RFR: 8355646: Optimize ObjectMonitor::exit

[Patricio Chilano Mateo]

On Mon, 28 Apr 2025 15:29:12 GMT, Fredrik Bredberg <fbredberg at openjdk.org> wrote:

> Today ObjectMonitor::exit always releases the lock before checking if it needs to wake up a waiting thread that is queued on the entry_list. In order to unpark a thread from the entry_list it needs to re-acquire the lock again.
> 
> If we (before releasing the lock) check if there is no assigned successor but there is a waiting thread in the entry_list we can immediately unpark the thread, thus saving us the trouble of releasing and re-acquiring the lock.
> 
> If there already is an assigned successor (spinning and wanting to acquire the lock) we still want to release the lock as soon as possible so that we don't lengthen the critical section thereby delaying any spinning threads from getting the lock in a high contention scenario.
> 
> This PR is first and foremost an optimization if the lock is lightly contended. It also removes some source reading confusion, as to why we release and then immediately re-acquire the lock.
> 
> When running performance tests we see that the number of test that has improved outnumber the tests that has not.
> 
> Passes tier1-7 with no added problems.

Looks reasonable to me too.

src/hotspot/share/runtime/objectMonitor.cpp line 1510:

> 1508: 
> 1509:   for (;;) {
> 1510:     assert(has_owner(current), "invariant");

Preexisting, but since we're touching this code, this assert seems redundant. We have already checked for this at the beginning of `exit()` and we check again at the end of the loop in case of reacquiring the lock.

-------------

PR Review: https://git.openjdk.org/jdk/pull/24927#pullrequestreview-2819406053
PR Review Comment: https://git.openjdk.org/jdk/pull/24927#discussion_r2076129622