下面的Java程序平均需要0.50到0.55秒的时间来运行:

public static void main(String[] args) {
    long startTime = System.nanoTime();
    int n = 0;
    for (int i = 0; i < 1000000000; i++) {
        n += 2 * (i * i);
    }
    System.out.println(
        (double) (System.nanoTime() - startTime) / 1000000000 + " s");
    System.out.println("n = " + n);
}

如果我将2 * (I * I)替换为2 * I * I,它将花费0.60到0.65秒的时间运行。如何来吗?

我把程序的每个版本都运行了15次,在两者之间交替运行。以下是调查结果:

 2*(i*i)  │  2*i*i
──────────┼──────────
0.5183738 │ 0.6246434
0.5298337 │ 0.6049722
0.5308647 │ 0.6603363
0.5133458 │ 0.6243328
0.5003011 │ 0.6541802
0.5366181 │ 0.6312638
0.515149  │ 0.6241105
0.5237389 │ 0.627815
0.5249942 │ 0.6114252
0.5641624 │ 0.6781033
0.538412  │ 0.6393969
0.5466744 │ 0.6608845
0.531159  │ 0.6201077
0.5048032 │ 0.6511559
0.5232789 │ 0.6544526

2 * i * i的最快运行时间比2 * (i * i)的最慢运行时间长。如果它们具有相同的效率,发生这种情况的概率将小于1/2^15 * 100% = 0.00305%。


当前回答

Kasperd在对公认答案的评论中问道:

Java和C示例使用了完全不同的寄存器名称。这两个例子都使用AMD64 ISA?

xor edx, edx
xor eax, eax
.L2:
mov ecx, edx
imul ecx, edx
add edx, 1
lea eax, [rax+rcx*2]
cmp edx, 1000000000
jne .L2

我没有足够的声誉在评论中回答这个问题,但这些都是相同的ISA。值得指出的是,GCC版本使用32位整数逻辑,而JVM编译版本内部使用64位整数逻辑。

R8 to R15 are just new X86_64 registers. EAX to EDX are the lower parts of the RAX to RDX general purpose registers. The important part in the answer is that the GCC version is not unrolled. It simply executes one round of the loop per actual machine code loop. While the JVM version has 16 rounds of the loop in one physical loop (based on rustyx answer, I did not reinterpret the assembly). This is one of the reasons why there are more registers being used since the loop body is actually 16 times longer.

其他回答

虽然与问题的环境没有直接关系,但出于好奇,我在。net Core 2.1 x64发布模式上做了同样的测试。

这是一个有趣的结果,证实了类似的现象(相反)发生在原力的黑暗面。代码:

static void Main(string[] args)
{
    Stopwatch watch = new Stopwatch();

    Console.WriteLine("2 * (i * i)");

    for (int a = 0; a < 10; a++)
    {
        int n = 0;

        watch.Restart();

        for (int i = 0; i < 1000000000; i++)
        {
            n += 2 * (i * i);
        }

        watch.Stop();

        Console.WriteLine($"result:{n}, {watch.ElapsedMilliseconds} ms");
    }

    Console.WriteLine();
    Console.WriteLine("2 * i * i");

    for (int a = 0; a < 10; a++)
    {
        int n = 0;

        watch.Restart();

        for (int i = 0; i < 1000000000; i++)
        {
            n += 2 * i * i;
        }

        watch.Stop();

        Console.WriteLine($"result:{n}, {watch.ElapsedMilliseconds}ms");
    }
}

结果:

2 * (i * i)

结果:119860736,438 ms 结果:119860736,433 ms 结果:119860736,437 ms 结果:119860736,435毫秒 结果:119860736,436 ms 结果:119860736,435毫秒 结果:119860736,435毫秒 结果:119860736,439 ms 结果:119860736,436 ms 结果:119860736,437 ms

2 * I * I

结果:119860736,417毫秒 结果:119860736,417毫秒 结果:119860736,417毫秒 结果:119860736,418 ms 结果:119860736,418 ms 结果:119860736,417毫秒 结果:119860736,418 ms 结果:119860736,416毫秒 结果:119860736,417毫秒 结果:119860736,418 ms

