GCC原子操作(Atomic Builtins)

/*
* =====================================================================================
*
*       Filename:  atomic.c
*
*    Description:  gcc-4.1.1: Built-in functions for atomic memory access
*                  http://gcc.gnu.org/onlinedocs/gcc-4.1.1/gcc/Atomic-Builtins.html

    compile:
         gcc atomic.c -lpthead
    run:
        time ./a.out 

    modify Macro VER to choose the version
            1. no lock
            2. pthread_mutex
            3. atomic
    ref:
        http://gcc.gnu.org/onlinedocs/gcc-4.1.1/gcc/Atomic-Builtins.html
        http://www.alexonlinux.com/multithreaded-simple-data-type-access-and-atomic-variables
		http://www.cnblogs.com/FrankTan/archive/2010/12/11/1903377.html

    the demo, create 4 threads. the odd thread do --, the even threads do ++. 
    so, the global_int should be zero. we can see:
            1. the "no lock version", fastest but not thread safe
            2. the "pthread mutex version" slowest, but thread safe
            3. the "atomic version", fast and thread safe

*        Version:  1.0
*        Created:  09/11/2012 05:28:13 PM
*       Revision:  none
*       Compiler:  gcc
*
*         Author:  TED (cn), guolb57@163.com
*       
* ===================================================================================
*/

#include <stdio.h>
#include <pthread.h>
#include <unistd.h>
#include <stdlib.h>
#include <errno.h>

/************************** for test the time cost begin *********************/
#include <sys/time.h>

struct timeval tp1, tp2;

static void start_timer()
{
    gettimeofday(&tp1, NULL);
}

/* return by seconds */
static double end_timer()
{
    gettimeofday(&tp2, NULL);
    return((tp2.tv_sec - tp1.tv_sec) + 
            (tp2.tv_usec - tp1.tv_usec) * 1.0e-6);
}

/* for test the time cost end */

#define LOOP_NUM 1000000

int global_int = 0;

#define VER              3  /* modify VER to choose the version here */

#define NO_LOCK_VER      1  /* quickest & not thread safe */
#define MUTEX_VER        2  /* slowest & thread safe */ 
#define ATOMIC_VER       3  /* quick & thread safe */

#if (VER == MUTEX_VER)
    pthread_mutex_t lock;
#endif

void *thread_proc( void *arg )
{
    int i;
    int flag = (int)arg;

    for (i = 0; i < LOOP_NUM; i++)	{
        if(flag % 2 == 0){
#if (VER == NO_LOCK_VER)
            global_int++;
#elif (VER == ATOMIC_VER)
            __sync_fetch_and_add( &global_int, 1 );
#elif (VER == MUTEX_VER)
            pthread_mutex_lock(&lock);
            global_int++;
            pthread_mutex_unlock(&lock);
#endif
        }
        else{
#if (VER == NO_LOCK_VER)
            global_int--;
#elif (VER == ATOMIC_VER)
            __sync_fetch_and_sub(&global_int, 1);
#elif (VER == MUTEX_VER)
            pthread_mutex_lock(&lock);
            global_int--;
            pthread_mutex_unlock(&lock);
#endif
        }
	}

	return NULL;
}

int main(int argc, const char *argv[])
{
   	int i;
	int thread_num = 4;
	pthread_t *pthread_id;

#if (VER == MUTEX_VER)
    pthread_mutex_init(&lock,NULL);
#endif

    int ntimes = 100;
    start_timer();
    while(ntimes--){
        pthread_id = (pthread_t*)malloc( sizeof( pthread_t ) * thread_num);
        if (NULL == pthread_id)	{
            perror( "malloc" );
            exit(1);
        }

        for (i = 0; i < thread_num; i++){
            if (pthread_create( &pthread_id[i], NULL, thread_proc, (void*)i)){
                perror( "pthread_create" );
                exit(1);
            }
        }

        for (i = 0; i < thread_num; i++){
            pthread_join(pthread_id[i], NULL);
        }

        free(pthread_id);

        printf( "global_int value is: %d\n", global_int );
        printf( "The Value should be: 0\n");

        global_int = 0;
    }

    printf("cost %g\n", (double)end_timer());

	return 0;
}