[Bug tree-optimization/71661] [7 Regression] wrong code at -O3

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71661

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This to me looks like some transformation of a loop that doesn't properly
adjust the number of iterations.  During vrp1 loop2 is:
loop_2 (header = 11, latch = 13, niter = c_23 + 1 <= 3 ? 2 - c_23 : 0,
upper_bound = 3, likely_upper_bound = 3)
{
  bb_11 (preds = {bb_5 bb_13 }, succs = {bb_6 })
  {
    <bb 11>:
    # c_5 = PHI <c_23(5), c_29(13)>
    # d_7 = PHI <d_24(5), d_8(13)>
    # .MEM_11 = PHI <.MEM_25(5), .MEM_12(13)>
    # .MEM_18 = VDEF <.MEM_11>
    a = 2;
    c_19 = c_5 + 1;

  }
  bb_6 (preds = {bb_11 }, succs = {bb_13 bb_7 })
  {
    <bb 6>:
    # c_6 = PHI <c_19(11)>
    # d_8 = PHI <d_7(11)>
    # .MEM_12 = PHI <.MEM_18(11)>
    if (c_6 <= 2)
      goto <bb 13>;
    else
      goto <bb 7>;

  }
  bb_13 (preds = {bb_6 }, succs = {bb_11 })
  {
    <bb 13>:
    c_29 = ASSERT_EXPR <c_6, c_6 <= 2>;
    goto <bb 11>;

  }
}

after dce it is:
loop_2 (header = 6, latch = 6, niter = , upper_bound = 3, likely_upper_bound =
3)
{
  bb_6 (preds = {bb_5 bb_6 }, succs = {bb_6 bb_7 })
  {
    <bb 6>:
    # c_5 = PHI <0(5), c_19(6)>
    # d_7 = PHI <d_24(5), d_7(6)>
    # .MEM_11 = PHI <.MEM_25(5), .MEM_18(6)>
    # .MEM_18 = VDEF <.MEM_11>
    a = 2;
    c_19 = c_5 + 1;
    if (c_19 <= 2)
      goto <bb 6>;
    else
      goto <bb 7>;

  }
}
so again reasonable.
but ch pass turns this into:
loop_2 (header = 11, latch = 7, niter = , upper_bound = 3, likely_upper_bound =
3)
{
  bb_7 (preds = {bb_8 bb_11 }, succs = {bb_11 })
  {
    <bb 7>:
    # c_9 = PHI <0(8), c_19(11)>
    # d_25 = PHI <d_20(8), d_7(11)>
    # .MEM_24 = PHI <.MEM_18(8), .MEM_18(11)>

  }
  bb_11 (preds = {bb_7 bb_2 bb_5 }, succs = {bb_7 bb_8 })
  {
    <bb 11>:
    # c_5 = PHI <c_9(7), 0(2), 0(5)>
    # d_7 = PHI <d_25(7), d_16(D)(2), 0(5)>
    # .MEM_11 = PHI <.MEM_24(7), .MEM_15(2), .MEM_10(5)>
    # .MEM_18 = VDEF <.MEM_11>
    a = 2;
    c_19 = c_5 + 1;
    if (c_19 <= 2)
      goto <bb 7>;
    else
      goto <bb 8>;

  }
  bb_8 (preds = {bb_11 }, succs = {bb_7 bb_10 })
  {
    <bb 8>:
    d_20 = d_7 + 1;
    if (d_20 <= 1)
      goto <bb 7>;
    else
      goto <bb 10>;

  }
}

(basically collapses 2 loops into one), and the upper_bound/likely_upper_bound
is just wrong here, because the loop latch is executed exactly 5 times
if the loop is entered from bb 5 (that is the case at runtime), or 2 times + UB
(if entered from bb 2) - with c_19 being 1, 2, 3, 1, 2.