[Bug c/12243] New: loop optimization incorrectly treats expression as constant

["selinger at mathstat dot uottawa dot ca" <gcc-bugzilla at gcc dot gnu dot org>]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12243

           Summary: loop optimization incorrectly treats expression as
                    constant
           Product: gcc
           Version: 3.3.1
            Status: UNCONFIRMED
          Severity: critical
          Priority: P2
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: selinger at mathstat dot uottawa dot ca
                CC: gcc-bugs at gcc dot gnu dot org,selinger at mathstat dot
                    uottawa dot ca
  GCC host triplet: any

The code below should terminate in the second loop iteration. When compiled with
-O2 or -O3 on certain architectures (see below), it enters an infinite loop
instead. Inspection of the generated assembly code reveals that the expression
marked (***) is wrongly treated as a constant expression and moved out of the loop.

Note that this is not a "floating point bug" in the usual sense: all floating
point numbers in this example are in fact small integers, and round-off errors
do not occur. Nevertheless, this particular bug seems to occur only with
floating point numbers.

Compiler versions where I have reproduced the bug:
   gcc version 3.3.1 (host: linux-i386, --target=sparc-sun-solaris2)
   gcc version 2.96 20000731 (Red Hat Linux 7.2 2.96-112.7.1)
   gcc version 2.95.4 20011002 (Debian prerelease)

I assume that this is an architecture independent bug which has been present
since at least 2.95.4; however, this particular sample code seems to reproduce
the bug only on certain architecture/compiler combinations. For example, I could
not reproduce it on gcc 3.3.1/i386 or gcc 2.95.4/sparc.

How to reproduce: gcc -O3 sample.c

Code: (no #include's are necessary)

int z=0;

int main() {
  double Q[2][2];
  double det;
  int i, j;
  double x;

  /* to the zero 2x2-matrix, keep adding the identity matrix until the
     determinant is non-zero */

  if (z == 0) {
    for (i=0; i<2; i++) {
      for (j=0; j<2; j++) {
        Q[i][j] = 0;
      }
    }
  }

  while(1) {

    det = Q[0][0]*Q[1][1] - Q[0][1]*Q[1][0];  /* (***) */
    if (det != 0.0) {
      x = 0;
      break;
    }

    for (i=0; i<2; i++) {
      for (j=0; j<2; j++) {
        Q[i][j] += (i==j ? 1 : 0);
      }
    }
  }

  return 0;
}