为什么std :: vector的速度是原始数组的两倍？包含完整代码

Question

结果:

矢量时间:7051

阵列时间:18944

我为此使用了MSVC释放模式,编译为32位.

在此测试之前,我正在查看GCC的矢量源代码并且感到惊讶,因为我认为operator[]检查了数组越界,但事实并非如此.但是,我没想到矢量这么快？!

完整代码:

#include <iostream>
#include <vector>

int main(){
    const int size = 10000;
    unsigned long long my_array[size];
    std::vector<unsigned long long> my_vec;

    my_vec.resize(size);

    //Populate containers
    for(int i=0; i<size; i++){
        my_vec[i] = i;
        my_array[i] = i;
    }

    //Initialise test variables
    unsigned long long sum = 0;
    unsigned long long time = 0;
    unsigned long long start = 0;
    unsigned long long finish = 0;

    //Time the vector
    start = __rdtsc();
    for(int i=0; i<size; i++){
        sum += my_vec[i];
    }
    finish = __rdtsc();


    time = finish - start;
    std::cout << "Vector time: " << time << "     " << sum << std::endl;


    sum = 0;

    //Time the array
    start = __rdtsc();
    for(int i=0; i<size; i++){
        sum += my_array[i];
    }
    finish = __rdtsc();

    time = finish - start;
    std::cout << "Array time: " << time << "     " << sum << std::endl;

    int t = 8;
    std::cin >> t;
    return 0;
}

Answer 1

以下是使用MSVC 2013.

对于矢量:

0019138E  mov         edi,edi  
  for (int i = 0; i<size; i++){
00191390  lea         ecx,[ecx+20h]  
    sum += my_vec[i];
00191393  movdqu      xmm0,xmmword ptr [ecx-20h]  
00191398  paddq       xmm1,xmm0  
0019139C  movdqu      xmm0,xmmword ptr [ecx-10h]  
001913A1  paddq       xmm2,xmm0  
001913A5  dec         esi  
001913A6  jne         main+0F0h (0191390h)  
  }

对于数组:

0019142D  lea         ecx,[ecx]  
  for (int i = 0; i<size; i++){
00191430  lea         ecx,[ecx+20h]  
    sum += my_array[i];
00191433  movdqu      xmm0,xmmword ptr [ecx-30h]  
00191438  paddq       xmm1,xmm0  
0019143C  movdqu      xmm0,xmmword ptr [ecx-20h]  
00191441  paddq       xmm2,xmm0  
00191445  dec         esi  
00191446  jne         main+190h (0191430h)  
  }

如您所见,内环是相同的.实际上,怀疑它是硬件的东西,我交换了两个循环,并且数组更快地到达相同的边距(实际上,在现实世界中,它们都不比其他更快或更慢).

我预测这是某种CPU缓存行为:https: //en.wikipedia.org/wiki/CPU_cache