[Bug rtl-optimization/27671] [4.2 Regression] optimization error on pentium4-Linux with %, regression from gcc-4.1.0

[kazu at gcc dot gnu dot org]

------- Comment #2 from kazu at gcc dot gnu dot org  2006-05-20 04:57 -------
The whole problem is reduced to the combiner doing wrong simplification
of XOR.  Here are some details.

Consider:

extern void abort (void) __attribute__ ((noreturn));
extern void exit (int) __attribute__ ((noreturn));

static int __attribute__((noinline))
foo (int a, int b)
{
  int c = a ^ b;
  if (c == a)
    abort ();
}

int
main (void)
{
  foo (0, 1);
  exit (0);
}

Note that

  c == a  <==>  a ^ b == a
          <==>  b == 0

So foo aborts if and only if b == 0.  In the above testcase, b gets 1,
so foo shouldn't abort, but it does.

The combiner tries to combine the following two insns.  (I'm omitting
clobbers for brevity.)

(set (reg:SI 61)
     (xor:SI (reg/v:SI 60 [ b ])
             (reg/v:SI 59 [ a ])))

(set (reg:CCZ 17 flags)
     (compare:CCZ (reg:SI 61)
                  (reg/v:SI 59 [ a ])))

The combiner produces

(set (reg:CCZ 17 flags)
     (compare:CCZ (reg/v:SI 59 [ a ])
                  (const_int 0 [0x0])))

Note that this is wrong.  What matters is whether b is 0 or not.

Once the problem is reduced to this level,
-march=pentium4 is no longer needed.
Is is just that with -march=pentium4, "k % 2" is expanded using XOR.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27671