Change compiler invocation example in gc.man to use dynamic libgc

* doc/gc.man (SYNOPSIS): Replace gc.a to -lgc option in the compiler
(cc) invocation example.

Add note of set_free_space_divisor, set_warn_proc ABI change after gc-7.1

Issue #197 (bdwgc).

GC_set_free_space_divisor() and GC_set_warn_proc() used to return
some value.  But starting from version 7.1alpha2, these API functions
do not return any value.

* include/gc.h (GC_set_free_space_divisor, GC_set_warn_proc): Add
comment that the setter used to returned the old value in gc-7.1 (and
previous versions).

Prevent DATASTART redefinition for NaCl
(fix commit 9738a14)

* include/private/gcconfig.h [NACL] (NO_PTHREAD_GETATTR_NP): Define.
* include/private/gcconfig.h [I386 && NACL] (OS_TYPE, etext, DATASTART,
_end, DATAEND, STACK_GRAN, HEURISTIC1, NO_PTHREAD_GETATTR_NP,
GETPAGESIZE, MAX_NACL_GC_THREADS): Remove.

Remove Android-specific code in gcconfig.h for M68K

* include/private/gcconfig.h [M68K && LINUX && __ELF__]
(SEARCH_FOR_DATA_START): Do not check PLATFORM_ANDROID macro.

Travis CI: Fix ASan build with Clang-4.0 (add sudo)

Initial support of TIZEN platform
(part of commit 9379c66 from Unity-Technologies/bdwgc)

Issue #173 (bdwgc).

* include/private/gcconfig.h [TIZEN] (PLATFORM_TIZEN): New macro.
* include/private/gcconfig.h [I386 && LINUX && __ELF__
&& PLATFORM_TIZEN] (SEARCH_FOR_DATA_START, GC_NO_SIGSETJMP): Define.
* include/private/gcconfig.h [ARM32 && LINUX && __ELF__
&& PLATFORM_TIZEN] (SEARCH_FOR_DATA_START): Likewise.
* include/private/gcconfig.h [PLATFORM_TIZEN && !GETPAGESIZE]
(GETPAGESIZE): Define to sysconf(_SC_PAGESIZE).
* include/private/gcconfig.h [CAN_HANDLE_FORK] (CAN_CALL_ATFORK): Do
not define if PLATFORM_TIZEN.

Fix global operator delete definition for C++14 in gc_cpp

Issue #195 (bdwgc).

Sized variants of global operator delete should be defined along with
the non-sized ones.  Otherwise compiler might invoke the one from the
standard library (as observed in e.g. Cygwin) leading to a crash.

* gc_cpp.cc [!_MSC_VER && __cplusplus>201103L] (operator delete):
Define sized variant (the size argument is ignored for now).
* gc_cpp.cc [!_MSC_VER && __cplusplus>201103L && GC_OPERATOR_NEW_ARRAY
&& !CPPCHECK] (operator delete[]): Likewise.

Fix last_reclaimed..gc_no interval comparison to threshold in unmap_old
(fix commit 14c324f8b)

* allchblk.c [USE_MUNMAP] (GC_unmap_old): Replace
(unsigned short)GC_gc_no-hb_last_reclaimed to
(unsigned short)(GC_gc_no-hb_last_reclaimed) to handle
value wrapping in hb_last_reclaimed (and the truncated GC_gc_no)
correctly.

Implement memory unmapping for Sony PS/3
(part of commit 9379c66 from Unity-Technologies/bdwgc)

Issue #173 (bdwgc).

* os_dep.c [USE_MUNMAP && !MSWIN32 && !MSWINCE && SN_TARGET_PS3]:
Include sys/mman.h instead of sys/memory.h.
* os_dep.c [USE_MUNMAP && !USE_WINALLOC && SN_TARGET_PS3]
(GC_unmap): Call ps3_free_mem() instead of mmap().

Remove obsolete advice about linking with _DYNAMIC=0 (Linux)

Issue #196 (bdwgc).

The mentioned GCC option is not needed since gc-5.0alpha3 (when
_DYNAMIC symbol was declared as weak in dyn_load.c).
Even more, the advice is problematic for Alpine Linux (and probably
other systems using musl) as building a program with gcc -static
produces a binary containing a _DYNAMIC symbol with a non-zero address,
and forcing the address to 0 causes a crash during GC initialization.

* doc/README.linux: Remove note that static executable should be
linked with the gcc option "-Wl,-defsym,_DYNAMIC=0".

Reduce the default MUNMAP_THRESHOLD value to 2 for Sony PS/3
(part of commit 9379c66 from Unity-Technologies/bdwgc)

Issue #173 (bdwgc).

* include/private/gcconfig.h [USE_MUNMAP && !MUNMAP_THRESHOLD
&& SN_TARGET_PS3] (MUNMAP_THRESHOLD): Define to 2; add comment.

Add check that gc_cpp operator delete is called (test_cpp)

Issue #195 (bdwgc).

* tests/test_cpp.cc (GC_CHECKED_DELETE): New macro.
* tests/test_cpp.cc (main): Use GC_CHECKED_DELETE() instead of operator
delete (where it is expected that the operator defined in gc_cpp.c/h
is called).

