(转)C协程实现的效率对比

前段时间实现的C协程依赖栈传递参数，在开启优化时会导致错误，于是实现了一个ucontext的版本,但ucontext的切换效率太差了，

在我的机器上执行4000W次切换需要11秒左右，这达不到我的要求，所以重新设计了实现，使得在开启优化时也能得到正确的结果.

并且效率也令人满意，4000W次切换仅需要730ms左右，足足比ucontext的实现快乐近15倍。

下面贴出实现:

#include "uthread.h"
#include <stdlib.h>
#include <ucontext.h>
#include <pthread.h>
#include "link_list.h"

struct uthread
{
    int32_t reg[8];//0:esp,1:ebp,2:eax,3:ebx,4:ecx,5:edx,6:edi,7:esi
    void *para;
    uthread_t parent;
    void*(*main_fun)(void*);
    void *stack;
    int32_t ssize;
    int8_t first_run;
};

#ifdef _DEBUG
//for debug version
void uthread_main_function()
{
    int32_t arg;
     __asm__ volatile(
        "movl %%eax,%0	
"
        :
        :"m"(arg)
    );    
    
    uthread_t u = (uthread_t)arg;
    void *ret = u->main_fun(u->para);
    if(u->parent)
        uthread_switch(u,u->parent,ret);
    else
        exit(0); 
}
#else
//for release version
void __attribute__((regparm(1))) uthread_main_function(void *arg)
{
    uthread_t u = (uthread_t)arg;
    void *ret = u->main_fun(u->para);
    if(u->parent)
        uthread_switch(u,u->parent,ret);
    else
        exit(0);
}
#endif
uthread_t uthread_create(uthread_t parent,void*stack,uint32_t stack_size,void*(*fun)(void*))
{
    uthread_t u = (uthread_t)calloc(1,sizeof(*u));
    u->parent = parent;
    u->main_fun = fun;
    u->stack = stack;
    u->ssize = stack_size;
    if(stack)
    {
        u->reg[0] = (int32_t)stack+stack_size-4;
        u->reg[1] = (int32_t)stack+stack_size-4;
    }
    if(u->main_fun)
        u->first_run = 1;
    return u;
}

