System.arraycopy 长度不变

System.arraycopy with constant length

我正在和 JMH (http://openjdk.java.net/projects/code-tools/jmh/) 一起玩,我偶然发现了一个奇怪的结果。

我正在对制作数组浅表副本的方法进行基准测试,我可以观察到预期的结果(遍历数组是个坏主意,#clone()System#arraycopy()Arrays#copyOf(),性能方面)。

除了 System#arraycopy() 当数组的长度被硬编码时会慢四分之一...等等,什么?这怎么能慢?

有谁知道可能是什么原因吗?

结果(吞吐量):

# JMH 1.11 (released 17 days ago)
# VM version: JDK 1.8.0_05, VM 25.5-b02
# VM invoker: /Library/Java/JavaVirtualMachines/jdk1.8.0_05.jdk/Contents/Home/jre/bin/java
# VM options: -Dfile.encoding=UTF-8 -Duser.country=FR -Duser.language=fr -Duser.variant
# Warmup: 20 iterations, 1 s each
# Measurement: 20 iterations, 1 s each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time

Benchmark                                            Mode  Cnt         Score         Error  Units
ArrayCopyBenchmark.ArraysCopyOf                     thrpt   20  67100500,319 ±  455252,537  ops/s
ArrayCopyBenchmark.ArraysCopyOf_Class               thrpt   20  65246374,290 ±  976481,330  ops/s
ArrayCopyBenchmark.ArraysCopyOf_Class_ConstantSize  thrpt   20  65068143,162 ± 1597390,531  ops/s
ArrayCopyBenchmark.ArraysCopyOf_ConstantSize        thrpt   20  64463603,462 ±  953946,811  ops/s
ArrayCopyBenchmark.Clone                            thrpt   20  64837239,393 ±  834353,404  ops/s
ArrayCopyBenchmark.Loop                             thrpt   20  21070422,097 ±  112595,764  ops/s
ArrayCopyBenchmark.Loop_ConstantSize                thrpt   20  24458867,274 ±  181486,291  ops/s
ArrayCopyBenchmark.SystemArrayCopy                  thrpt   20  66688368,490 ±  582416,954  ops/s
ArrayCopyBenchmark.SystemArrayCopy_ConstantSize     thrpt   20  48992312,357 ±  298807,039  ops/s

和基准 class:

import java.util.Arrays;
import java.util.concurrent.TimeUnit;

import org.openjdk.jmh.annotations.Benchmark;
import org.openjdk.jmh.annotations.BenchmarkMode;
import org.openjdk.jmh.annotations.Mode;
import org.openjdk.jmh.annotations.OutputTimeUnit;
import org.openjdk.jmh.annotations.Scope;
import org.openjdk.jmh.annotations.Setup;
import org.openjdk.jmh.annotations.State;

@State(Scope.Benchmark)
@BenchmarkMode(Mode.Throughput)
@OutputTimeUnit(TimeUnit.SECONDS)
public class ArrayCopyBenchmark {

    private static final int LENGTH = 32;

    private Object[] array;

    @Setup
    public void before() {
        array = new Object[LENGTH];
        for (int i = 0; i < LENGTH; i++) {
            array[i] = new Object();
        }
    }

    @Benchmark
    public Object[] Clone() {
        Object[] src = this.array;
        return src.clone();
    }

    @Benchmark
    public Object[] ArraysCopyOf() {
        Object[] src = this.array;
        return Arrays.copyOf(src, src.length);
    }

    @Benchmark
    public Object[] ArraysCopyOf_ConstantSize() {
        Object[] src = this.array;
        return Arrays.copyOf(src, LENGTH);
    }

    @Benchmark
    public Object[] ArraysCopyOf_Class() {
        Object[] src = this.array;
        return Arrays.copyOf(src, src.length, Object[].class);
    }

    @Benchmark
    public Object[] ArraysCopyOf_Class_ConstantSize() {
        Object[] src = this.array;
        return Arrays.copyOf(src, LENGTH, Object[].class);
    }

    @Benchmark
    public Object[] SystemArrayCopy() {
        Object[] src = this.array;
        int length = src.length;
        Object[] array = new Object[length];
        System.arraycopy(src, 0, array, 0, length);
        return array;
    }