New field (expl_freed_bytes_since_gc) in public prof_stats_s

* include/gc.h (GC_prof_stats_s): Add expl_freed_bytes_since_gc field
(to the end of the structure).
* misc.c [!GC_GET_HEAP_USAGE_NOT_NEEDED] (fill_prof_stats): Store
GC_bytes_freed value to pstats->expl_freed_bytes_since_gc.

New API function (get_expl_freed_bytes_since_gc)

Note: this function could be used in test_cpp to check that the proper
operator delete is called.

* include/gc.h (GC_get_expl_freed_bytes_since_gc): New function
prototype.
* mallocx.c (GC_get_expl_freed_bytes_since_gc): New function definition
(which returns GC_bytes_freed).

Prevent multiple sem_post calls for a thread in suspend_handler
(fix commit af409e4bd)

Issue #181 (bdwgc).

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& THREAD_SANITIZER] (GC_suspend_handler_inner): Call pthread_sigmask()
after last_stop_count update (thus preventing duplicate sem_post() call
in case of GC_suspend_handler_inner is re-entered (if GC_retry_signals);
refine comment.

Do not call sem_getvalue in stop_world if one thread exists

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL] (GC_stop_world):
Do not call sem_getvalue() if n_live_threads is zero (matters only if
GC_retry_signals).

Add assertion for suspend_ack_sem in start_world

This is to check that sem_post() is not called twice for a thread
(to stop) in GC_suspend_handler.

* pthread_stop_world.c [!NACL && GC_ASSERTIONS && !GC_OPENBSD_UTHREADS]
(GC_start_world): Add assertion that GC_suspend_ack_sem has zero value.

Travis CI: Test GC_NETBSD_THREADS_WORKAROUND

Allow GC_NETBSD_THREADS_WORKAROUND macro manual definition

* include/private/gcconfig.h [GC_NETBSD_THREADS]
(GC_NETBSD_THREADS_WORKAROUND): Define; copy comment from
pthread_stop_world.c.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_NETBSD_THREADS] (GC_NETBSD_THREADS_WORKAROUND): Remove.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_restart_ack_sem): Define only if GC_NETBSD_THREADS_WORKAROUND.

Fix error code in abort message if sem_wait failed in start_world (NetBSD)

* pthread_stop_world.c [GC_NETBSD_THREADS_WORKAROUND] (GC_start_world):
Print errno instead of the value returned by sem_wait if the latter has
failed; remove "code" local variable.

Travis CI: Do not use 'trusty' build environment

Update ChangeLog file

Update ChangeLog file (add gc-7.4.8 release date)

Update ChangeLog file (add gc-7.2i release date)

Update ChangeLog file (v7.6 changes only)

Update ChangeLog file (v7.2 - v7.4 changes only)
(Cherry-pick commit b15b52be from 'release-7_4' branch.)

Remove duplicate check whether MPROTECT_VDB is defined
(fix commit db26b7e4b)

* tests/test.c [GC_PTHREADS && MPROTECT_VDB && !REDIRECT_MALLOC
&& !MAKE_BACK_GRAPH && !USE_PROC_FOR_LIBRARIES && !NO_INCREMENTAL]
(main): Remove duplicate defined(MPROTECT_VDB).

Workaround TSan hang in free_inner when called from at-fork child handler

* pthread_support.c [CAN_HANDLE_FORK] (GC_remove_all_threads_but_me):
Do not call GC_INTERNAL_FREE(p) if THREAD_SANITIZER and CAN_CALL_ATFORK;
add comment.

Eliminate UBSan warning of overflow during descr subtraction in mark_from

There is no unsigned overflow because descr has a negative value
actually when it is subtracted from type_descr.

* mark.c (GC_mark_from): Cast descr to signed_word
in (type_descr-(descr+...)).

Workaround TSan false positive in remove_all_threads_but_me

* pthread_support.c [CAN_HANDLE_FORK] (store_to_threads_table): New
static function (defined with GC_ATTR_NO_SANITIZE_THREAD attribute
if CAN_CALL_ATFORK); add comment.
* pthread_support.c [CAN_CALL_ATFORK] (GC_remove_all_threads_but_me):
Call store_to_threads_table instead of GC_threads[hv]=me.

Do not add no_sanitize_thread to fork at-handlers if no pthread_atfork
(fix commit 3a52469)

* pthread_support.c [THREAD_SANITIZER] (GC_wait_for_gc_completion): Do
not skip I_HOLD_LOCK() assertion unless CAN_CALL_ATFORK.
* pthread_support.c [CAN_HANDLE_FORK && GC_ASSERTIONS]
(fork_prepare_proc, fork_parent_proc, fork_child_proc): Do not add
GC_ATTR_NO_SANITIZE_THREAD attribute unless CAN_CALL_ATFORK.

Do not call set_handle_fork(1) in gctest if pthread_atfork not supported

* tests/test.c [!NO_TEST_HANDLE_FORK && HANDLE_FORK]
(INIT_FORK_SUPPORT): Do not define to GC_set_handle_fork(1) unless
CAN_CALL_ATFORK.

