branch_predictor.c

#if 0
In these examples, let's assume that passing more than 2 args to the program is 
*not* the case we are optimizing this for.

With that in mind, lets build a sample program with
 and without __builtin_expect and compare assembly output. 
#endif

/*
   no predictor:

sub    $0x8,%rsp
cmp    $0x3,%edi
jle    0x4004e8 <main+24>   ; <===== fail
int3                        ; <===== fail
xor    %eax,%eax
add    $0x8,%rsp
retq   
nopl   0x0(%rax)
mov    $0x4005ec,%edi
callq  0x4003c0 <puts@plt>
xor    %eax,%eax
add    $0x8,%rsp
retq

*/

#include <stdio.h>

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

  if (argc > 3)
    asm("int $0x3\n");
  else
    printf("yo dog\n");

  return 0;
}


===

/* 
  with predictor:

sub    $0x8,%rsp
cmp    $0x3,%edi
jg     0x4004ea <main+26> ; <==== win
mov    $0x4005dc,%edi     ; <==== win
callq  0x4003c0 <puts@plt>
xor    %eax,%eax
add    $0x8,%rsp
retq   
int3   
jmp    0x4004e3 <main+19>

*/

#include <stdio.h>

#define unlikely(x)     __builtin_expect((x),0)

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

  if (unlikely(argc > 3))
    asm("int $0x3\n");
  else
    printf("yo dog\n");

  return 0;
}