[Bug tree-optimization/37102] [4.3/4.4 Regression] possible integer codegen bug

[jakub at gcc dot gnu dot org]

------- Comment #3 from jakub at gcc dot gnu dot org  2008-08-14 10:02 -------
Somewhat shorter:
extern void abort (void);

unsigned int a, b = 1, c;

void __attribute__ ((noinline))
foo (int x)
{
  if (x != 5)
    abort ();
}

int
main ()
{
  unsigned int d, e;
  for (d = 1; d < 5; d++)
    if (c)
      a = b;
  a = b;
  e = a << 1;
  if (e)
    e = (e << 1) ^ 1;
  foo (e);
  return 0;
}

At *.tailc looks sane to me.
<bb 2>:
  pretmp.35_15 = c;
  b.1_5 = b;
  a_lsm.48_4 = a;
  if (pretmp.35_15 != 0)
    goto <bb 3>;
  else
    goto <bb 4>;

<bb 3>:

<bb 4>:
  # a_lsm.48_28 = PHI <b.1_5(3), a_lsm.48_4(2)>
  a = b.1_5;
  e_9 = b.1_5 << 1;
  if (e_9 != 0)
    goto <bb 5>;
  else
    goto <bb 6>;
so a_lsm.48_28 is either b or a, depending on whether c != 0 or not, and a = b.
Then copyrename4 decides to replace a = b.1_5; with a = a_lsm_48_5; and at
*.uncprop we have:
<bb 2>:
  pretmp.35_15 = c;
  a_lsm.48_5 = b;
  a_lsm.48_4 = a;
  if (pretmp.35_15 != 0)
    goto <bb 3>;
  else
    goto <bb 4>;

<bb 3>:

<bb 4>:
  # a_lsm.48_28 = PHI <a_lsm.48_5(3), a_lsm.48_4(2)>
  a = a_lsm.48_5;
  e_9 = a_lsm.48_5 << 1;
  if (e_9 != 0)
    goto <bb 5>;
  else
    goto <bb 6>;
which looks correct too.  But *.optimized is already wrong:
<bb 2>:
  a_lsm.49 = b;
  a_lsm.48 = a;
  if (c != 0)
    goto <bb 3>;
  else
    goto <bb 7>;

<bb 7>:
  a_lsm.49 = a_lsm.48;
  goto <bb 4>;

<bb 3>:

<bb 4>:
  a = a_lsm.49;
  e = a_lsm.49 << 1;
  if (e != 0)
    goto <bb 5>;
  else
    goto <bb 8>;
While in *.uncprop a was set unconditionally to b, now it is sometimes b and
sometimes a.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37102