Travis CI: Test explicit at-fork handlers invocation (TSan)

Travis CI: Test with increased number of test threads

Revert 'Workaround TSan hang in GC_free_inner when called from at-fork'

This reverts commit d8e271ea8ab4477fc3ac8edb884bbf858d401343.

Because there is a better solution not to skip GC_INTERNAL_FREE in
GC_remove_all_threads_but_me if called from GC_atfork_child.

Really prevent hang in sem_wait in stop_world if compiled with TSan
(fix commit af409e4)

Issue #181 (bdwgc).

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& THREAD_SANITIZER] (GC_retry_signals): Initialize to TRUE.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_store_stack_ptr): Update comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (GC_suspend_thread): Likewise.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& THREAD_SANITIZER] (GC_stop_world): Do not use sem_trywait+sleep
workaround (not needed if GC_retry_signals).

Fix gctest failure if compiled with TSan and parallel marker

As of clang-4.0, Thread Sanitizer does not support creation of threads
in the forked process (before exec).  So, GC_start_mark_threads()
is a no-op now if TSan is enabled.

* misc.c [THREADS && PARALLEL_MARK && CAN_HANDLE_FORK]
(GC_start_mark_threads): Do not call GC_start_mark_threads_inner()
if THREAD_SANITIZER; add comment.

Use same macro (NTHREADS) across all tests to specify number of threads
(code refactoring)

* tests/disclaim_test.c [GC_PTHREADS] (THREAD_CNT): Rename to NTHREADS;
do not define if already defined.
* tests/disclaim_test.c [!GC_PTHREADS] (THREAD_CNT): Rename to
NTHREADS; undefine it first.
* tests/disclaim_test.c (MUTATE_CNT, main): Rename THREAD_CNT to
NTHREADS.
* tests/subthread_create.c [!NTHREADS] (INITIAL_THREAD_COUNT): Rename
to NTHREADS; add comment.
* tests/subthread_create.c (main): Rename INITIAL_THREAD_COUNT to
NTHREADS.
* tests/threadkey_test.c (LIMIT, main): Rename LIMIT to NTHREADS.

Travis CI: Workaround disclaim_test failure when TSan is used
(fix commit 153a2c7)

GC incremental mode is not supported if compiled with TSan for now.

Workaround TSan data race report in check_fl_marks

* alloc.c [GC_ASSERTIONS && THREAD_LOCAL_ALLOC
&& AO_HAVE_load_acquire_read] (GC_check_fl_marks): Skip the whole check
if THREAD_SANITIZER; add comment.

Use __builtin_expect in SIZET_SAT_ADD macro

* include/private/gc_priv.h (EXPECT): Move the definition to be before
SIZET_SAT_ADD.
* include/private/gc_priv.h (SIZET_SAT_ADD): Use EXPECT().

Turn on incremental mode (if available) in disclaim_test

* tests/disclaim_test.c [NO_INCREMENTAL] (main): Call
GC_enable_incremental() afer GC_init_finalized_malloc() call.

Fix marking of disclaim-reachable objects in the incremental mode

Issue #192 (bdwgc).

Unconditional marking is done analogously to marking from uncollectible
blocks, since they are meant to secure access to data for disclaim
notifiers after the data is no longer marked.

* mark.c [ENABLE_DISCLAIM] (GC_push_next_marked_dirty): Honor
MARK_UNCONDITIONALLY when rescanning dirty blocks during incremental
marking.

Travis CI: Add a test with TSan with parallel marker (Linux/clang)

Fix data race in do_local_mark when comparing active_count to helper_count

* mark.c [PARALLEL_MARK] (has_inactive_helpers): New static function
(which compares GC_active_count to GC_helper_count with the mark lock
held).
* mark.c [PARALLEL_MARK && GC_ASSERTIONS] (GC_do_local_mark): Remove
GC_acquire_mark_lock() with immediate GC_acquire_mark_unlock(); add
the appropriate comment to the function description instead.
* mark.c [PARALLEL_MARK] (GC_do_local_mark): Call has_inactive_helpers()
instead of GC_active_count<GC_helper_count; do not call
has_inactive_helpers() unless local_top>local_mark_stack+1.

Fix data race in mark_thread when updating mark_no

* mark.c [PARALLEL_MARK] (GC_mark_local, GC_help_marker): Update
description comment; remove GC_acquire_mark_lock() call at the function
beginning; remove GC_release_mark_lock() call on return.
* mark.c [PARALLEL_MARK] (GC_do_parallel_mark): Remove
GC_acquire/release_mark_lock() calls around GC_mark_local() call.
* pthread_support.c [PARALLEL_MARK] (GC_mark_thread): Remove
GC_release_mark_lock() call.
* win32_threads.c [PARALLEL_MARK] (GC_mark_thread): Likewise.

Workaround TSan warning about data race in generic_malloc_many

It is acceptable to update GC_bytes_found counter without
synchronization (so, it is OK to have inaccurate value in
GC_bytes_found in favor of performance).

* mallocx.c [PARALLEL_MARK && THREAD_SANITIZER]
(GC_generic_malloc_many): Acquire the allocation lock to update
GC_bytes_found (after successful GC_reclaim_generic).
* mallocx.c [PARALLEL_MARK && !THREAD_SANITIZER]
(GC_generic_malloc_many): Move GC_bytes_found update into the block
executed with the mark lock held.

