EDIT 2: Editing...

EDIT 1: as spotted by @pbsd this was not working good with some values...

I have added a way of checking the overflow and I have tested it:

  #define BITS (8*sizeof(int))-1

  int checked_add(int a, int b, int *rp)
  {
      *rp = a+b;
      
      return (a^b) < 0 ? 0 : (a^(*rp)) < 0  ? 1 : 0) ;
  }

It is around 30% faster than the fastest in the blog and I believe the code is way cleaner. What we are doing is checking if all of the values has the same sign (the "^" is just a "xor"), if so, no overflow in other case, overflow.

This is the testing program (g++ compliant):

  #include <stdio.h>
  #include <time.h>
  #include <stdlib.h>

  #define BITS (8*sizeof(int))-1

  int checked_add2(int a, int b, int *rp) {
    uint ur = (uint)a + (uint)b;
    uint sr = ur >> (BITS-1);
    uint sa = (uint)a >> (BITS-1);
    uint sb = (uint)b >> (BITS-1);
    *rp = ur;
    return
      (sa && sb && !sr) ||
      (!sa && !sb && sr);
  }


  int checked_add(int a, int b, int *rp)
  {
      *rp = a+b;
      return ((a^b) | (a^(*rp)) < 0)  ? 1 : 0 ;
  }


  int main(int argc, char* argv[])
  {
      int a, b, c;
      long clong;
      srand(time(NULL));


      for(unsigned int i = 0; i < 50000000; ++i)
      {
          bool overflow = false;
          a = rand();
          b = rand();
          clong = (long)a + (long)b;

          overflow = checked_add(a,b,&c);

          if(clong != (long)c && !overflow)
              printf("Overflow not detected %i + %i = %i \n", a, b, c);

      }

      return 0;
  }

Time for "checked_add": 2.694s

Time for "checked_add2": 3.200s

This is the dump (6 instructions so far): http://goo.gl/cnBKdS