Missed optimization in gcc 3.4?

[Martin Reinecke <martin at MPA-Garching dot MPG dot DE>]

Hi,

the piece of code below shows a case where the loop optimizer and/or inliner
of gcc 3.4 does a suboptimal job.


------- snip

class RNG
  {
  private:
    unsigned int x,y,z,w;

  public:
    /*! initializes the generator with 1 to 4 seed values. */
    RNG (unsigned int x1=123456789, unsigned int y1=362436069,
         unsigned int z1=521288629, unsigned int w1=88675123)
      : x(x1), y(y1), z(z1), w(w1)  {}


    unsigned int int_rand_uni()
      {
      unsigned int t = x^(x<<11);
      x = y;
      y = z;
      z = w;

      return w=(w^(w>>19))^(t^(t>>8));
      }

    void skip()
      {
      for (int m=0; m<100000000; ++m)
        int_rand_uni();
      }
  };

int main()
  {
  RNG rng;

#ifdef OUTER_LOOP

  for (int m=0; m<100000000; ++m)
    rng.int_rand_uni();

#else

  rng.skip();

#endif

  return rng.int_rand_uni()%255;
  }

------- snip

When compiling with -O3, the runtime of the resulting executable
is about three times as long when OUTER_LOOP is defined
than when it is undefined, despite the fact that int_rand_uni()
is inlined in both cases. Is there a way to generate the same,
efficient code in both cases, or is this an inherent limitation of
gcc's optimizer?

Thanks,
  Martin