[Bug target/59371] [5/6/7 Regression] Performance regression in GCC 4.8 and later versions.

[aldyh at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59371

Aldy Hernandez <aldyh at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |aldyh at gcc dot gnu.org

--- Comment #21 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
Assuming that changing the condition of the loop to !=, would fix the
regression, could we use range info to determine this?

For instance, right before the ivopts pass, we have:

  <bb 2> [15.00%]:
  # RANGE [0, 65535] NONZERO 65535
  _19 = (int) c_11(D);
  if (_19 > 0)
    goto <bb 3>; [85.00%]
  else
    goto <bb 7>; [15.00%]

  <bb 3> [12.75%]:

  <bb 4> [85.00%]:
  # p_20 = PHI <p_14(6), p_10(D)(3)>
  # i_21 = PHI <i_15(6), 0(3)>
  # x_22 = PHI <x_13(6), 0(3)>
  _1 = *p_20;
  x_13 = _1 + x_22;
  p_14 = p_20 + 4;
  # RANGE [0, 65535]
  i.1_2 = (unsigned short) i_21;
  # RANGE [0, 65535]
  _3 = i.1_2 + 1;
  i_15 = (short int) _3;
  # RANGE [-32768, 32767]
  _4 = (int) i_15;
  if (_4 < _19)
    goto <bb 6>; [85.00%]
  else
    goto <bb 5>; [15.00%]

The range for the max value `c' (_19) is [0..65535].  So if we know that the
induction variable `i' (_4) starts at the lower end of that (0), and we
increment the IV by 1, then we could always replace the < by a !=, right?  This
would avoid having to do loop versioning.

If this is acceptable, would we do this in the ivopts pass?  If so, I guess we
could start by explaining to the pass that _4 is an IV.  Because right now I
see that we determine that everything but _4 is an IV: i_21, i.1_2, _3, i_15.

Thoughts/hints greatly appreciated :).