libstdc++: Fix proxy wait detection in atomic wait

libstdc++: Fix proxy wait detection in atomic wait

A failed assertion was observed with std::atomic<bool>::wait when the
loop in __atomic_wait_address is entered and calls _M_setup_wait a
second time, after waking from __wait_impl. When the first call to
_M_setup_wait makes a call to _M_setup_proxy_wait that function decides
that a proxy wait is needed for an object of type bool, and it updates
the _M_obj and _M_obj_size members to refer to the futex in the proxy
state, instead of referring to the bool object itself. The next time
_M_setup_wait is called it calls _M_setup_proxy_wait again but now it
sees _M_obj_size == sizeof(futex) and so this time decides a proxy wait
is *not* needed, and then fails the __glibcxx_assert(_M_obj == addr)
check.

The problem is that _M_setup_proxy_wait wasn't correctly handling the
case where it's called a second time, after the decision to use a proxy
wait has already been made. That can be fixed in _M_setup_proxy_wait by
checking if _M_obj != addr, which implies that a proxy wait has already
been set up by a previous call. In that case, _M_setup_proxy_wait should
only update _M_old to the latest value of the proxy _M_ver.

This change means that _M_setup_proxy_wait is safe to call repeatedly
for a proxy wait, and will only update _M_wait_state, _M_obj, and
_M_obj_size on the first call. On the second and subsequent calls, those
variables are already correctly set for the proxy wait so don't need to
be set again.

For non-proxy waits, calling _M_setup_proxy_wait more than once is safe,
but pessimizes performance. The caller shouldn't make a second call to
_M_setup_proxy_wait because we don't need to check again if a proxy wait
should be used (the answer won't change) and we don't need to load a
value from the proxy _M_ver.

However, it was difficult to detect the case of a non-proxy wait,
because _M_setup_wait doesn't know if it's being called the first time
(when _M_setup_proxy_wait is called to make the initial decision) or a
subsequent time (in which case _M_obj == addr implies a non-proxy wait
was already decided on). As a result, _M_setup_proxy_wait was being used
every time to see if it's a proxy wait. We can resolve this by splitting
the _M_setup_wait function into _M_setup_wait and _M_on_wake, where the
former is only called once to do the initial setup and the latter is
called after __wait_impl returns, to prepare to check the predicate and
possibly wait again.  The new _M_on_wake function can avoid unnecessary
calls to _M_setup_proxy_wait by checking _M_obj == addr to identify a
non-proxy wait.

The three callers of _M_setup_wait are updated to use _M_on_wake instead
of _M_setup_wait after waking from a waiting function. This change
revealed a latent performance bug in __atomic_wait_address_for which was
not passing __res to _M_setup_wait, so a new value was always loaded
even when __res._M_has_val was true. By splitting _M_on_wake out of
_M_setup_wait this problem became more obvious, because we no longer
have _M_setup_wait doing two different jobs, depending on whether it was
passed the optional third argument or not.

libstdc++-v3/ChangeLog:

	* include/bits/atomic_timed_wait.h (__atomic_wait_address_until):
	Use _M_on_wake instead of _M_setup_wait after waking.
	(__atomic_wait_address_for): Likewise.
	* include/bits/atomic_wait.h (__atomic_wait_address): Likewise.
	(__wait_args::_M_setup_wait): Remove third parameter and move
	code to update _M_old to ...
	(__wait_args::_M_on_wake): New member function to update _M_old
	after waking, only calling _M_setup_proxy_wait if needed.
	(__wait_args::_M_store): New member function to update _M_old
	from a value, for non-proxy waits.
	* src/c++20/atomic.cc (__wait_args::_M_setup_proxy_wait): If
	_M_obj is not addr, only load a new value and return true.

Reviewed-by: Tomasz Kamiński <tkaminsk@redhat.com>