CODEWITHSUNDEEP
linuxassemblyx86-64cpu-registersatt
Larry
Larry

Reputation: 1765

Assembly - compiler will not see my inline changes?

Context :

Linux 64.

GCC 4.8.2 (with -O3 -march=native)

The x86_64 abi under my left hand, opened at page 21.

The C code :

int main (int argc, char ** argv) {

    int16_t h = atoi(argv[1]) ;
    int16_t p;

  __asm__ ("mov %2, %0\n\t"
    "rol $8,%1\n\t"          
     : "=r" (p)    /* output operands */
     : "0"(p),"g"(h)/* input operands */
     :"cc");        /* clobbered operands */

  printf("%d %d\n", h, p);
  return 0;

}

The assembly code (lines underlying the problem):

    ...
    movl    $10, %edx
    movq    8(%rsi), %rdi
    xorl    %esi, %esi
    call    strtol
    xorl    %edx, %edx
    movl    $.LC0, %edi
#APP
# 1627 "test2ptr.c" 1
    movl %ax, %dx          <- set in %dx
    rol $8,%dx

# 0 "" 2
#NO_APP
    movswl  %ax, %esi
    movswl  %dx, %edx <- Then this line should not appear
    xorl    %eax, %eax
    call    printf
    xorl    %eax, %eax
    ...

If I comment it, the result is fine.

But I cannot rely on modifying the source (unmaintainable : every time one change something in the source, one would have to come back in that spot to be sure it is still working.. no go).

The question :

Why is the line movswl %dx, %edx kept ?

It should move a long in a word. But it is already done by me and cost me one superfluous clock.

Is there any workaround ?

Is it just an option I didn't set ?

Thanks

Upvotes: 1

Views: 40

Answers (2)

Larry
Larry

Reputation: 1765

Ok,

So the workaround is to set int32_t and not int16_t.

Now the code is... 2 cpu cycles faster.

That is ridiculous.

But I like assembly so much now :)

Upvotes: 0

Jester
Jester

Reputation: 58762

Since you specified a 16 bit type, but printf expects 32 bit integers, your result needs to be sign extended which is what that code does. Nevertheless, if you used the proper format for printf both versions should produce identical output.

As usual, you don't need inline asm for a rotate, and also if you ever use a mov in inline asm chances are you are doing it wrong.

Upvotes: 1

Related Questions