void uthread_destroy(uthread_t *u)
{
    free(*u);
    *u = NULL;
}

#ifdef _DEBUG
void* __attribute__((regparm(3))) uthread_switch(uthread_t from,uthread_t to,void *para)
{
    if(!from)
        return NULL;
    to->para = para;
    int32_t esp,ebp,eax,ebx,ecx,edx,edi,esi;
    //save current registers
    //the order is important    
     __asm__ volatile(
        "movl %%eax,%2	
"
        "movl %%ebx,%3	
"
        "movl %%ecx,%4	
"
        "movl %%edx,%5	
"
        "movl %%edi,%6	
"
        "movl %%esi,%7	
"     
        "movl %%ebp,%1	
"
        "movl %%esp,%0	
"
        :
        :"m"(esp),"m"(ebp),"m"(eax),"m"(ebx),"m"(ecx),"m"(edx),"m"(edi),"m"(esi)
    );
    from->reg[0] = esp;
    from->reg[1] = ebp;
    from->reg[2] = eax;
    from->reg[3] = ebx;
    from->reg[4] = ecx;
    from->reg[5] = edx;
    from->reg[6] = edi;
    from->reg[7] = esi;    
    if(to->first_run)
    {
       to->first_run = 0;
       esp = to->reg[0];
       //use eax to pass arg
       eax = (int32_t)to;
        __asm__ volatile (
            "movl %1,%%eax	
"
            "movl %0,%%ebp	
"
            "movl %%ebp,%%esp	
"
            :
            :"m"(esp),"m"(eax)
        );       
       uthread_main_function();
    }
    else
    {
        esp = to->reg[0];
        ebp = to->reg[1];
        eax = to->reg[2];
        ebx = to->reg[3];
        ecx = to->reg[4];
        edx = to->reg[5];
        edi = to->reg[6];
        esi = to->reg[7];
        //the order is important
        __asm__ volatile (
            "movl %2,%%eax	
"
            "movl %3,%%ebx	
"
            "movl %4,%%ecx	
"
            "movl %5,%%edx	
"
            "movl %6,%%edi	
"
            "movl %7,%%esi	
"        
            "movl %1,%%ebp	
"
            "movl %0,%%esp	
"
            :
            :"m"(esp),"m"(ebp),"m"(eax),"m"(ebx),"m"(ecx),"m"(edx),"m"(edi),"m"(esi)
        );
    }    
    return from->para;
}
#else
void* __attribute__((regparm(3))) uthread_switch(uthread_t from,uthread_t to,void *para)
{
    if(!from)
        return NULL;
    to->para = para;
    int32_t esp,ebp,edi,esi;
    //save current registers
    //the order is important    
     __asm__ volatile(
        "movl %%eax,%2	
"
        "movl %%ebx,%3	
"
        "movl %%ecx,%4	
"
        "movl %%edx,%5	
"
        "movl %%edi,%6	
"
        "movl %%esi,%7	
"     
        "movl %%ebp,%1	
"
        "movl %%esp,%0	
"
        :
        :"m"(from->reg[0]),"m"(from->reg[1]),"m"(from->reg[2]),"m"(from->reg[3])
        ,"m"(from->reg[4]),"m"(from->reg[5]),"m"(from->reg[6]),"m"(from->reg[7])
    );
    if(to->first_run)
    {
       to->first_run = 0;   
       //change stack
       //the order is important
        __asm__ volatile (
            "movl %0,%%ebp	
"
            "movl %%ebp,%%esp	
"
            :
            :"m"(to->reg[0])
        );               
       uthread_main_function((void*)to);
    }
    else
    {
        esp = to->reg[0];
        ebp = to->reg[1];
        edi = to->reg[6];
        esi = to->reg[7];
        //the order is important
        __asm__ volatile (
            "movl %2,%%eax	
"
            "movl %3,%%ebx	
"
            "movl %4,%%ecx	
"
            "movl %5,%%edx	
"
            "movl %6,%%edi	
"
            "movl %7,%%esi	
"        
            "movl %1,%%ebp	
"
            "movl %0,%%esp	
"
            :
            :"m"(esp),"m"(ebp),"m"(to->reg[2]),"m"(to->reg[3])
            ,"m"(to->reg[4]),"m"(to->reg[5]),"m"(edi),"m"(esi)
        );
    }    
    return from->para;
}
#endif

test.c

#include <stdio.h>
#include "uthread.h"
#include "SysTime.h"
#include <stdlib.h>
void* ufun2(void *arg)
{
    printf("ufun2
");
    char **tmp = (char**)arg;
    uthread_t self = (uthread_t)tmp[0];
    uthread_t parent = (uthread_t)tmp[1];
    volatile void *ptr = self;
    while(ptr)
    {
        ptr = uthread_switch(self,parent,NULL);
    }
    return NULL;
}

char *stack1;
char *stack2;

void* ufun1(void *arg)
{
    uthread_t self = (uthread_t)arg;
    uthread_t u = uthread_create(self,stack2,4096,ufun2);
    char* _arg[2];
    _arg[0] = (char*)u;
    _arg[1] = (char*)self;
    int i = 0;
    uint32_t tick = GetSystemMs();
    for( ; i < 20000000; ++i)
    {
        uthread_switch(self,u,&_arg[0]);
    }
    printf("%d
",GetSystemMs()-tick);
    uthread_switch(self,u,NULL);
    return arg;
}

int main()
{
    stack1 = (char*)malloc(4096);
    stack2 = (char*)malloc(4096);
    /*
     * if use ucontext version
    char dummy_stack[4096];
    uthread_t p = uthread_create(NULL,dummy_stack,0,NULL);
    */
    uthread_t p = uthread_create(NULL,NULL,0,NULL);
    uthread_t u = uthread_create(p,stack1,4096,ufun1);
    uthread_switch(p,u,u);
    printf("main end
");
    return 0;
};

转自：https://www.cnblogs.com/sniperHW/archive/2012/08/05/2624334.html