Workaround TSan false positives in add_to_black_list_normal/stack

* blacklst.c (GC_add_to_black_list_normal, GC_add_to_black_list_stack):
Call set_pht_entry_from_index_concurrent() instead of
set_pht_entry_from_index().
* include/private/gc_priv.h [PARALLEL_MARK && THREAD_SANITIZER]
(set_pht_entry_from_index_concurrent): New macro (use AO_or).
* include/private/gc_priv.h [!PARALLEL_MARK || !THREAD_SANITIZER]
(set_pht_entry_from_index_concurrent): Define to
set_pht_entry_from_index.

Document base and size of objects allocated by finalized_malloc

* include/gc_disclaim.h (GC_finalized_malloc): Refine comment (add
a note about GC_base and GC_size).

Update comment in finalized_disclaim to match FINALIZER_CLOSURE_FLAG
(fix commit 012da3f)

* fnlz_mlc.c [ENABLE_DISCLAIM] (GC_finalized_disclaim): Update/refine
comment (about the lowest 2 bits of the first word).

Remove done_init static variable from fnlz_mlc.c
(code refactoring)

GC_init_finalized_malloc!=0 is used instead of boolean done_init.

* fnlz_mlc.c [ENABLE_DISCLAIM] (done_init): Remove static variable.
* fnlz_mlc.c [ENABLE_DISCLAIM] (GC_init_finalized_malloc,
GC_finalized_malloc): Use GC_finalized_kind instead of done_init.
* fnlz_mlc.c [ENABLE_DISCLAIM] (GC_init_finalized_malloc): Add
assertion that the result of GC_new_kind_inner() is non-zero.

Fix assertion violation about disabled cancel in try_to_collect_inner

* misc.c [!GC_DISABLE_INCREMENTAL && !KEEP_BACK_PTRS]
(GC_enable_incremental): Wrap GC_gcollect_inner and GC_read_dirty calls
into DISABLE/RESTORE_CANCEL.

Workaround TSan false positive about read of mark_lock_holder

The assertion should not be violated in GC_acquire_mark_lock even
reading from GC_mark_lock_holder is done before acquiring the mark lock.

* pthread_support.c [PARALLEL_MARK && NUMERIC_THREAD_ID_UNIQUE]
(GC_acquire_mark_lock): Skip assertion on GC_mark_lock_holder (before
acquiring the lock) if THREAD_SANITIZER.
* win32_threads.c [PARALLEL_MARK] (GC_acquire_mark_lock): Likewise.

Add a sanity check that load_acquire and store_release are available

* pthread_stop_world.c [(!AO_HAVE_load_acquire
|| !AO_HAVE_store_release) && !CPPCHECK]: Issue error that
libatomic_ops does not define AO_load_acquire and/or AO_store_release
so the client should manually define AO_REQUIRE_CAS macro.

Improve comment in store_stack_ptr in pthread_stop_world.c
(fix commit a194093)

* pthread_stop_world.c (GC_store_stack_ptr): Refine comment about the
atomic primitives usage.

Remove explicit case of TRUE/FALSE to AO_t in suspend/resume_thread
(fix commit ce09fd5)

* pthread_stop_world.c [GC_ENABLE_SUSPEND_THREAD
&& !GC_OPENBSD_UTHREADS && !NACL] (GC_suspend_thread,
GC_resume_thread): Do not case TRUE/FALSE to AO_t explicitly.

Guard against USE_PTHREAD_LOCKS and USE_SPIN_LOCK are both defined

USE_SPIN_LOCK might be specified by client. As it and USE_PTHREAD_LOCKS
are mutually exclusive, one of them should be undefined.

* include/private/gc_locks.h [(!AO_HAVE_test_and_set_acquire
|| GC_RTEMS_PTHREADS || SN_TARGET_PS3 || GC_WIN32_THREADS || LINT2)
&& GC_PTHREADS] (USE_SPIN_LOCK): Undefine (as USE_PTHREAD_LOCKS is
defined).

Fix data race in make_descriptor when setting explicit_typing_initialized
(fix commit 5f350a0)

* typd_mlc.c [AO_HAVE_load_acquire] (GC_explicit_typing_initialized):
Add volatile qualifier.
* typd_mlc.c [AO_HAVE_load_acquire] (GC_make_descriptor): Remove
cast of &GC_explicit_typing_initialized; do not check THREADS macro;
fallback to locked checking of GC_explicit_typing_initialized if
AO_HAVE_store_release is not defined; use AO_store_release to set
GC_explicit_typing_initialized (to true).
* typd_mlc.c (GC_make_descriptor): Reformat code dealing with
GC_explicit_typing_initialized (check whether AO_HAVE_load_acquire is
defined only once).

Revert 'Eliminate TSan warning about data race in make_descriptor'

This reverts commit a3ee14468b807ecf2d6fa279699b7a9879443bdf.

Because it just hides a potential data race.

Travis CI: Add tests with TSan (only without parallel mark, Linux/clang)

Travis CI: Remove -D NO_INCREMENTAL if --enable-munmap

The incremental mode is disabled in gctest automatically now
if compiled with memory unmapping support.

