUILTINS,
don't unlock ws->lock, otherwise lock it.
(gomp_loop_ordered_dynamic_start, gomp_loop_ordered_guided_start): Call
gomp_work_share_init_done after gomp_loop_init. Lock ws->lock.
(gomp_parallel_loop_start): Call gomp_new_team instead of
gomp_new_work_share. Call gomp_loop_init on &team->work_shares[0].
Adjust gomp_team_start caller. Pass 0 as second argument to
gomp_resolve_num_threads.
(gomp_loop_init): For GFS_DYNAMIC, multiply ws->chunk_size by incr.
If adding ws->chunk_size nthreads + 1 times after end won't
overflow, set ws->mode to 1.
* libgomp_g.h (GOMP_loop_ull_static_start, GOMP_loop_ull_dynamic_start,
GOMP_loop_ull_guided_start, GOMP_loop_ull_runtime_start,
GOMP_loop_ull_ordered_static_start,
GOMP_loop_ull_ordered_dynamic_start,
GOMP_loop_ull_ordered_guided_start,
GOMP_loop_ull_ordered_runtime_start, GOMP_loop_ull_static_next,
GOMP_loop_ull_dynamic_next, GOMP_loop_ull_guided_next,
GOMP_loop_ull_runtime_next, GOMP_loop_ull_ordered_static_next,
GOMP_loop_ull_ordered_dynamic_next, GOMP_loop_ull_ordered_guided_next,
GOMP_loop_ull_ordered_runtime_next, GOMP_task, GOMP_taskwait): New
prototypes.
* libgomp.map: Export lock routines also @@OMP_2.0.
(GOMP_loop_ordered_dynamic_first,
GOMP_loop_ordered_guided_first, GOMP_loop_ordered_runtime_first,
GOMP_loop_ordered_static_first): Remove.
(GOMP_loop_ull_dynamic_next, GOMP_loop_ull_dynamic_start,
GOMP_loop_ull_guided_next, GOMP_loop_ull_guided_start,
GOMP_loop_ull_ordered_dynamic_next,
GOMP_loop_ull_ordered_dynamic_start,
GOMP_loop_ull_ordered_guided_next,
GOMP_loop_ull_ordered_guided_start,
GOMP_loop_ull_ordered_runtime_next,
GOMP_loop_ull_ordered_runtime_start,
GOMP_loop_ull_ordered_static_next,
GOMP_loop_ull_ordered_static_start,
GOMP_loop_ull_runtime_next, GOMP_loop_ull_runtime_start,
GOMP_loop_ull_static_next, GOMP_loop_ull_static_start,
GOMP_task, GOMP_taskwait): Export @@GOMP_2.0.
(omp_set_schedule, omp_get_schedule,
omp_get_thread_limit, omp_set_max_active_levels,
omp_get_max_active_levels, omp_get_level,
omp_get_ancestor_thread_num, omp_get_team_size, omp_get_active_level,
omp_set_schedule_, omp_set_schedule_8_,
omp_get_schedule_, omp_get_schedule_8_, omp_get_thread_limit_,
omp_set_max_active_levels_, omp_set_max_active_levels_8_,
omp_get_max_active_levels_, omp_get_level_,
omp_get_ancestor_thread_num_, omp_get_ancestor_thread_num_8_,
omp_get_team_size_, omp_get_team_size_8_, omp_get_active_level_):
New exports @@OMP_3.0.
* omp.h.in (omp_sched_t): New type.
(omp_set_schedule, omp_get_schedule, omp_get_thread_limit,
omp_set_max_active_levels, omp_get_max_active_levels,
omp_get_level, omp_get_ancestor_thread_num, omp_get_team_size,
omp_get_active_level): New prototypes.
* env.c (gomp_spin_count_var, gomp_throttled_spin_count_var,
gomp_available_cpus, gomp_managed_threads, gomp_max_active_levels_var,
gomp_thread_limit_var, gomp_remaining_threads_count,
gomp_remaining_threads_lock): New variables.
(parse_spincount): New function.
(initialize_env): Call gomp_init_num_threads unconditionally.
Initialize gomp_available_cpus. Call parse_spincount,
initialize gomp_{,throttled_}spin_count_var
depending on presence and value of OMP_WAIT_POLICY and
GOMP_SPINCOUNT env vars. Handle GOMP_BLOCKTIME env var.
Handle OMP_WAIT_POLICY, OMP_MAX_ACTIVE_LEVELS,
OMP_THREAD_LIMIT, OMP_STACKSIZE env vars. Handle unit specification
for GOMP_STACKSIZE. Initialize gomp_remaining_threads_count and
gomp_remaining_threads_lock if needed. Use gomp_global_icv.
(gomp_nthreads_var, gomp_dyn_var, gomp_nest_var,
gomp_run_sched_var, gomp_run_sched_chunk): Remove.
(gomp_global_icv): New.
(parse_schedule): Use it. Parse "auto".
(omp_set_num_threads): Use gomp_icv.
