🚀 go-pugleaf

Christoph Brunner wrote:
> Hi,
> 
> I'm currently trying to get into extended inline ASM in gcc (gcc version 
> 3.2.2, gas 2.13.90.0.18).
> However, I'm having some trouble: The following code, using SSE 
> instructions, segfaults, and I don't really know why.
> 
> Perhaps someone could give me a hint?
> 
> void add_asm(float *a, float *b, float *c, int n)
> {
>    // a,b,c properly aligned; (n % 4) = 0
>    int iter = 0;
>     
>    __asm__ __volatile__
>    (
>     "calc:\n"
>     "subl $4, %%ecx\n"
>        
>     // perform package load/addition/store
>     "movaps (%3), %%xmm0\n"
>     "addps (%4), %%xmm0\n"
>     "movaps %%xmm0, (%0)\n"
OK. %0 should be an input operand.
>     // increment number of iterations
>     "addl $1, %1\n"
>     // advance pointers
>     // removing the following lines prevents the segfault
>     "addl $16, %3\n"
>     "addl $16, %4\n"
>     "addl $16, %0\n"
> 
>     "jnz calc\n"
> 
>     : "+r"(c), "+r"(iter), "+c" (n)
>     : "r"(a), "r"(b)
>    );
>    printf("iter:%d\n", iter); // STRANGE: always prints "0"
> }
> 

Perhaps:
     void add_asm (float *a, float *b, float *c, int n){
         int iter = 0;
         __asm__ __volatile__ (
             "1:\n\t"
             "cmpl %0, %1\n\t"
             "jae 2f\n\t"
             "movaps (%3,%0,4), %%xmm0\n\t"
             "addps (%4,%0,4), %%xmm0\n\t"
             "movaps %%xmm0, (%2,%0,4)\n\t"
             "addl $4, %0\n\t"
             "jmp 1b\n\t"
             "2:"
             : "+r" (iter)
             : "rm" (n), "r" (c), "r" (a), "r" (b)
         );
         /* equivalent to:
            for (iter = 0; (unsigned)iter < (unsigned)n; iter += 4){
                __asm__ __volatile__ (
                    "movaps %1, %%xmm0\n\t"
                    "addps %2, %%xmm0\n\t"
                    "movaps %%xmm0, %0"
                    : "o" (c[iter])
                    : "o" (a[iter]), "o" (b[iter])
                );
            }
          */
         iter >>= 2;
         printf ("iter:%d\n", iter);
     }