Use byte padding to avoid fields re-ordering on JDK 15 and above

[lantalex]

Same thing as previous PR but for 3.4.x

On modern JDK (15+) good old trick using long fields as padding is no longer reliable. To prevent false-sharing we could use byte padding

[simonbasle]

fixes #2057.

[simonbasle]

@lantalex I tried to come up with a simple JOL-based test to validate the layout but it didn't fail even with original long padding in Java 17:

	@Test
	@Tag("slow")
	void objectPadding() {
		ClassLayout layout = ClassLayout.parseClass(SpscArrayQueue.class);

		AtomicLong currentPaddingSize = new AtomicLong();
		List<String> interestingFields = new ArrayList<>();

		layout.fields().forEach(f -> {
			if (f.name().startsWith("pad")) {
				currentPaddingSize.addAndGet(f.size());
			}
			else {
				if (currentPaddingSize.get() > 0) {
					interestingFields.add("padding of " + currentPaddingSize.getAndSet(0));
				}
				interestingFields.add(f.name());
			}
		});
		if (currentPaddingSize.get() > 0) {
			interestingFields.add("padding of " + currentPaddingSize.getAndSet(0));
		}

		assertThat(interestingFields).containsExactly(
			"array",
			"mask",
			"padding of 120",
			"producerIndex",
			"padding of 120",
			"consumerIndex",
			"padding of 120"
		);
	}

Compare to a similar test in another class that has padding (and could be also modified in your PR btw), QueueDrainSubscriber:

	@Test
	@Tag("slow")
	void objectPadding2() {
		ClassLayout layout = ClassLayout.parseClass(QueueDrainSubscriber.class);

		AtomicLong currentPaddingSize = new AtomicLong();
		List<String> interestingFields = new ArrayList<>();

		layout.fields().forEach(f -> {
			if (f.name().startsWith("pad")) {
				currentPaddingSize.addAndGet(f.size());
			}
			else {
				if (currentPaddingSize.get() > 0) {
					interestingFields.add("padding of " + currentPaddingSize.getAndSet(0));
				}
				interestingFields.add(f.name());
			}
		});
		if (currentPaddingSize.get() > 0) {
			interestingFields.add("padding of " + currentPaddingSize.getAndSet(0));
		}

		assertThat(interestingFields).containsExactly(
			"padding of 120",
			"wip",
			"padding of 120",
			"requested",
			"padding of 120",
			"cancelled",
			"done",
			"actual",
			"queue",
			"error"
		);
	}

which fails on Java 17, having this layout detected instead:

"wip",
        "padding of 240",
        "requested",
        "padding of 120",
        "cancelled",
        "done",
        "actual",
        "queue",
        "error"

Note that in both cases I've renamed the padding fields for better identification to all use pad prefix rather than q/p/b

[simonbasle]

looking like a good start, but see my comment in PR discussion.

in addition, I would:

• use pad prefix instead of b
• add the objectLayout() test (see PR comment, needs edition for 128 padding instead of 120)
• make a similar change in QueueDrainSubscriber (using pad prefix too)
• replace the QueueDrainSubscriber#objectLayout() test with the simplified one (see objectLayout2() in my PR comment)

[lantalex]

@simonbasle
Should I revert my changes for SpscArrayQueue? Looks like long padding still works for this class, test for object layout was added (passing both for JDK17/JDK11)

[simonbasle]

@lantalex nah, I'd say let's keep it that way, just in case it becomes the inspiration for a future padded class 😜

[simonbasle]

nice ! I wouldn't filter on the paddingSizes assertions but otherwise LGTM

[lantalex]

Hello @simonbasle !
PR was updated. I changed a little bit SpscArrayQueueP1 padding (added extra 4 bytes) to make SpscArrayQueue object fields layout consistent on modern (JDK15+) and classic versions of JDK

[reactorbot]

@simonbasle this PR seems to have been merged on a maintenance branch, please ensure the change is merge-forwarded to intermediate maintenance branches and up to main 🙇

[simonbasle]

thanks for the PR @lantalex 🎁 🙇

[lantalex]

thanks for the PR @lantalex gift bow

Thank you for review and fast feedback! 😀