Do not enable mprotect-based incremental mode if unmapping is on (gctest)

Incremental mode based on mprotect() is not compatible with memory
unmapping for now.

* tests/test.c [!PCR && !GC_WIN32_THREADS && !MAKE_BACK_GRAPH
&& !NO_INCREMENTAL] (main): Do not call GC_enable_incremental() if
MPROTECT_VDB && USE_MUNMAP.

Eliminate TSan false positive for stop_info.stack_ptr (v2)

Without this patch, Thread Sanitizer reports a data race between
GC_has_other_debug_info() and the code which sets stop_info.stack_ptr.

* include/private/pthread_support.h [THREAD_SANITIZER]: Include
dbg_mlc.h file.
* include/private/pthread_support.h [THREAD_SANITIZER] (GC_Thread_Rep):
Add dummy field (as the first field of the structure).
* pthread_support.c [THREAD_SANITIZER && CPPCHECK] (GC_new_thread):
Call GC_noop1(first_thread.dummy[0]) (to suppress cppcheck warning
about unused field).

Revert 'Eliminate TSan false positive for stop_info.stack_ptr'

This reverts commit 80ad68c54e8f339204e7f3b4c305d390e7f44006.

Because it does not eliminate the warning fully, a better solution
would be to add a dummy "oh" header to GC_Thread_Rep.

Eliminate 'value exceeds maximum object size' GCC warning in huge_test

* tests/huge_test.c [!GC_ATTR_ALLOC_SIZE] (GC_ATTR_ALLOC_SIZE): Define
to empty (before include gc.h).

Eliminate 'this statement may fall through' GCC warnings

* cord/cordprnt.c (extract_conv_spec): Eliminate fall through in
a switch statement.
* cord/tests/de.c (do_command): Replace "fall through:" comment with
"FALLTHRU" formal comment (on a new line).
* win32_threads.c [!GC_PTHREADS && !GC_NO_THREADS_DISCOVERY]
(GC_DllMain): Likewise.

Fix 'comparison of 255 with expr of type bool' error in gc_atomic_ops.h

* include/private/gc_atomic_ops.h [GC_BUILTIN_ATOMIC && (CPPCHECK
|| !__GCC_ATOMIC_TEST_AND_SET_TRUEVAL)] (AO_TS_SET): Define to 1
instead of 0xff.

Travis CI: Test with ASan and GCC

Remove some redundant macro definitions from gc_atomic_ops.h
(fix commit 46a2411)

* include/private/gc_atomic_ops.h [GC_BUILTIN_ATOMIC] (AO_HAVE_load,
AO_HAVE_store, AO_HAVE_store_release_write): Remove.
* include/private/gc_atomic_ops.h [GC_BUILTIN_ATOMIC && AO_REQUIRE_CAS]
(AO_HAVE_compare_and_swap): Likewise.
* include/private/gc_atomic_ops.h [!GC_BUILTIN_ATOMIC]: Add comment.
* include/private/gc_priv.h (word_ptr_ao_u): Replace
"ifdef AO_HAVE_load" to "ifdef PARALLEL_MARK".
* tests/test_atomic_ops.c [AO_HAVE_compare_and_swap_release] (main):
Do not check whether AO_REQUIRE_CAS and AO_HAVE_compare_and_swap are
defined.
* typd_mlc.c [GC_FORCE_INCLUDE_ATOMIC_OPS || GC_BUILTIN_ATOMIC]
(include gc_atomic_ops.h): Do not check whether AO_HAVE_load_acquire
is defined.

Adjust AO_HAVE_x check to match AO_fetch_and_add primitive variant used
(fix commit 6043293)

* tests/subthread_create.c: Replace "ifdef AO_HAVE_fetch_and_add" with
"ifdef AO_HAVE_fetch_and_add1".

Prevent test_atomic_ops link fail if AO_nop_full is not defined
(fix commit 066e0ca)

* tests/test_atomic_ops.c (main): Call AO_nop_full() only
if AO_HAVE_nop_full is defined.

Remove redundant THREADS macro checks
(code refactoring)

THREAD_LOCAL_ALLOC macro implies THREADS one.

* alloc.c [GC_ASSERTIONS && THREAD_LOCAL_ALLOC] (GC_check_fl_marks,
GC_check_tls, GC_finish_collection): Do not check whether THREADS macro
is defined.
* include/private/gc_priv.h  [GC_ASSERTIONS && THREAD_LOCAL_ALLOC]
(GC_check_fl_marks): Likewise.

Fix lack of barriers to synchronize memory for suspend_handler

pthread_kill is not on the list of functions which synchronize memory.

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_world_is_stopped): Reformat comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_suspend_handler_inner): Use AO_load_acquire instead of AO_load
to fetch the value of GC_stop_count; add comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_start_world): Use AO_store_release instead of AO_store to reset
GC_world_is_stopped; add comment.

Eliminate TSan false positive related to stop_info.stack_ptr access

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_store_stack_ptr): New inline function definition (move the code
from GC_suspend_handler_inner but stop_info.stack_ptr is now stored
atomically); add comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_suspend_handler_inner): Call GC_store_stack_ptr().
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_push_all_stacks): Use AO_load to fetch p->stop_info.stack_ptr.

