我写了自己的分支预测实验：

public class BranchPrediction {
    public static void main(final String[] args) {
        long start;
        long sum = 0;
 
        /* No branch */
        start = System.nanoTime();
        sum = 0;
        for (long i = 0; i < 10000000000L; ++i)
            sum += i;
        System.out.println(System.nanoTime() - start);
        System.out.println(sum);
 
        /* With branch */
        start = System.nanoTime();
        sum = 0;
        for (long i = 0; i < 10000000000L; ++i)
            if (i >= 0)
                sum += i;
        System.out.println(System.nanoTime() - start);
        System.out.println(sum);
 
        /* No branch (again) */
        start = System.nanoTime();
        sum = 0;
        for (long i = 0; i < 10000000000L; ++i)
            sum += i;
        System.out.println(System.nanoTime() - start);
        System.out.println(sum);
 
        /* With branch (again) */
        start = System.nanoTime();
        sum = 0;
        for (long i = 0; i < 10000000000L; ++i)
            if (i >= 0)
                sum += i;
        System.out.println(System.nanoTime() - start);
        System.out.println(sum);
    }
}

结果让我感到困惑：根据程序输出,带分支的循环比没有分支循环的循环可靠得快.

输出示例：

7949691477
-5340232226128654848
6947699555
-5340232226128654848
7920972795
-5340232226128654848
7055459799
-5340232226128654848

为什么会这样？

编辑：

>反汇编类显示Java编译器没有优化(遗漏)任何东西(https://gist.github.com/HouzuoGuo/5692424)
> Why is it faster to process a sorted array than an unsorted array?作者使用的Java基准技术与我的相同.
>该机器是Intel核心i7,运行Linux 3.2 64位和Oracle JVM 1.7 64位
>当我超过循环迭代的次数时,with-branch循环比非分支循环运行多SECONDS.