为什么Java中的“或”慢于“和”？

Question

今天，我在java中遇到了相当惊人的行为， or慢于and ！

我甚至做了一个你可以在下面看到的测试用例。现在我想知道为什么会这样？ 我做错了什么或者只是在我的电脑上发生了什么？ 我看不出有任何理由， or应慢and特意用这个显著差异。 我想用其他一些语言测试这种现象，你对这个一般有什么想法吗？

public class TestClass {

    public static void main(String[] args) {

        long[] or = new long[10];
        long[] and = new long[10];
        long lStartTime, lEndTime, difference = 0;
        for (int idx = 0; idx < 10; idx++) {

            lStartTime = System.nanoTime();
            for (int i= 0; i < 1000000000; i++) { 
                int j = i | i+1 ;
            }
            lEndTime = System.nanoTime();
            difference = lEndTime - lStartTime;
            System.out.println("Elapsed milliseconds: " + difference/1000000);
            or[idx] = difference;

            lStartTime = System.nanoTime();
            for (int i= 0; i < 1000000000; i++) {
                int j = i & i+1 ;
            }
            lEndTime = System.nanoTime();
            difference = lEndTime - lStartTime;
            System.out.println("Elapsed milliseconds: " + difference/1000000);  
            and[idx] = difference;
            System.out.println("------------------------------------" );

        }

        long tmp = 0;
        for (long l : or) {
            tmp += l;
        }
        tmp /= 10;
        System.out.println("Elapsed milliseconds for or: " + tmp/1000000);
        tmp = 0;
        for (long l : and) {
            tmp += l;
        }
        tmp /= 10;
        System.out.println("Elapsed milliseconds for and: " + tmp/1000000);
    }
}

结果：

Elapsed milliseconds: 1600
Elapsed milliseconds: 1332
------------------------------------
Elapsed milliseconds: 1609
Elapsed milliseconds: 1335
------------------------------------
Elapsed milliseconds: 1609
Elapsed milliseconds: 1335
------------------------------------
Elapsed milliseconds: 1542
Elapsed milliseconds: 1314
------------------------------------
Elapsed milliseconds: 1705
Elapsed milliseconds: 1324
------------------------------------
Elapsed milliseconds: 1559
Elapsed milliseconds: 1315
------------------------------------
Elapsed milliseconds: 1526
Elapsed milliseconds: 1314
------------------------------------
Elapsed milliseconds: 1568
Elapsed milliseconds: 1340
------------------------------------
Elapsed milliseconds: 1551
Elapsed milliseconds: 1318
------------------------------------
Elapsed milliseconds: 1574
Elapsed milliseconds: 1321
------------------------------------
Elapsed milliseconds for or: 1584
Elapsed milliseconds for and: 1325

Answer 1

Bit-Wise OR 并不比Bit-Wise AND慢！

在代码中的时间测量片段之间切换，您将得到相反的结果：

for (int idx = 0; idx < 10; idx++) {

    lStartTime = System.nanoTime();
    for (int i= 0; i < 1000000000; i++) {
        int j = i & i+1 ;
    }
    lEndTime = System.nanoTime();

    difference = lEndTime - lStartTime;
    System.out.println("Elapsed milliseconds: " + difference/1000000);  
    and[idx] = difference;

    lStartTime = System.nanoTime();
    for (int i= 0; i < 1000000000; i++) { 
        int j = i | i+1 ;
    }
    lEndTime = System.nanoTime();

    difference = lEndTime - lStartTime;
    System.out.println("Elapsed milliseconds: " + difference/1000000);
    or[idx] = difference;

    System.out.println("------------------------------------");
}

我倾向于猜测JVM在第二个片段中应用了某种运行时优化。

Answer 2

编写正确的微基准测试非常耗时且容易出错。 我建议使用像Caliper这样的现有库。

这是在Caliper中完成的相应基准测试（编译时需要的最新git版本）：

public class BitwiseOperatorPerformance {
    @Benchmark
    public int timeOr(int reps){
        int dummy = 0;
        for (int i = 0; i < reps; i++) {
            dummy |= i+1;
        }
        return dummy;
    }
    @Benchmark
    public int timeAnd(int reps){
        int dummy = 0;
        for (int i = 0; i < reps; i++) {
            dummy &= i+1;
        }
        return dummy;
    }
}

这是测试结果： 链接

结果表明， AND和OR运算符的性能完全相同 。