(omp_set_dynamic, omp_get_dynamic, omp_set_nested, omp_get_nested):
Likewise.
(omp_get_max_threads): Move from parallel.c.
(omp_set_schedule, omp_get_schedule, omp_get_thread_limit,
omp_set_max_active_levels, omp_get_max_active_levels): New functions,
add ialias.
(parse_stacksize, parse_wait_policy): New functions.
* fortran.c: Rewrite lock wrappers, if symbol versioning provide
both wrappers for compatibility and new locks.
(omp_set_schedule, omp_get_schedule,
omp_get_thread_limit, omp_set_max_active_levels,
omp_get_max_active_levels, omp_get_level,
omp_get_ancestor_thread_num, omp_get_team_size,
omp_get_active_level): New ialias_redirect.
(omp_set_schedule_, omp_set_schedule_8_,
omp_get_schedule_, omp_get_schedule_8_, omp_get_thread_limit_,
omp_set_max_active_levels_, omp_set_max_active_levels_8_,
omp_get_max_active_levels_, omp_get_level_,
omp_get_ancestor_thread_num_, omp_get_ancestor_thread_num_8_,
omp_get_team_size_, omp_get_team_size_8_, omp_get_active_level_):
New functions.
* parallel.c: Include limits.h.
(gomp_resolve_num_threads): Add count argument. Rewritten.
(GOMP_parallel_start): Call gomp_new_team and pass that as last
argument to gomp_team_start. Pass 0 as second argument to
gomp_resolve_num_threads.
(GOMP_parallel_end): Decrease gomp_remaining_threads_count
if gomp_thread_limit_var != ULONG_MAX.
(omp_in_parallel): Implement using ts.active_level.
(omp_get_max_threads): Move to env.c.
(omp_get_level, omp_get_ancestor_thread_num,
omp_get_team_size, omp_get_active_level): New functions,
add ialias.
* sections.c (GOMP_sections_start): Call gomp_work_share_init_done
after gomp_sections_init. If HAVE_SYNC_BUILTINS, call
gomp_iter_dynamic_next instead of the _locked variant and don't take
lock around it, otherwise acquire it before calling
gomp_iter_dynamic_next_locked.
(GOMP_sections_next): If HAVE_SYNC_BUILTINS, call
gomp_iter_dynamic_next instead of the _locked variant and don't take
lock around it.
(GOMP_parallel_sections_start): Call gomp_new_team instead of
gomp_new_work_share. Call gomp_sections_init on &team->work_shares[0].
Adjust gomp_team_start caller. Pass count as second argument to
gomp_resolve_num_threads, don't adjust num_threads after the call.
Use gomp_icv.
* iter.c (gomp_iter_dynamic_next_locked): Don't multiply
ws->chunk_size by incr.
(gomp_iter_dynamic_next): Likewise. If ws->mode, use more efficient
code.
* libgomp_f.h.in (omp_lock_25_arg_t, omp_nest_lock_25_arg_t): New
types.
(omp_lock_25_arg, omp_nest_lock_25_arg): New macros.
(omp_check_defines): Check even the compat defines.
* config/linux/ptrlock.c: New file.
* config/linux/ptrlock.h: New file.
* config/linux/wait.h: New file.
* config/posix/ptrlock.c: New file.
* config/posix/ptrlock.h: New file.
* config/linux/bar.h (gomp_team_barrier_wait,
gomp_team_barrier_wait_end, gomp_team_barrier_wake): New prototypes.
(gomp_team_barrier_set_task_pending,
gomp_team_barrier_clear_task_pending,
gomp_team_barrier_set_waiting_for_tasks,
gomp_team_barrier_waiting_for_tasks,
gomp_team_barrier_done): New inlines.
(gomp_barrier_t): Rewritten.
(gomp_barrier_state_t): New typedef.
(gomp_barrier_init, gomp_barrier_reinit, gomp_barrier_destroy,
gomp_barrier_wait_start): Rewritten.
(gomp_barrier_wait_end): Change second argument to
gomp_barrier_state_t.
(gomp_barrier_last_thread, gomp_barrier_wait_last): New static
inlines.
* config/linux/bar.c: Include wait.h instead of libgomp.h and
futex.h.
(gomp_barrier_wait_end): Rewritten.
(gomp_team_barrier_wait, gomp_team_barrier_wait_end,
gomp_team_barrier_wake, gomp_barrier_wait_last): New functions.
* config/posix/bar.h (gomp_barrier_t): Add generation field.
(gomp_barrier_state_t): New typedef.
(gomp_team_barrier_wait,
gomp_team_barrier_wait_end, gomp_team_barrier_wake): New prototypes.
(gomp_barrier_wait_start): Or all but low 2 bits from generation
into the return value. Return gomp_barrier_state_t.
(gomp_team_barrier_set_task_pending,
gomp_team_barrier_clear_task_pending,
gomp_team_barrier_set_waiting_for_tasks,
gomp_team_barrier_waiting_for_tasks,
gomp_team_barrier_done): New inlines.
