Fix concurrent bitmap update in GC_dirty

Fix concurrent bitmap update in GC_dirty

Issue #235 (bdwgc).

* darwin_stop_world.c [!GC_NO_THREADS_DISCOVERY]
(GC_suspend_thread_list): Call GC_acquire_dirty_lock and
GC_release_dirty_lock around thread_suspend().
* darwin_stop_world.c (GC_stop_world): Likewise.
* include/private/gc_priv.h (set_pht_entry_from_index_safe): Remove
unused macro.
* include/private/gc_priv.h [THREADS] (GC_acquire_dirty_lock,
GC_release_dirty_lock): Define new macro; move comment from
win32_threads.c.
* include/private/gc_priv.h [THREADS && !GC_DISABLE_INCREMENTAL]
(GC_fault_handler_lock): Declare global variable (even if not
MPROTECT_VDB).
* os_dep.c [!GC_DISABLE_INCREMENTAL] (async_set_pht_entry_from_index):
Define even if not MPROTECT_VDB; reformat comments.
* os_dep.c [!GC_DISABLE_INCREMENTAL && AO_HAVE_test_and_set_acquire]
(GC_fault_handler_lock): Likewise.
* os_dep.c [THREADS && AO_HAVE_test_and_set_acquire]
(async_set_pht_entry_from_index): Use GC_acquire_dirty_lock and
GC_release_dirty_lock; remove wrong comment.
* os_dep.c [THREADS && !AO_HAVE_test_and_set_acquire]
(currently_updating, async_set_pht_entry_from_index): Remove incorrect
implementation.
* os_dep.c [!GC_DISABLE_INCREMENTAL] (GC_dirty_inner): Call
async_set_pht_entry_from_index instead of set_pht_entry_from_index;
remove FIXME.
* pthread_stop_world.c [GC_ENABLE_SUSPEND_THREAD] (GC_suspend_thread):
Refine comment for AO_store_release call; call GC_acquire_dirty_lock
and GC_release_dirty_lock around a block of RAISE_SIGNAL() and
sem_wait().
* pthread_stop_world.c [GC_OPENBSD_UTHREADS] (GC_suspend_all): Call
GC_acquire_dirty_lock and GC_release_dirty_lock around
pthread_suspend_np().
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS] (GC_suspend_all): Add
comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL] (GC_stop_world):
If GC_manual_vdb then call GC_acquire_dirty_lock and
GC_release_dirty_lock around a block of GC_suspend_all() and
suspend_restart_barrier(); add comment.
* win32_threads.c (GC_suspend): Use GC_acquire_dirty_lock and
GC_release_dirty_lock (regardless of MPROTECT_VDB).