[Bug middle-end/41963] [4.5 Regression] 177.mesa in SPEC CPU 2K is miscompiled

["matz at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #4 from matz at gcc dot gnu dot org  2009-11-06 13:07 -------
It's a simple miscompilation of the cse_reciprocals pass that is exposed
by my change:

% cat pr41963.c
#include <math.h>

static __attribute__((noinline)) void f (float *dst, float *src)
{
  int i, j;
  for (i = 0; i < 2; i++)
    {
      float len;
      dst[0] = src[0];
      dst[1] = src[1];
      len = sqrtf (dst[0] * dst[0] + dst[1] * dst[1]);
      if (len > 0.5f)
        {
          len = 1.0f / len;
          dst[0] *= len;
          dst[1] *= len;
        }
    }
}

int main()
{
  float dst[2], src[2];
  src[0] = 2.0f;
  src[1] = 5.0f;
  f (dst, src);
  printf ("%f %f -> %f\n", dst[0], dst[1], dst[0] * dst[0] + dst[1] * dst[1]);
  return 0;
}

The reciprocal pass overwrites the statement
  len = sqrtf(temp)
  (and a later use of len in 1/len)
with
  len = rsqrtf(temp)
  (and removed the 1/len computation)
But it does this without caring for _any_ other use of len, e.g. in the
compare or the second loop iteration.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41963