[Bug middle-end/71474] New: PRED_LOOP_IV_COMPARE wrongly calculates number of iterations of a loop

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71474

            Bug ID: 71474
           Summary: PRED_LOOP_IV_COMPARE wrongly calculates number of
                    iterations of a loop
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: marxin at gcc dot gnu.org
  Target Milestone: ---

Hello.

During investigation of poor predictor 'loop iv compare' heuristics, I've come
to a problematic test-case:

subroutine rubber(arg1, arg2)

integer k,l
real*8 arg1(3,3,3), arg2(3,3,3)

  do l=1,3
     do k=1,l
        arg1(k,l,1) = arg2(k,l,1)
     enddo
  enddo

end

042t.profile_estimate gimple looks as follows:

  <bb 3>:
  # l_13 = PHI <1(2), l_28(8)>
  if (l_13 > 0)
    goto <bb 4>;
  else
    goto <bb 6>;

  <bb 4>:
  # k_6 = PHI <1(3)>

  <bb 5>:
  # k_12 = PHI <k_6(4), k_26(9)>
  _1 = (integer(kind=8)) l_13;
  _2 = _1 * 3;
  _3 = (integer(kind=8)) k_12;
  _4 = _2 + _3;
  _5 = _4 + -4;
  _11 = *arg2_22(D)[_5];
  *arg1_23(D)[_5] = _11;
  k_26 = k_12 + 1;
  if (k_12 == l_13)
    goto <bb 6>;
  else
    goto <bb 9>;

  <bb 9>:
  goto <bb 5>;

  <bb 6>:
  l_28 = l_13 + 1;
  if (l_13 == 3)
    goto <bb 7>;
  else
    goto <bb 8>;

  <bb 8>:
  goto <bb 3>;

  <bb 7>:
  return;

Problem is that is_comparison_with_loop_invariant_p is called for the gimple
compare in BB 3 (if (l_13 > 0)),
which does not determine a number of iterations of the loop. As a result we get
confusing loop_bound_step, loop_iv_base
that are eventually used by predict_iv_comparison to calculate the misleading
probability:

  loop iv compare heuristics of edge 3->4: 0.0%

While the proper value should be 100% as l_13 is always a positive number.