字节码的顺序略有不同。

2 * (i * i):

     iconst_2
     iload0
     iload0
     imul
     imul
     iadd

Vs 2 * I * I:

     iconst_2
     iload0
     imul
     iload0
     imul
     iadd

乍一看,这没什么区别;第二个版本更优,因为它少使用了一个插槽。

因此,我们需要更深入地挖掘较低级别(JIT)1。

请记住,JIT倾向于非常积极地展开小循环。事实上,我们观察到2 * (i * i)情况下的16倍展开:

030   B2: # B2 B3 <- B1 B2  Loop: B2-B2 inner main of N18 Freq: 1e+006
030     addl    R11, RBP    # int
033     movl    RBP, R13    # spill
036     addl    RBP, #14    # int
039     imull   RBP, RBP    # int
03c     movl    R9, R13 # spill
03f     addl    R9, #13 # int
043     imull   R9, R9  # int
047     sall    RBP, #1
049     sall    R9, #1
04c     movl    R8, R13 # spill
04f     addl    R8, #15 # int
053     movl    R10, R8 # spill
056     movdl   XMM1, R8    # spill
05b     imull   R10, R8 # int
05f     movl    R8, R13 # spill
062     addl    R8, #12 # int
066     imull   R8, R8  # int
06a     sall    R10, #1
06d     movl    [rsp + #32], R10    # spill
072     sall    R8, #1
075     movl    RBX, R13    # spill
078     addl    RBX, #11    # int
07b     imull   RBX, RBX    # int
07e     movl    RCX, R13    # spill
081     addl    RCX, #10    # int
084     imull   RCX, RCX    # int
087     sall    RBX, #1
089     sall    RCX, #1
08b     movl    RDX, R13    # spill
08e     addl    RDX, #8 # int
091     imull   RDX, RDX    # int
094     movl    RDI, R13    # spill
097     addl    RDI, #7 # int
09a     imull   RDI, RDI    # int
09d     sall    RDX, #1
09f     sall    RDI, #1
0a1     movl    RAX, R13    # spill
0a4     addl    RAX, #6 # int
0a7     imull   RAX, RAX    # int
0aa     movl    RSI, R13    # spill
0ad     addl    RSI, #4 # int
0b0     imull   RSI, RSI    # int
0b3     sall    RAX, #1
0b5     sall    RSI, #1
0b7     movl    R10, R13    # spill
0ba     addl    R10, #2 # int
0be     imull   R10, R10    # int
0c2     movl    R14, R13    # spill
0c5     incl    R14 # int
0c8     imull   R14, R14    # int
0cc     sall    R10, #1
0cf     sall    R14, #1
0d2     addl    R14, R11    # int
0d5     addl    R14, R10    # int
0d8     movl    R10, R13    # spill
0db     addl    R10, #3 # int
0df     imull   R10, R10    # int
0e3     movl    R11, R13    # spill
0e6     addl    R11, #5 # int
0ea     imull   R11, R11    # int
0ee     sall    R10, #1
0f1     addl    R10, R14    # int
0f4     addl    R10, RSI    # int
0f7     sall    R11, #1
0fa     addl    R11, R10    # int
0fd     addl    R11, RAX    # int
100     addl    R11, RDI    # int
103     addl    R11, RDX    # int
106     movl    R10, R13    # spill
109     addl    R10, #9 # int
10d     imull   R10, R10    # int
111     sall    R10, #1
114     addl    R10, R11    # int
117     addl    R10, RCX    # int
11a     addl    R10, RBX    # int
11d     addl    R10, R8 # int
120     addl    R9, R10 # int
123     addl    RBP, R9 # int
126     addl    RBP, [RSP + #32 (32-bit)]   # int
12a     addl    R13, #16    # int
12e     movl    R11, R13    # spill
131     imull   R11, R13    # int
135     sall    R11, #1
138     cmpl    R13, #999999985
13f     jl     B2   # loop end  P=1.000000 C=6554623.000000