(gomp_barrier_wait_end): Change second argument to
gomp_barrier_state_t.
(gomp_barrier_last_thread, gomp_barrier_wait_last): New static
inlines.
* config/posix/bar.c (gomp_barrier_init): Clear generation field.
(gomp_barrier_wait_end): Change second argument to
gomp_barrier_state_t.
(gomp_team_barrier_wait, gomp_team_barrier_wait_end,
gomp_team_barrier_wake): New functions.
* config/linux/mutex.c: Include wait.h instead of libgomp.h and
futex.h.
(gomp_futex_wake, gomp_futex_wait): New variables.
(gomp_mutex_lock_slow): Call do_wait instead of futex_wait.
* config/linux/lock.c: Rewrite to make locks task owned,
for backwards compatibility provide the old entrypoints
if symbol versioning. Include wait.h instead of libgomp.h and
futex.h.
(gomp_set_nest_lock_25): Call do_wait instead of futex_wait.
* config/posix95/lock.c: Rewrite to make locks task owned,
for backwards compatibility provide the old entrypoints
if symbol versioning.
* config/posix/lock.c: Rewrite to make locks task owned,
for backwards compatibility provide the old entrypoints
if symbol versioning.
* config/linux/proc.c (gomp_init_num_threads): Use gomp_global_icv.
(get_num_procs, gomp_dynamic_max_threads): Use gomp_icv.
* config/posix/proc.c, config/mingw32/proc.c: Similarly.
* config/linux/powerpc/futex.h (FUTEX_WAIT, FUTEX_WAKE): Remove.
(sys_futex0): Return error code.
(futex_wake, futex_wait): If ENOSYS was returned, clear
FUTEX_PRIVATE_FLAG in gomp_futex_wa{ke,it} and retry.
(cpu_relax, atomic_write_barrier): New static inlines.
* config/linux/alpha/futex.h (FUTEX_WAIT, FUTEX_WAKE): Remove.
(futex_wake, futex_wait): If ENOSYS was returned, clear
FUTEX_PRIVATE_FLAG in gomp_futex_wa{ke,it} and retry.
(cpu_relax, atomic_write_barrier): New static inlines.
* config/linux/x86/futex.h (FUTEX_WAIT, FUTEX_WAKE): Remove.
(sys_futex0): Return error code.
(futex_wake, futex_wait): If ENOSYS was returned, clear
FUTEX_PRIVATE_FLAG in gomp_futex_wa{ke,it} and retry.
(cpu_relax, atomic_write_barrier): New static inlines.
* config/linux/s390/futex.h (FUTEX_WAIT, FUTEX_WAKE): Remove.
(sys_futex0): Return error code.
(futex_wake, futex_wait): If ENOSYS was returned, clear
FUTEX_PRIVATE_FLAG in gomp_futex_wa{ke,it} and retry.
(cpu_relax, atomic_write_barrier): New static inlines.
* config/linux/ia64/futex.h (FUTEX_WAIT, FUTEX_WAKE): Remove.
(sys_futex0): Return error code.
(futex_wake, futex_wait): If ENOSYS was returned, clear
FUTEX_PRIVATE_FLAG in gomp_futex_wa{ke,it} and retry.
(cpu_relax, atomic_write_barrier): New static inlines.
* config/linux/sparc/futex.h (FUTEX_WAIT, FUTEX_WAKE): Remove.
(sys_futex0): Return error code.
(futex_wake, futex_wait): If ENOSYS was returned, clear
FUTEX_PRIVATE_FLAG in gomp_futex_wa{ke,it} and retry.
(cpu_relax, atomic_write_barrier): New static inlines.
* config/linux/sem.c: Include wait.h instead of libgomp.h and
futex.h.
(gomp_sem_wait_slow): Call do_wait instead of futex_wait.
* config/linux/affinity.c: Assume HAVE_SYNC_BUILTINS.
* config/linux/omp-lock.h (omp_lock_25_t, omp_nest_lock_25_t): New
types.
(omp_nest_lock_t): Change owner into void *, add lock field.
* config/posix95/omp-lock.h: Include semaphore.h.
(omp_lock_25_t, omp_nest_lock_25_t): New types.
(omp_lock_t): Use sem_t instead of mutex if semaphores
aren't broken.
(omp_nest_lock_t): Likewise. Change owner to void *.
* config/posix/omp-lock.h: Include semaphore.h.
(omp_lock_25_t, omp_nest_lock_25_t): New types.
(omp_lock_t): Use sem_t instead of mutex if semaphores
aren't broken.
(omp_nest_lock_t): Likewise. Add owner field.
Jakub
-------------- next part --------------
A non-text attachment was scrubbed...
Name: gomp3-merge4.patch.bz2
Type: application/x-bzip2
Size: 37651 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20080605/d09d2256/attachment.bz2>
More information about the Gcc-patches
mailing list