    @Benchmark
    public Object[] SystemArrayCopy_ConstantSize() {
        Object[] src = this.array;
        Object[] array = new Object[LENGTH];
        System.arraycopy(src, 0, array, 0, LENGTH);
        return array;
    }

    @Benchmark
    public Object[] Loop() {
        Object[] src = this.array;
        int length = src.length;
        Object[] array = new Object[length];
        for (int i = 0; i < length; i++) {
            array[i] = src[i];
        }
        return array;
    }

    @Benchmark
    public Object[] Loop_ConstantSize() {
        Object[] src = this.array;
        Object[] array = new Object[LENGTH];
        for (int i = 0; i < LENGTH; i++) {
            array[i] = src[i];
        }
        return array;
    }
}

像往常一样,通过研究生成的代码可以快速回答此类问题。 JMH 在 Linux 上为您提供 -prof perfasm,在 Windows 上提供 -prof xperfasm。如果你 运行 JDK 8u40 上的基准测试,那么你会看到(注意我使用 -bm avgt -tu ns 使分数更容易理解):

Benchmark                         Mode  Cnt   Score   Error  Units
ACB.SystemArrayCopy               avgt   25  13.294 ± 0.052  ns/op
ACB.SystemArrayCopy_ConstantSize  avgt   25  16.413 ± 0.080  ns/op

为什么这些基准测试的表现不同?我们先做-prof perfnorm来剖析(我去掉了无关紧要的行):

Benchmark                                     Mode  Cnt    Score    Error  Units
ACB.SAC                                       avgt   25   13.466 ±  0.070  ns/op
ACB.SAC:·CPI                                  avgt    5    0.602 ±  0.025   #/op
ACB.SAC:·L1-dcache-load-misses                avgt    5    2.346 ±  0.239   #/op
ACB.SAC:·L1-dcache-loads                      avgt    5   24.756 ±  1.438   #/op
ACB.SAC:·L1-dcache-store-misses               avgt    5    2.404 ±  0.129   #/op
ACB.SAC:·L1-dcache-stores                     avgt    5   14.929 ±  0.230   #/op
ACB.SAC:·LLC-loads                            avgt    5    2.151 ±  0.217   #/op
ACB.SAC:·branches                             avgt    5   17.795 ±  1.003   #/op
ACB.SAC:·cycles                               avgt    5   56.677 ±  3.187   #/op
ACB.SAC:·instructions                         avgt    5   94.145 ±  6.442   #/op

ACB.SAC_ConstantSize                          avgt   25   16.447 ±  0.084  ns/op
ACB.SAC_ConstantSize:·CPI                     avgt    5    0.637 ±  0.016   #/op
ACB.SAC_ConstantSize:·L1-dcache-load-misses   avgt    5    2.357 ±  0.206   #/op
ACB.SAC_ConstantSize:·L1-dcache-loads         avgt    5   25.611 ±  1.482   #/op
ACB.SAC_ConstantSize:·L1-dcache-store-misses  avgt    5    2.368 ±  0.123   #/op
ACB.SAC_ConstantSize:·L1-dcache-stores        avgt    5   25.593 ±  1.610   #/op
ACB.SAC_ConstantSize:·LLC-loads               avgt    5    1.050 ±  0.038   #/op
ACB.SAC_ConstantSize:·branches                avgt    5   17.853 ±  0.697   #/op
ACB.SAC_ConstantSize:·cycles                  avgt    5   66.680 ±  2.049   #/op
ACB.SAC_ConstantSize:·instructions            avgt    5  104.759 ±  4.831   #/op

所以,ConstantSize 以某种方式做了更多的 L1-dcache-stores,但少了一个 LLC-load。嗯,这就是我们正在寻找的,在常量情况下有更多的商店。 -prof perfasm 方便地突出显示装配中的热点:

default:

  4.32%    6.36%   0x00007f7714bda2dc: movq   [=12=]x1,(%rax)            ; alloc
  0.09%    0.04%   0x00007f7714bda2e3: prefetchnta 0x100(%r9)
  2.95%    1.48%   0x00007f7714bda2eb: movl   [=12=]xf80022a9,0x8(%rax)
  0.38%    0.18%   0x00007f7714bda2f2: mov    %r11d,0xc(%rax)
  1.56%    3.02%   0x00007f7714bda2f6: prefetchnta 0x140(%r9)
  4.73%    2.71%   0x00007f7714bda2fe: prefetchnta 0x180(%r9)

ConstantSize:

  0.58%    1.22%   0x00007facf921132b: movq   [=13=]x1,(%r14)            ; alloc
  0.84%    0.72%   0x00007facf9211332: prefetchnta 0xc0(%r10)
  0.11%    0.13%   0x00007facf921133a: movl   [=13=]xf80022a9,0x8(%r14)
  0.21%    0.68%   0x00007facf9211342: prefetchnta 0x100(%r10)
  0.50%    0.87%   0x00007facf921134a: movl   [=13=]x20,0xc(%r14)
  0.53%    0.82%   0x00007facf9211352: mov    [=13=]x10,%ecx
  0.04%    0.14%   0x00007facf9211357: xor    %rax,%rax
  0.34%    0.76%   0x00007facf921135a: shl    [=13=]x3,%rcx
  0.50%    1.17%   0x00007facf921135e: rex.W rep stos %al,%es:(%rdi) ; zeroing
 29.49%   52.09%   0x00007facf9211361: prefetchnta 0x140(%r10)
  1.03%    0.53%   0x00007facf9211369: prefetchnta 0x180(%r10)  

所以那个讨厌的 rex.W rep stos %al,%es:(%rdi) 消耗了大量时间。这会将新分配的数组归零。在 ConstantSize 测试中,JVM 无法关联您正在覆盖整个目标数组,因此它必须在进入实际数组副本之前将其预置零。

如果您查看 JDK 9b82(最新的可用代码)上生成的代码,那么您会看到它在非归零副本中折叠了两种模式,正如您在 -prof perfasm 中看到的那样,也可以用 -prof perfnorm:

确认
Benchmark                                     Mode  Cnt    Score    Error  Units
ACB.SAC                                       avgt   50   14.156 ±  0.492  ns/op
ACB.SAC:·CPI                                  avgt    5    0.612 ±  0.144   #/op
ACB.SAC:·L1-dcache-load-misses                avgt    5    2.363 ±  0.341   #/op
ACB.SAC:·L1-dcache-loads                      avgt    5   28.350 ±  2.181   #/op
ACB.SAC:·L1-dcache-store-misses               avgt    5    2.287 ±  0.607   #/op
ACB.SAC:·L1-dcache-stores                     avgt    5   16.922 ±  3.402   #/op
ACB.SAC:·branches                             avgt    5   21.242 ±  5.914   #/op
ACB.SAC:·cycles                               avgt    5   67.168 ± 20.950   #/op
ACB.SAC:·instructions                         avgt    5  109.931 ± 35.905   #/op

ACB.SAC_ConstantSize                          avgt   50   13.763 ±  0.067  ns/op
ACB.SAC_ConstantSize:·CPI                     avgt    5    0.625 ±  0.024   #/op
ACB.SAC_ConstantSize:·L1-dcache-load-misses   avgt    5    2.376 ±  0.214   #/op
ACB.SAC_ConstantSize:·L1-dcache-loads         avgt    5   28.285 ±  2.127   #/op
ACB.SAC_ConstantSize:·L1-dcache-store-misses  avgt    5    2.335 ±  0.223   #/op
ACB.SAC_ConstantSize:·L1-dcache-stores        avgt    5   16.926 ±  1.467   #/op
ACB.SAC_ConstantSize:·branches                avgt    5   19.469 ±  0.869   #/op
ACB.SAC_ConstantSize:·cycles                  avgt    5   62.395 ±  3.898   #/op
ACB.SAC_ConstantSize:·instructions            avgt    5   99.891 ±  5.435   #/op

当然,所有这些用于数组复制的纳米基准都容易受到向量化复制存根中由奇怪的对齐引起的性能差异的影响,但这是另一个(恐怖的)故事,我没有勇气说出来。