我们看到有1个寄存器被“溢出”到堆栈上。

对于2 * i * i版本:

05a   B3: # B2 B4 <- B1 B2  Loop: B3-B2 inner main of N18 Freq: 1e+006
05a     addl    RBX, R11    # int
05d     movl    [rsp + #32], RBX    # spill
061     movl    R11, R8 # spill
064     addl    R11, #15    # int
068     movl    [rsp + #36], R11    # spill
06d     movl    R11, R8 # spill
070     addl    R11, #14    # int
074     movl    R10, R9 # spill
077     addl    R10, #16    # int
07b     movdl   XMM2, R10   # spill
080     movl    RCX, R9 # spill
083     addl    RCX, #14    # int
086     movdl   XMM1, RCX   # spill
08a     movl    R10, R9 # spill
08d     addl    R10, #12    # int
091     movdl   XMM4, R10   # spill
096     movl    RCX, R9 # spill
099     addl    RCX, #10    # int
09c     movdl   XMM6, RCX   # spill
0a0     movl    RBX, R9 # spill
0a3     addl    RBX, #8 # int
0a6     movl    RCX, R9 # spill
0a9     addl    RCX, #6 # int
0ac     movl    RDX, R9 # spill
0af     addl    RDX, #4 # int
0b2     addl    R9, #2  # int
0b6     movl    R10, R14    # spill
0b9     addl    R10, #22    # int
0bd     movdl   XMM3, R10   # spill
0c2     movl    RDI, R14    # spill
0c5     addl    RDI, #20    # int
0c8     movl    RAX, R14    # spill
0cb     addl    RAX, #32    # int
0ce     movl    RSI, R14    # spill
0d1     addl    RSI, #18    # int
0d4     movl    R13, R14    # spill
0d7     addl    R13, #24    # int
0db     movl    R10, R14    # spill
0de     addl    R10, #26    # int
0e2     movl    [rsp + #40], R10    # spill
0e7     movl    RBP, R14    # spill
0ea     addl    RBP, #28    # int
0ed     imull   RBP, R11    # int
0f1     addl    R14, #30    # int
0f5     imull   R14, [RSP + #36 (32-bit)]   # int
0fb     movl    R10, R8 # spill
0fe     addl    R10, #11    # int
102     movdl   R11, XMM3   # spill
107     imull   R11, R10    # int
10b     movl    [rsp + #44], R11    # spill
110     movl    R10, R8 # spill
113     addl    R10, #10    # int
117     imull   RDI, R10    # int
11b     movl    R11, R8 # spill
11e     addl    R11, #8 # int
122     movdl   R10, XMM2   # spill
127     imull   R10, R11    # int
12b     movl    [rsp + #48], R10    # spill
130     movl    R10, R8 # spill
133     addl    R10, #7 # int
137     movdl   R11, XMM1   # spill
13c     imull   R11, R10    # int
140     movl    [rsp + #52], R11    # spill
145     movl    R11, R8 # spill
148     addl    R11, #6 # int
14c     movdl   R10, XMM4   # spill
151     imull   R10, R11    # int
155     movl    [rsp + #56], R10    # spill
15a     movl    R10, R8 # spill
15d     addl    R10, #5 # int
161     movdl   R11, XMM6   # spill
166     imull   R11, R10    # int
16a     movl    [rsp + #60], R11    # spill
16f     movl    R11, R8 # spill
172     addl    R11, #4 # int
176     imull   RBX, R11    # int
17a     movl    R11, R8 # spill
17d     addl    R11, #3 # int
181     imull   RCX, R11    # int
185     movl    R10, R8 # spill
188     addl    R10, #2 # int
18c     imull   RDX, R10    # int
190     movl    R11, R8 # spill
193     incl    R11 # int
196     imull   R9, R11 # int
19a     addl    R9, [RSP + #32 (32-bit)]    # int
19f     addl    R9, RDX # int
1a2     addl    R9, RCX # int
1a5     addl    R9, RBX # int
1a8     addl    R9, [RSP + #60 (32-bit)]    # int
1ad     addl    R9, [RSP + #56 (32-bit)]    # int
1b2     addl    R9, [RSP + #52 (32-bit)]    # int
1b7     addl    R9, [RSP + #48 (32-bit)]    # int
1bc     movl    R10, R8 # spill
1bf     addl    R10, #9 # int
1c3     imull   R10, RSI    # int
1c7     addl    R10, R9 # int
1ca     addl    R10, RDI    # int
1cd     addl    R10, [RSP + #44 (32-bit)]   # int
1d2     movl    R11, R8 # spill
1d5     addl    R11, #12    # int
1d9     imull   R13, R11    # int
1dd     addl    R13, R10    # int
1e0     movl    R10, R8 # spill
1e3     addl    R10, #13    # int
1e7     imull   R10, [RSP + #40 (32-bit)]   # int
1ed     addl    R10, R13    # int
1f0     addl    RBP, R10    # int
1f3     addl    R14, RBP    # int
1f6     movl    R10, R8 # spill
1f9     addl    R10, #16    # int
1fd     cmpl    R10, #999999985
204     jl     B2   # loop end  P=1.000000 C=7419903.000000

在这里,我们观察到更多的“溢出”和更多的堆栈访问[RSP +…]],因为需要保留更多的中间结果。

因此,问题的答案很简单:2 * (i * i)比2 * i * i快,因为JIT为第一种情况生成了更优的程序集代码。


当然,很明显,第一个版本和第二个版本都不好;因为任何x86-64 CPU都至少支持SSE2,所以循环真的可以从向量化中受益。

这是优化器的问题;就像通常的情况一样,它展开得过于激进,搬起石头砸自己的脚,同时错过了其他各种机会。

事实上,现代x86-64 cpu将指令进一步分解为微操作(μ ops),并具有寄存器重命名、μ op缓存和循环缓冲区等功能,循环优化需要更多的技巧,而不是简单的展开以获得最佳性能。根据Agner Fog的优化指南:

由于μ op缓存在性能上的增益是相当大的 如果平均指令长度超过4字节,则相当可观。 下面的方法可以优化使用μ op缓存 被认为是: 确保关键循环足够小,以适应µop缓存。 将最关键的循环条目和函数条目对齐32。 避免不必要的循环展开。 避免有额外加载时间的指令 ……

关于这些加载时间——即使是最快的L1D命中也需要4个周期,一个额外的寄存器和μ op,所以是的,即使是少量的内存访问也会在紧循环中损害性能。

但回到向量化的机会——看看它能有多快,我们可以用GCC编译一个类似的C应用程序,它完全向量化它(AVX2显示,SSE2类似)2:

  vmovdqa ymm0, YMMWORD PTR .LC0[rip]
  vmovdqa ymm3, YMMWORD PTR .LC1[rip]
  xor eax, eax
  vpxor xmm2, xmm2, xmm2
.L2:
  vpmulld ymm1, ymm0, ymm0
  inc eax
  vpaddd ymm0, ymm0, ymm3
  vpslld ymm1, ymm1, 1
  vpaddd ymm2, ymm2, ymm1
  cmp eax, 125000000      ; 8 calculations per iteration
  jne .L2
  vmovdqa xmm0, xmm2
  vextracti128 xmm2, ymm2, 1
  vpaddd xmm2, xmm0, xmm2
  vpsrldq xmm0, xmm2, 8
  vpaddd xmm0, xmm2, xmm0
  vpsrldq xmm1, xmm0, 4
  vpaddd xmm0, xmm0, xmm1
  vmovd eax, xmm0
  vzeroupper

运行时间:

SSE: 0.24秒,或2倍的速度。 AVX: 0.15秒,或3倍的速度。 AVX2: 0.08秒,快5倍。


要获得JIT生成的程序集输出,请获取一个调试JVM并使用-XX:+PrintOptoAssembly运行

2 C版本使用-fwrapv标志编译,这使GCC能够将有符号整数溢出视为两个补码的自动换行。

添加的两种方法生成的字节代码略有不同:

  17: iconst_2
  18: iload         4
  20: iload         4
  22: imul
  23: imul
  24: iadd

对于2 * (i * i) vs:

  17: iconst_2
  18: iload         4
  20: imul
  21: iload         4
  23: imul
  24: iadd

对于2 * i * i。

当像这样使用JMH基准时:

@Warmup(iterations = 5, batchSize = 1)
@Measurement(iterations = 5, batchSize = 1)
@Fork(1)
@BenchmarkMode(Mode.AverageTime)
@OutputTimeUnit(TimeUnit.MILLISECONDS)
@State(Scope.Benchmark)
public class MyBenchmark {

    @Benchmark
    public int noBrackets() {
        int n = 0;
        for (int i = 0; i < 1000000000; i++) {
            n += 2 * i * i;
        }
        return n;
    }

    @Benchmark
    public int brackets() {
        int n = 0;
        for (int i = 0; i < 1000000000; i++) {
            n += 2 * (i * i);
        }
        return n;
    }

}

区别很明显:

# JMH version: 1.21
# VM version: JDK 11, Java HotSpot(TM) 64-Bit Server VM, 11+28
# VM options: <none>

Benchmark                      (n)  Mode  Cnt    Score    Error  Units
MyBenchmark.brackets    1000000000  avgt    5  380.889 ± 58.011  ms/op
MyBenchmark.noBrackets  1000000000  avgt    5  512.464 ± 11.098  ms/op

你观察到的是正确的,而不仅仅是你的基准测试风格的异常(例如,没有热身,参见如何用Java编写正确的微基准测试?)

再次与Graal一起运行:

# JMH version: 1.21
# VM version: JDK 11, Java HotSpot(TM) 64-Bit Server VM, 11+28
# VM options: -XX:+UnlockExperimentalVMOptions -XX:+EnableJVMCI -XX:+UseJVMCICompiler

Benchmark                      (n)  Mode  Cnt    Score    Error  Units
MyBenchmark.brackets    1000000000  avgt    5  335.100 ± 23.085  ms/op
MyBenchmark.noBrackets  1000000000  avgt    5  331.163 ± 50.670  ms/op

您可以看到,结果更加接近,这是有道理的,因为Graal是一个整体性能更好、更现代的编译器。

因此,这实际上只是取决于JIT编译器优化特定代码段的能力,并不一定有逻辑上的原因。

使用Java 11并使用以下VM选项关闭循环展开的有趣观察:

-XX:LoopUnrollLimit=0

带有2 * (i * i)表达式的循环会产生更紧凑的本机代码1:

L0001: add    eax,r11d
       inc    r8d
       mov    r11d,r8d
       imul   r11d,r8d
       shl    r11d,1h
       cmp    r8d,r10d
       jl     L0001

与2 * I * I版本相比:

L0001: add    eax,r11d
       mov    r11d,r8d
       shl    r11d,1h
       add    r11d,2h
       inc    r8d
       imul   r11d,r8d
       cmp    r8d,r10d
       jl     L0001

Java版本:

java version "11" 2018-09-25
Java(TM) SE Runtime Environment 18.9 (build 11+28)
Java HotSpot(TM) 64-Bit Server VM 18.9 (build 11+28, mixed mode)

基准测试结果:

Benchmark          (size)  Mode  Cnt    Score     Error  Units
LoopTest.fast  1000000000  avgt    5  694,868 ±  36,470  ms/op
LoopTest.slow  1000000000  avgt    5  769,840 ± 135,006  ms/op

基准测试源代码:

@BenchmarkMode(Mode.AverageTime)
@OutputTimeUnit(TimeUnit.MILLISECONDS)
@Warmup(iterations = 5, time = 5, timeUnit = TimeUnit.SECONDS)
@Measurement(iterations = 5, time = 5, timeUnit = TimeUnit.SECONDS)
@State(Scope.Thread)
@Fork(1)
public class LoopTest {

    @Param("1000000000") private int size;

    public static void main(String[] args) throws RunnerException {
        Options opt = new OptionsBuilder()
            .include(LoopTest.class.getSimpleName())
            .jvmArgs("-XX:LoopUnrollLimit=0")
            .build();
        new Runner(opt).run();
    }

    @Benchmark
    public int slow() {
        int n = 0;
        for (int i = 0; i < size; i++)
            n += 2 * i * i;
        return n;
    }

    @Benchmark
    public int fast() {
        int n = 0;
        for (int i = 0; i < size; i++)
            n += 2 * (i * i);
        return n;
    }
}

1 -虚拟机选项使用:- xx:+UnlockDiagnosticVMOptions - xx:+PrintAssembly - xx:LoopUnrollLimit=0

(编者注:正如另一个答案所示,这个答案与来自asm的证据相矛盾。这个猜测得到了一些实验的支持,但结果并不正确。)


当乘法是2 * (i * i)时,JVM能够从循环中分解出乘法2,从而得到等效但更高效的代码:

int n = 0;
for (int i = 0; i < 1000000000; i++) {
    n += i * i;
}
n *= 2;

但是当乘法是(2 * i) * i时,JVM不会优化它,因为乘以常数不再恰好在n +=加法之前。

以下是我认为这种情况的几个原因:

在循环开始时添加if (n == 0) n = 1语句会导致两个版本的效率一样高,因为分解乘法不再保证结果相同 优化后的版本(通过分解2的乘法)与2 * (i * i)版本一样快

下面是我用来得出这些结论的测试代码:

public static void main(String[] args) {
    long fastVersion = 0;
    long slowVersion = 0;
    long optimizedVersion = 0;
    long modifiedFastVersion = 0;
    long modifiedSlowVersion = 0;

    for (int i = 0; i < 10; i++) {
        fastVersion += fastVersion();
        slowVersion += slowVersion();
        optimizedVersion += optimizedVersion();
        modifiedFastVersion += modifiedFastVersion();
        modifiedSlowVersion += modifiedSlowVersion();
    }

    System.out.println("Fast version: " + (double) fastVersion / 1000000000 + " s");
    System.out.println("Slow version: " + (double) slowVersion / 1000000000 + " s");
    System.out.println("Optimized version: " + (double) optimizedVersion / 1000000000 + " s");
    System.out.println("Modified fast version: " + (double) modifiedFastVersion / 1000000000 + " s");
    System.out.println("Modified slow version: " + (double) modifiedSlowVersion / 1000000000 + " s");
}

private static long fastVersion() {
    long startTime = System.nanoTime();
    int n = 0;
    for (int i = 0; i < 1000000000; i++) {
        n += 2 * (i * i);
    }
    return System.nanoTime() - startTime;
}

private static long slowVersion() {
    long startTime = System.nanoTime();
    int n = 0;
    for (int i = 0; i < 1000000000; i++) {
        n += 2 * i * i;
    }
    return System.nanoTime() - startTime;
}

private static long optimizedVersion() {
    long startTime = System.nanoTime();
    int n = 0;
    for (int i = 0; i < 1000000000; i++) {
        n += i * i;
    }
    n *= 2;
    return System.nanoTime() - startTime;
}

private static long modifiedFastVersion() {
    long startTime = System.nanoTime();
    int n = 0;
    for (int i = 0; i < 1000000000; i++) {
        if (n == 0) n = 1;
        n += 2 * (i * i);
    }
    return System.nanoTime() - startTime;
}

private static long modifiedSlowVersion() {
    long startTime = System.nanoTime();
    int n = 0;
    for (int i = 0; i < 1000000000; i++) {
        if (n == 0) n = 1;
        n += 2 * i * i;
    }
    return System.nanoTime() - startTime;
}

结果如下:

Fast version: 5.7274411 s
Slow version: 7.6190804 s
Optimized version: 5.1348007 s
Modified fast version: 7.1492705 s
Modified slow version: 7.2952668 s