Avoid 'Unexpected heap growth' fail if compiled by GCC with ASan (gctest)

* tests/test.c [ADDRESS_SANITIZER && !__clang__] (check_heap_stats):
Multiply max_heap_sz by ~1.66.

Fix data race in last_stop_count access (suspend_handler_inner)

* include/private/pthread_stop_world.h [!GC_OPENBSD_UTHREADS]
(thread_stop_info.last_stop_count): Do not define if NACL; change the
type from word to AO_t; add volatile qualifier; fix a typo in comment
("GC_stop_count").
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(update_last_stop_count): Remove.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_suspend_handler_inner): Replace update_last_stop_count() call with
AO_store_release(&me->stop_info.last_stop_count,my_stop_count).
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_suspend_all): Replace p->stop_info.last_stop_count with
AO_load(&p->stop_info.last_stop_count).

Fix data race in GC_suspend/resume_thread

* include/private/pthread_support.h [GC_ENABLE_SUSPEND_THREAD
&& !GC_DARWIN_THREADS && !GC_OPENBSD_UTHREADS && !NACL]
(GC_Thread_Rep.suspended_ext): New field.
* include/private/pthread_support.h (SUSPENDED_EXT): Remove macro.
* pthread_stop_world.c [GC_ENABLE_SUSPEND_THREAD
&& !GC_OPENBSD_UTHREADS && !NACL] (GC_suspend_handler_inner): Replace
me->flags & SUSPENDED_EXT with AO_load(&me->suspended_ext).
* pthread_stop_world.c [GC_ENABLE_SUSPEND_THREAD
&& !GC_OPENBSD_UTHREADS && !NACL] (suspend_self_inner): Replace
me->flags & SUSPENDED_EXT with AO_load_acquire(&me->suspended_ext).
* pthread_stop_world.c [GC_ENABLE_SUSPEND_THREAD
&& !GC_OPENBSD_UTHREADS && !NACL] (GC_suspend_thread): Replace
t->flags & SUSPENDED_EXT with t->suspended_ext; replace
t->flags|=SUSPENDED_EXT with AO_store_release(&t->suspended_ext, TRUE);
add comment.
* pthread_stop_world.c [GC_ENABLE_SUSPEND_THREAD
&& !GC_OPENBSD_UTHREADS && !NACL] (GC_resume_thread): Replace
t->flags&=~SUSPENDED_EXT with AO_store(&t->suspended_ext, FALSE).
* pthread_stop_world.c [GC_ENABLE_SUSPEND_THREAD
&& !GC_OPENBSD_UTHREADS && !NACL] (GC_is_thread_suspended,
GC_suspend_all, GC_start_world): Replace t->flags & SUSPENDED_EXT with
t->suspended_ext.
* pthread_support.c (GC_register_my_thread): Remove assertion that uses
SUSPENDED_EXT.

Revert 'Workaround TSan false positives in suspend/resume_thread'

This reverts commit 918500dcd5f629391e442f68620873909c733ea7.

Because it just hides the data race issues in GC_suspend/resume_thread.

Workaround TSan hang in GC_free_inner when called from at-fork handler

* pthread_support.c [CAN_HANDLE_FORK] (GC_remove_all_threads_but_me):
Do not call GC_INTERNAL_FREE(p) if THREAD_SANITIZER; add comment.

Workaround TSan false positives for at-fork handlers accessing lock_holder

* pthread_support.c (GC_wait_for_gc_completion): Skip I_HOLD_LOCK()
assertion if THREAD_SANITIZER and CAN_HANDLE_FORK; add comment.
* pthread_support.c [CAN_HANDLE_FORK && GC_ASSERTIONS]
(fork_prepare_proc, fork_parent_proc, fork_child_proc): Add
GC_ATTR_NO_SANITIZE_THREAD attribute.

Workaround TSan false positive in lookup_thread called by suspend_handler

* include/private/pthread_support.h (THREAD_TABLE_INDEX): Move the
macro from pthread_support.c file.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& THREAD_SANITIZER] (GC_lookup_thread_async): New static function
(same implementation as of GC_lookup_thread but with
GC_ATTR_NO_SANITIZE_THREAD attribute); add comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& !THREAD_SANITIZER] (GC_lookup_thread_async): Define as macro.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL]
(GC_suspend_handler_inner): Call GC_lookup_thread_async() instead of
GC_lookup_thread(); move the corresponding comment to
GC_lookup_thread_async.

Workaround TSan false positive about clear_hdr_marks/realloc data race

* mallocx.c (hb_sz_async_grow_within_hblk): New static function (with
GC_ATTR_NO_SANITIZE_THREAD attribute).
* mallocx.c (GC_realloc): Call hb_sz_async_grow_within_hblk instead of
hhdr->hb_sz=sz.

Eliminate TSan warning about data race in default_on_abort

The warning is about a potential race between GC_default_on_abort
(which updates GC_find_leak) and GC_finish_collection (which reads the
variable).

* misc.c [DONT_USE_ATEXIT && !PCR && !SMALL_CONFIG] (skip_gc_atexit):
New static variable; add comment.
* misc.c [DONT_USE_ATEXIT && (PCR || SMALL_CONFIG)] (skip_gc_atexit):
Define as macro (to FALSE).
* misc.c [DONT_USE_ATEXIT] (GC_exit_check): Do not call GC_gcollect
if skip_gc_atexit.
* misc.c [!PCR && !SMALL_CONFIG] (GC_default_on_abort): Do not modify
GC_find_leak.
* misc.c [DONT_USE_ATEXIT && !PCR && !SMALL_CONFIG]
(GC_default_on_abort): Set skip_gc_atexit to true.

Eliminate TSan warning about data race in make_descriptor

* typd_mlc.c [THREADS) && AO_HAVE_load_acquire && AO_HAVE_store]
(GC_make_descriptor): Use AO_store for setting
GC_explicit_typing_initialized (to true).

Support AddressSanitizer (GCC)

* include/private/gcconfig.h [!__has_feature && __SANITIZE_ADDRESS__]
(ADDRESS_SANITIZER): Define.

Fix 'unknown attribute no_sanitize' compiler warnings for GCC
(fix commits ce75cf1, b850e1c)

* include/private/gc_priv.h [!GC_ATTR_NO_SANITIZE_ADDR
&& ADDRESS_SANITIZER && !AO_CLANG_PREREQ(3,8)]
(GC_ATTR_NO_SANITIZE_ADDR): Define to
__attribute__((no_sanitize_address)).
* include/private/gc_priv.h [!GC_ATTR_NO_SANITIZE_MEMORY
&& MEMORY_SANITIZER && !AO_CLANG_PREREQ(3,8)]
(GC_ATTR_NO_SANITIZE_MEMORY): Define to
__attribute__((no_sanitize_memory)).
* include/private/gc_priv.h [!GC_ATTR_NO_SANITIZE_THREAD
&& THREAD_SANITIZER && !AO_CLANG_PREREQ(3,8)]
(GC_ATTR_NO_SANITIZE_THREAD): Define to
__attribute__((no_sanitize_thread)).

Place no_sanitize attributes in a GCC-compliant way

Now the attributes precede the function prototypes (the format is:
<attribute> void f(void) <body>).

* finalize.c (GC_should_invoke_finalizers): Move
GC_ATTR_NO_SANITIZE_THREAD attribute to the beginning of the function
prototype (even before GC_API, GC_INNER or static).
* malloc.c (fill_size_map): Likewise.
* mark.c (GC_noop1, GC_mark_from, GC_push_all_eager): Likewise.
* mark.c [WRAP_MARK_SOME && PARALLEL_MARK] (GC_push_conditional_eager):
Likewise.
* misc.c [THREADS] (next_random_no): Likewise.
* os_dep.c [MPROTECT_VDB && THREADS && AO_HAVE_test_and_set_acquire]
(async_set_pht_entry_from_index): Likewise.
* os_dep.c [MPROTECT_VDB] (is_header_found_async): Likewise.
* pthread_stop_world.c [!NACL && !GC_OPENBSD_UTHREADS]
(update_last_stop_count): Likewise.
* pthread_stop_world.c [!NACL && !GC_OPENBSD_UTHREADS
&& GC_ENABLE_SUSPEND_THREAD] (set_suspended_ext_flag,
GC_resume_thread): Likewise.
* pthread_support.c [THREAD_SANITIZER && (USE_SPIN_LOCK
|| !NO_PTHREAD_TRYLOCK)] (is_collecting): Likewise.
* pthread_support.c [USE_SPIN_LOCK] (set_last_spins_and_high_spin_max,
reset_spin_max): Likewise.
* mark.c (GC_mark_from, GC_push_all_eager): Move
GC_ATTR_NO_SANITIZE_ADDR, GC_ATTR_NO_SANITIZE_MEMORY attributes to the
beginning of the function prototype.
* mach_dep.c (GC_with_callee_saves_pushed): Likewise.
* mark.c [WRAP_MARK_SOME && PARALLEL_MARK] (GC_push_conditional_eager):
Likewise.

Workaround hangs in sigsuspend and sem_wait if compiled with TSan

Issue #181 (bdwgc).

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& THREAD_SANITIZER] (GC_suspend_handler_inner): Call
pthread_sigmask(SIG_SETMASK) with an empty set (thus unmask all
signals); add comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& THREAD_SANITIZER] (GC_stop_world): Call sem_trywait() repeatedly
(with a delay of 100 microseconds) while getting EAGAIN error (instead
of a sem_wait call).

Revert 'Workaround a hang in sigsuspend if compiled with Thread Sanitizer'

This reverts commit 04fd7b5309596423e4342502d3123028d836d551.

Because it does not avoid hangs in all cases when stopping the world.
A better solution is to re-enable signals in the suspend handler.

Allow to turn on spin locking even if thread-local allocations are used

* include/private/gc_locks.h [GC_PTHREADS && !GC_WIN32_THREADS
&& THREAD_LOCAL_ALLOC && !USE_PTHREAD_LOCKS] (UNCOND_LOCK,
UNCOND_UNLOCK): Use spin locks if USE_SPIN_LOCK is defined by client.

