[Bug tree-optimization/68963] [4.9/5/6 Regression] O3 vs. O2 discards part of loop and terminates early

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68963

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
      Known to work|                            |4.7.4
           Keywords|                            |wrong-code
   Last reconfirmed|                            |2015-12-18
          Component|c                           |tree-optimization
                 CC|                            |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1
            Summary|O3 vs. O2 discards part of  |[4.9/5/6 Regression] O3 vs.
                   |loop and terminates early   |O2 discards part of loop
                   |                            |and terminates early
   Target Milestone|---                         |4.9.4
      Known to fail|                            |4.8.5, 4.9.3, 5.3.0, 6.0

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  This is complete peeling doing sth wrong:

test123.c:10:2: note: loop with 4 iterations completely unrolled
Last iteration exit edge was proved true.

We have bogus range info in .pcom:

  <bb 5>:
  # RANGE [0, 3] NONZERO 3
  # i_23 = PHI <0(4), i_20(9)>
  if (K_5(D) > i_23)
    goto <bb 6>;
  else
    goto <bb 7>;
...
  <bb 8>:
  # RANGE [1, 4] NONZERO 7
  i_20 = i_23 + 1;
  if (i_20 < _21)
    goto <bb 9>;
  else
    goto <bb 3>;

  <bb 9>:
  goto <bb 5>;

but disabling VRP1 doesn't help here (it uses niter analysis to derive the
above so niter analysis is wrong in some way).