Question

基本上这是在试图回答另一个问题时出现的。假设这段代码：

AtomicInteger i = new AtomicInteger(0);
AtomicInteger count = new AtomicInteger(0);
IntStream.generate(() -> i.incrementAndGet())
        .parallel()
        .peek(x -> count.incrementAndGet())
        .limit(5)
        .forEach(System.out::println);

System.out.println("count = " + count);

我理解IntStream#generate是无序无限流的事实，并且要完成它必须有一个短路操作（在这种情况下为limit）。我也理解Supplier可以自由地调用Stream实现在达到该限制之前的次数。

在java-8下运行，将始终count打印512（可能并非总是如此，但我的机器上也是如此）。

在对比运行中，这在java-10下很少超过5。所以我的问题是内部发生了什么变化，短路发生得更好（我试图通过拥有源代码并尝试做一些差异来自己解决这个问题......）

Answer 1

这种变化发生在Java 9，beta 103和Java 9，beta 120（JDK‑8154387）之间。

负责的班级是StreamSpliterators.UnorderedSliceSpliterator.OfInt，分别是。它的超级StreamSpliterators.UnorderedSliceSpliterator。

该类的旧版本看起来像

abstract static class UnorderedSliceSpliterator<T, T_SPLITR extends Spliterator<T>> {
    static final int CHUNK_SIZE = 1 << 7;

    // The spliterator to slice
    protected final T_SPLITR s;
    protected final boolean unlimited;
    private final long skipThreshold;
    private final AtomicLong permits;

    UnorderedSliceSpliterator(T_SPLITR s, long skip, long limit) {
        this.s = s;
        this.unlimited = limit < 0;
        this.skipThreshold = limit >= 0 ? limit : 0;
        this.permits = new AtomicLong(limit >= 0 ? skip + limit : skip);
    }

    UnorderedSliceSpliterator(T_SPLITR s,
                              UnorderedSliceSpliterator<T, T_SPLITR> parent) {
        this.s = s;
        this.unlimited = parent.unlimited;
        this.permits = parent.permits;
        this.skipThreshold = parent.skipThreshold;
    }

...

        @Override
        public void forEachRemaining(Consumer<? super T> action) {
            Objects.requireNonNull(action);

            ArrayBuffer.OfRef<T> sb = null;
            PermitStatus permitStatus;
            while ((permitStatus = permitStatus()) != PermitStatus.NO_MORE) {
                if (permitStatus == PermitStatus.MAYBE_MORE) {
                    // Optimistically traverse elements up to a threshold of CHUNK_SIZE
                    if (sb == null)
                        sb = new ArrayBuffer.OfRef<>(CHUNK_SIZE);
                    else
                        sb.reset();
                    long permitsRequested = 0;
                    do { } while (s.tryAdvance(sb) && ++permitsRequested < CHUNK_SIZE);
                    if (permitsRequested == 0)
                        return;
                    sb.forEach(action, acquirePermits(permitsRequested));
                }
                else {
                    // Must be UNLIMITED; let 'er rip
                    s.forEachRemaining(action);
                    return;
                }
            }
        }

正如我们所看到的，它试图在每个分裂器中缓冲最多CHUNK_SIZE = 1 << 7个元素，这可能最终达到“CPU核心数”×128个元素。

相比之下，新版本看起来像

abstract static class UnorderedSliceSpliterator<T, T_SPLITR extends Spliterator<T>> {
    static final int CHUNK_SIZE = 1 << 7;

    // The spliterator to slice
    protected final T_SPLITR s;
    protected final boolean unlimited;
    protected final int chunkSize;
    private final long skipThreshold;
    private final AtomicLong permits;

    UnorderedSliceSpliterator(T_SPLITR s, long skip, long limit) {
        this.s = s;
        this.unlimited = limit < 0;
        this.skipThreshold = limit >= 0 ? limit : 0;
        this.chunkSize = limit >= 0 ? (int)Math.min(CHUNK_SIZE,
            ((skip + limit) / AbstractTask.LEAF_TARGET) + 1) : CHUNK_SIZE;
        this.permits = new AtomicLong(limit >= 0 ? skip + limit : skip);
    }

    UnorderedSliceSpliterator(T_SPLITR s,
                              UnorderedSliceSpliterator<T, T_SPLITR> parent) {
        this.s = s;
        this.unlimited = parent.unlimited;
        this.permits = parent.permits;
        this.skipThreshold = parent.skipThreshold;
        this.chunkSize = parent.chunkSize;
    }

...

        @Override
        public void forEachRemaining(Consumer<? super T> action) {
            Objects.requireNonNull(action);

            ArrayBuffer.OfRef<T> sb = null;
            PermitStatus permitStatus;
            while ((permitStatus = permitStatus()) != PermitStatus.NO_MORE) {
                if (permitStatus == PermitStatus.MAYBE_MORE) {
                    // Optimistically traverse elements up to a threshold of chunkSize
                    if (sb == null)
                        sb = new ArrayBuffer.OfRef<>(chunkSize);
                    else
                        sb.reset();
                    long permitsRequested = 0;
                    do { } while (s.tryAdvance(sb) && ++permitsRequested < chunkSize);
                    if (permitsRequested == 0)
                        return;
                    sb.forEach(action, acquirePermits(permitsRequested));
                }
                else {
                    // Must be UNLIMITED; let 'er rip
                    s.forEachRemaining(action);
                    return;
                }
            }
        }

现在有一个实例字段chunkSize。当存在定义的限制且表达式((skip + limit) / AbstractTask.LEAF_TARGET) + 1的计算值小于CHUNK_SIZE时，将使用该较小的值。因此，当限制较小时，chunkSize会小得多。对于限制为5的情况，块大小始终为1。

限制和无序流的内部更改

1 个答案: