c++ - 为什么内联汇编程序在此代码中有时更快而有时更慢？每次运行的执行时间变化很大

Question

我已经编写了一些 C++ 代码来测试 C++ 和内联汇编代码的时间。最初我只是玩得开心，但后来我注意到每次运行我的程序时，我都会得到不同的结果。有时 C++ 更快，有时内联汇编代码更快，有时它们都一样。

这是怎么回事？

这是带有程序输出的代码:

#define TRIALS 1000000
#include <iostream>
using namespace std;
typedef std::chrono::high_resolution_clock Clock;
int main()
{
  auto t1 = Clock::now();
  auto t2 = Clock::now();
  int X3=17;
  int X2=17;
  int X4=17;
  int X=17;



  int sum=0;
  int avg=0;
  cout << "=================================" << endl;
  cout << "| var*=10;                      |" << endl;
  cout << "=================================" << endl;

  for( int i=0; i<TRIALS; i++ )
    {
      X3=17;
      t1 = Clock::now();  
      X3*=10;
      t2 = Clock::now();
      sum+=chrono::duration_cast<std::chrono::nanoseconds>(t2 - t1).count();
    }
  avg=sum/TRIALS;
  cout << "| Product:  " << X3<< "  "<< avg << " nanoseconds |" << endl;
  cout << "=================================" << endl;
  cout << endl << endl;

  avg=sum=0;
  cout << "=================================" << endl;
  cout << "| use inline assembler with shl |" << endl;
  cout << "=================================" << endl;

  for( int i=0; i<TRIALS; i++ )
    {
      X=17;
      t1 = Clock::now();
      asm /*volatile*/ (
            "movl %0, %%eax;" // X->ax
            "shll %%eax;"// ax*=2
            "movl %%eax, %%ebx;" // ax->bx
            "shll %%eax;" // ax*=2
            "shll %%eax;" // ax*=2
            "add %%ebx, %%eax;" // bx+ax->ax
            : "=a" (X)
            : "a" (X)
            : "%ebx"
            );
      t2 = Clock::now();
      sum+=chrono::duration_cast<std::chrono::nanoseconds>(t2 - t1).count();
    }
  avg=sum/TRIALS;
  cout << "| Product:  " << X << "  "<< avg << " nanoseconds |" << endl;
  cout << "=================================" << endl;
  cout << endl << endl;
  avg=sum=0;

  cout << "=================================" << endl;
  cout << "| var=var*10                    |" << endl;
  cout << "=================================" << endl;

  for( int i=0; i<TRIALS; i++ )
    {
      X2=17;
      t1 = Clock::now();
      X2=X2*10;
      t2 = Clock::now();
      sum+=chrono::duration_cast<std::chrono::nanoseconds>(t2 - t1).count();
    }
  avg=sum/TRIALS;
  cout << "| Product:  " << X3<< "  "<< avg << " nanoseconds |" << endl;
  cout << "=================================" << endl;
  cout << endl << endl;

  avg=sum=0;


  cout << "=================================" << endl;
  cout << "| use inline assembler with mul |" << endl;
  cout << "=================================" << endl;
  for( int i=0; i<TRIALS; i++ )
    {
      X4=17;
      t1 = Clock::now();
      asm  (
    "movl %0, %%eax;" // X->ax
    "movl $0x0A, %%ebx;" // 10->bx
    "mul %%ebx;" // 10*ax->ax
    : "=a" (X4)
    : "a" (X4)
    : "%ebx"
    );
      t2 = Clock::now();
      sum+=chrono::duration_cast<std::chrono::nanoseconds>(t2 - t1).count();
    }
  avg=sum/TRIALS;
  cout << "| Product:  " << X4<< "  "<< avg << " nanoseconds |" << endl;
  cout << "=================================" << endl;
  cout << endl;

  return(0);
}

程序输出#1:

=================================
| var*=10;                      |
=================================
| Product:  170  50 nanoseconds |
=================================


=================================
| use inline assembler with shl |
=================================
| Product:  170  50 nanoseconds |
=================================


=================================
| var=var*10                    |
=================================
| Product:  170  50 nanoseconds |
=================================


=================================
| use inline assembler with mul |
=================================
| Product:  170  50 nanoseconds |
=================================

输出#2:

=================================
| var*=10;                      |
=================================
| Product:  170  62 nanoseconds |
=================================


=================================
| use inline assembler with shl |
=================================
| Product:  170  57 nanoseconds |
=================================


=================================
| var=var*10                    |
=================================
| Product:  170  59 nanoseconds |
=================================


=================================
| use inline assembler with mul |
=================================
| Product:  170  58 nanoseconds |
=================================

Answer 1

这些更像是提示，而不仅仅是“解决方案”:

1) 将 TRAILS 提高几个数量级以实际测量秒范围内的东西

2) 重复测量几次(n=100 或更多)并取平均值(如果您关心统计数据，则平均值误差 = rms/sqrt(n))

3) 实际测量你想测量的东西:至少只把你感兴趣的代码放到 TRAILS 循环中，即:

t1 = Clock::now();  
for( int i=0; i<TRIALS; i++ )
    {
     ... only code relevant for your calculation here ...
    }
t2 = Clock::now();
sum = chrono::duration_cast<std::chrono::nanoseconds>(t2 - t1).count();

4) 最后，考虑 Godbolt 编译器资源管理器服务 https://godbolt.org/您可以在其中检查代码的汇编器输出以了解各种优化器设置。对于像你的代码一样简单的代码(我试过)它只是这样做(使用 -O3):mov eax,170 所以你看:编译器很聪明，你不能轻易用内联汇编器打败他!这肯定是非平凡示例的情况。