Eliminate TSan false positive for stop_info.stack_ptr

Without this patch, Thread Sanitizer reports a data race between
GC_has_other_debug_info() and the code which sets stop_info.stack_ptr.

* include/private/pthread_support.h (GC_Thread_Rep.status): Move the
field to be before stop_info one (so, the latter now has the offset
of at least 3 in words, while oh.oh_sz has the offset of 2 in words
(when KEEP_BACK_PTRS, MAKE_BACK_GRAPH and NEED_CALLINFO are not
defined)).

Eliminate TSan warning about data race when accessing GC_debugging_started

Now GC_debugging_started variable is double-checked with the allocation
lock held before calling GC_start_debugging_inner.

* dbg_mlc.c (GC_start_debugging_inner): Define as GC_INNER (instead of
STATIC).
* dbg_mlc.c [THREADS] (GC_start_debugging): Define as STATIC; do not
call GC_start_debugging_inner() if GC_debugging_started.
* dbg_mlc.c [!THREADS] (GC_start_debugging): Define as macro (redirect
to GC_start_debugging_inner).
* gcj_mlc.c (GC_debug_gcj_malloc): Call GC_start_debugging_inner
(holding the lock) instead of GC_start_debugging.
* include/private/gc_priv.h (GC_start_debugging): Rename to
GC_start_debugging_inner; improve usage comment.

Fix typo in a function name in README.sgi

* doc/README.sgi (4): Fix a typo ("pthread_cond_timedwait").

Workaround TSan false positives in extend_size_map

Thread Sanitizer reports data races between fill_size_map (called from
GC_extend_size_map) and GC_gcj_malloc[_ignore_off_page],
GC_malloc_kind_global, GC_generic_malloc_uncollectable,
GC_malloc_explicitly_typed_ignore_off_page which read a value from
GC_size_map before acquiring the allocation lock.  These races could
be ignored as the value is verified after acquiring the lock.

* alloc.c: Refine comment about GC_allocobj (and GC_size_map) usage
for the case of a multi-threaded environment.
* malloc.c (fill_size_map): New static function (with
GC_ATTR_NO_SANITIZE_THREAD attribute).
* malloc.c (GC_extend_size_map): Use fill_size_map() to fill in
a region of GC_size_map.

Remove GENERAL_MALLOC internal macro
(code refactoring)

* alloc.c: Update comment (about GC_allocobj usage).
* include/private/gc_priv.h (GENERAL_MALLOC): Remove.
* include/private/gc_priv.h (GENERAL_MALLOC_IOP): Move to typd_mlc.c.
* malloc.c (GC_malloc_kind_global): Expand GENERAL_MALLOC macro; copy
the comment from gc_priv.h.

Workaround a hang in sigsuspend if compiled with Thread Sanitizer

Issue #181 (bdwgc).

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& THREAD_SANITIZER] (GC_suspend_handler_inner): Use sched_yield()
instead of sigsuspend(&suspend_handler_mask); add TODO item.

Workaround TSan false positives in suspend/resume_thread

* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (set_suspended_ext_flag): New static
function (with GC_ATTR_NO_SANITIZE_THREAD attribute); add comment.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (GC_suspend_thread): Call
set_suspended_ext_flag() instead of t->flags|=SUSPENDED_EXT.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (GC_resume_thread): Add
GC_ATTR_NO_SANITIZE_THREAD attribute; add comment.

Fix data race when getting object size in explicitly-typed allocators

* typd_mlc.c (COMPLEX): Reformat comment.
* typd_mlc.c (GC_malloc_explicitly_typed,
GC_malloc_explicitly_typed_ignore_off_page, GC_calloc_explicitly_typed):
Always use BYTES_TO_GRANULES(GC_size(op)) instead of GC_size_map[lb] to
determine size of the allocated object in granules (because the value
of GC_size_map[lb] might be updated by another thread since the value
use in GC_malloc_kind); add comment.

Remove 'ifdef USE_SPIN_LOCK' which is always false in gc_locks.h
(fix commit 9a20535)

* include/private/gc_locks.h [!GC_WIN32_THREADS && GC_PTHREADS
&& USE_PTHREAD_LOCKS && !GC_ASSERTIONS && NO_PTHREAD_TRYLOCK
&& USE_SPIN_LOCK] (UNCOND_LOCK): Define to
pthread_mutex_lock(&GC_allocate_ml) instead of GC_lock().

Workaround TSan false positive in write fault handler

The data races reported between GC_write_fault_handler,
catch_exception_raise and GC_install_header, GC_remove_header are OK
as the added or removed header should be already unprotected.

* os_dep.c [MPROTECT_VDB && THREADS] (is_header_found_async): New
static function (with GC_ATTR_NO_SANITIZE_THREAD attribute).
* os_dep.c [MPROTECT_VDB && !THREADS] (is_header_found_async): New
macro.
* os_dep.c [MPROTECT_VDB && !DARWIN] (GC_write_fault_handler): Use
is_header_found_async() instead of HDR(addr)!=0.
* os_dep.c [MPROTECT_VDB && DARWIN] (catch_exception_raise): Likewise.