OpenMP 原子指令设计与实现

#pragma omp atomic
表达式;

x binop = 表达式;
x++;
x--;
++x;
--x;

#include <stdio.h>
#include <omp.h>

int main()
{
  int data = 1;
#pragma omp parallel num_threads(4) shared(data) default(none)
  {
#pragma omp atomic
    data += data * 2;
  }
  printf("data = %d\n", data);
  return 0;
}

data = 81

#include <stdio.h>
#include <omp.h>

int main()
{
  int data = 0;
#pragma omp parallel num_threads(4) shared(data) default(none)
  {
#pragma omp atomic
    data += 1;
  }
  printf("data = %d\n", data);
  return 0;
}

0000000000401193 <main._omp_fn.0>:
  401193:       55                      push   %rbp
  401194:       48 89 e5                mov    %rsp,%rbp
  401197:       48 89 7d f8             mov    %rdi,-0x8(%rbp)
  40119b:       48 8b 45 f8             mov    -0x8(%rbp),%rax
  40119f:       48 8b 00                mov    (%rax),%rax
  4011a2:       f0 83 00 01             lock addl $0x1,(%rax) # 这就是编译出来的原子指令——对应x86平台
  4011a6:       5d                      pop    %rbp
  4011a7:       c3                      retq   
  4011a8:       0f 1f 84 00 00 00 00    nopl   0x0(%rax,%rax,1)
  4011af:       00 

0000000000401193 <main._omp_fn.0>:
  401193:       55                      push   %rbp
  401194:       48 89 e5                mov    %rsp,%rbp
  401197:       48 89 7d f8             mov    %rdi,-0x8(%rbp)
  40119b:       48 8b 45 f8             mov    -0x8(%rbp),%rax
  40119f:       48 8b 00                mov    (%rax),%rax
  4011a2:       f0 83 28 01             lock subl $0x1,(%rax)
  4011a6:       5d                      pop    %rbp
  4011a7:       c3                      retq   
  4011a8:       0f 1f 84 00 00 00 00    nopl   0x0(%rax,%rax,1)
  4011af:       00 

#include <stdio.h>
#include <omp.h>

int main()
{
  int data = 1;
#pragma omp parallel num_threads(4) shared(data) default(none)
  {
#pragma omp atomic
    data *= 2;
  }
  printf("data = %d\n", data);
  return 0;
}

0000000000401193 <main._omp_fn.0>:
  401193:       55                      push   %rbp
  401194:       48 89 e5                mov    %rsp,%rbp
  401197:       48 89 7d f8             mov    %rdi,-0x8(%rbp)
  40119b:       48 8b 45 f8             mov    -0x8(%rbp),%rax
  40119f:       48 8b 08                mov    (%rax),%rcx
  4011a2:       8b 01                   mov    (%rcx),%eax
  4011a4:       89 c2                   mov    %eax,%edx
  4011a6:       8d 34 12                lea    (%rdx,%rdx,1),%esi # 这条语句的含义为 data *= 2
  4011a9:       89 d0                   mov    %edx,%eax
  4011ab:       f0 0f b1 31             lock cmpxchg %esi,(%rcx)
  4011af:       89 d6                   mov    %edx,%esi
  4011b1:       89 c2                   mov    %eax,%edx
  4011b3:       39 f0                   cmp    %esi,%eax
  4011b5:       75 ef                   jne    4011a6 <main._omp_fn.0+0x13>
  4011b7:       5d                      pop    %rbp
  4011b8:       c3                      retq   
  4011b9:       0f 1f 80 00 00 00 00    nopl   0x0(%rax)

#include <stdio.h>
#include <stdbool.h>
#include <stdatomic.h>

// 这个函数对应上面的汇编程序
void atomic_multiply(int* data)
{
  int oldval = *data;
  int write = oldval * 2;
  // __atomic_compare_exchange_n 这个函数的作用就是
  // 将 data 指向的值和 old 的值进行比较，如果相等就将 write 的值写入 data
  // 指向的内存地址 如果操作成功返回 true 否则返回 false
  while (!__atomic_compare_exchange_n (data, &oldval, write, false,
                                      __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
  {
    oldval = *data;
    write = oldval * 2;
  }
}

int main()
{
  int data = 2;
  atomic_multiply(&data);
  printf("data = %d\n", data);
  return 0;
}

0000000000401193 <main._omp_fn.0>:
  401193:       55                      push   %rbp
  401194:       48 89 e5                mov    %rsp,%rbp
  401197:       48 89 7d f8             mov    %rdi,-0x8(%rbp)
  40119b:       48 8b 45 f8             mov    -0x8(%rbp),%rax
  40119f:       48 8b 08                mov    (%rax),%rcx
  4011a2:       8b 01                   mov    (%rcx),%eax
  4011a4:       89 c2                   mov    %eax,%edx
  4011a6:       89 d0                   mov    %edx,%eax
  4011a8:       c1 e8 1f                shr    $0x1f,%eax
  4011ab:       01 d0                   add    %edx,%eax
  4011ad:       d1 f8                   sar    %eax
  4011af:       89 c6                   mov    %eax,%esi
  4011b1:       89 d0                   mov    %edx,%eax
  4011b3:       f0 0f b1 31             lock cmpxchg %esi,(%rcx)
  4011b7:       89 d6                   mov    %edx,%esi
  4011b9:       89 c2                   mov    %eax,%edx
  4011bb:       39 f0                   cmp    %esi,%eax
  4011bd:       75 e7                   jne    4011a6 <main._omp_fn.0+0x13>
  4011bf:       5d                      pop    %rbp
  4011c0:       c3                      retq   
  4011c1:       66 2e 0f 1f 84 00 00    nopw   %cs:0x0(%rax,%rax,1)
  4011c8:       00 00 00 
  4011cb:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)