Performance overhead of DebugProbesImpl for non-concurrent coroutines

[qwwdfsad]

Snippet to reproduce:

class DebugProbesTest : DebugTestBase() {
    fun <Node> generateRecursiveSequence(initialSequence: Sequence<Node>, children: (Node) -> Sequence<Node>): Sequence<Node> {
        return sequence {
            val initialIterator = initialSequence.iterator()
            if (!initialIterator.hasNext()) {
                return@sequence
            }
            val visited = HashSet<Node>()
            val sequences = ArrayDeque<Sequence<Node>>()
            sequences.addLast(initialIterator.asSequence())
            while (sequences.isNotEmpty()) {
                val currentSequence = sequences.removeFirst()
                for (node in currentSequence) {
                    if (visited.add(node)) {
                        yield(node)
                        sequences.addLast(children(node))
                    }
                }
            }
        }
    }

    @Volatile
    var a = 2
    @Test
    fun stressGenerateRecursive() {
        while (true) {
            runBlocking {
                repeat(8) {
                    launch(Dispatchers.Default) {
                        val seq = generateRecursiveSequence((1..100).asSequence()) {
                            (1..it).asSequence()
                        }

                        for (i in seq) {
                            a = i
                        }
                    }
                }
            }
        }
    }
}

Profile:

Up to 30% is occupied by updateState

[qwwdfsad]

It seems like a ReentrantLock is a rudiment from pre-concurrent implementation times.
It indeed incurs some non-trivial overhead, while providing only a single benefit -- strongly consistent snapshot in read operations (dump/hierarchyToString etc.). The guarantee seems to be way too strong for such an overhead, and the proposed solution is just to remove this lock.

The trick here is to ensure that we do have a proper ordered access to memory locations everywhere

[maxmedvedev]

Hey, when is this fix going to be released?

[qwwdfsad]

It's already been released as part of 1.7.0-Beta