asked on 1/26/2009

gcc assembly help in converting windows assembly code to linux gcc.

64-bit counter increment/decrement on 32-bit system.

atomically in_value is incremented and updated value is returned.
Pseudo code...
__int64* increment(__Int64 *in_value)
{
__int64 *temp = NULL;
 
__asm  {
               mov edi, in_value
                mov eax, [edi]
                mov edx, [edi+4]
           retry:
                mov ecx, edx
                mov ebx, eax
                add ebx, 1
                adc ecx, 0
                lock cmpxchg8b [edi]
                jnz retry
                mov eax, temp
                xor edx, edx
                mov [eax+edx*4], ebx
                mov [eax+edx*4+4], ecx
     }
return *temp;
}
 
It works great. I want it be converted into gcc assembly.
 
I unsuccessfully converted it in to gcc assembly. But its wrong, i dont know how to pass in_value  (hi and low) in one ready.
 
#define low(x)       *(((unsigned int*)&(x))+0)
#define high(x)      *(((unsigned int*)&(x))+1)
 
asm volatile (  "mov   %0,  %%edi \n"
                "mov   %1,  %%eax  \n"
                "mov   %2,  %%edx \n"
                "1: \n"
                "mov   %%edx,  %%ecx\n"
             "mov   %%eax, %%ebx \n"
              "add   $1, %%ebx \n"
            "adc   $0, %%ecx \n"
            "lock;  cmpxchg8b %0\n"
            "jnz   1b \n"
                "mov   %%edx, %%edx \n"
                "mov   %3, %%ebx \n"     //Error
                "mov   %3, %%ecx \n"     //Error
               : "=m" (*in_value)
   : "m" (low(*in_value)), "m" (high(*in_value)), "m" (*in_value)
               : "memory", "ebx", "ecx", "eax", "edi");
 
in_value should not re-read for one cycle to preserve the atomicity.
thank you

Last Comment

1/27/2009

1/26/2009

have You tried gcc built-ins? http://gcc.gnu.org/onlinedocs/gcc-4.1.0/gcc/Atomic-Builtins.html
Also please refer http://www.exit.com/blog/archives/000361.html discussion.

Do You still want the bit2bit translation from windows?

SimonHowald

1/26/2009

ASKER

Yes I did look into the second link but it does not compile on linux. I am looking for compilable version of your second link or bit2bit conversion from windows.
Thanks

ravenpl

1/26/2009

Right - about gcc correctness

#include <stdio.h>
int main()
{
    volatile long long target = 7;
    unsigned int incr = 7;
 
    __asm__ __volatile__(
    " movl %0, %%eax\n"
    " movl 4+%0, %%edx\n"
    "1: movl %1, %%ebx\n"
    " xorl %%ecx, %%ecx\n"
    " addl %%eax, %%ebx\n"
    " adcl %%edx, %%ecx\n"
    "lock; cmpxchg8b %0\n"
    " jnz 1b"
    : "+o" (target)
    : "m" (incr)
    : "memory", "eax", "ebx", "ecx", "edx", "cc");
 
    printf("%lld\n", target);
    return 0;
}

Open in new window

SimonHowald

1/26/2009

ASKER

thank you I am yet to test but I think you solved my problem and there is one more thing to be done. I need to return the updated counter.
the before the update counter value is there in b and c register. This value should put into local varriable and return it.
int64 increment(__int64 *in_value)
{
return temp;
}

ravenpl

1/26/2009

//untested yet
long long inc64(volatile long long *in)
{
    long long rv;
    __asm__ __volatile__(
    " movl 0+%0, %%eax\n"
    " movl 4+%0, %%edx\n"
    "1: movl $1, %%ebx\n"
    " xorl %%ecx, %%ecx\n"
    " addl %%eax, %%ebx\n"
    " adcl %%edx, %%ecx\n"
    "lock; cmpxchg8b %0\n"
    " jnz 1b\n"
    " movl %%ebx, %1\n"
    " movl %%ecx, 4+%1\n"
    : "+o" (*in)
    : "m"(rv)
    : "memory", "eax", "ebx", "ecx", "edx", "cc");
    return rv;
}

Open in new window

ASKER CERTIFIED SOLUTION

ravenpl

1/26/2009

THIS SOLUTION IS ONLY AVAILABLE TO MEMBERS.

View this solution by signing up for a free trial.

Members can start a 7-Day free trial and enjoy unlimited access to the platform.

See Pricing Options

Start Free Trial

ravenpl

1/26/2009

Although the code works fine(due to volatile), the rv should be placed on the output parametrs, rather than input
: "+o" (*in), "=m" (rv) : : "memory" ....

SimonHowald

1/26/2009

ASKER

Thank you all worked. Now to change it to decrement. I just need to change
addl to subl and adc to sbb right?

ravenpl

1/27/2009

I would rather add -1
< "1: movl $1, %%ebx\n"
< " xorl %%ecx, %%ecx\n"
> "1: movl $-1, %%ebx\n"
> " movl %%ebx, %%ecx\n"

SimonHowald

1/27/2009

ASKER

ravenpl is awesome. I salute his spirit of responding query on time. Your answer works great .. Looking forward to your help in future.