inline assembly -> Windows to Linux port

[selectany]

Hi!

I'd like to port my existing source code of chess to work on Linux using GCC.
The problem is with inline assembler code for LSB routine.
The following code works properly only in Debug mode.
If I use some kind of optimization, the result of routine is bogus:
- in Release build routine always returns 0;
- if I force inlining with __inline__ __attribute__((always_inline)) ), the result of the routine is random.

Here is the code:

Code: Select all

int LSB(unsigned long long arg1)
{
        int hword = (int)(arg1 >> 32);

        asm(
        "bsfl %0,%%eax               \n\t"
        "jnz 1f                      \n\t"
        "bsfl %1,%%eax               \n\t"
        "addl $32,%%eax              \n\t"
        "1:                          \n\t"
        :
        :"m"(arg1),"m"(hword)
        :"%eax");
}

PP: Platform: ArchLinux, IDE: NetBeans

[tvrzsky]

Try this:

Code: Select all

int LSBssf(unsigned long long arg1) 
{ 
        int hword = (int)(arg1 >> 32); 
        int retval;
        asm(
        "bsfl %1,%0               \n\t" 
        "jnz 1f                      \n\t" 
        "bsfl %2,%0               \n\t" 
        "addl $32,%0              \n\t" 
        "1:                          \n\t" 
        : "=&q" (retval)
        :"m"(arg1),"q"(hword) 
        : );
        return retval;
}

Or this:

Code: Select all

int LSBb(unsigned long long arg)
{
    if ((int)arg) return __builtin_ctz((arg));
    else return (__builtin_ctz((arg>>32)));
}

[tvrzsky]

Oops, of course I meant:

Code: Select all

int LSBb(unsigned long long arg)
{
    if ((int)arg) return __builtin_ctz((arg));
    else return (__builtin_ctz((arg>>32))) + 32;
}

[selectany]

Thanks

I didn't know about these built-in things.
I found even more suitable one: __builtin_ctzll.
Is it better to use it?

[tvrzsky]

Thanks

I didn't know about these built-in things.
I found even more suitable one: __builtin_ctzll.
Is it better to use it?

Surely, I did not take notice of this one. At least migration to 64bits will be effortless.