[Bug tree-optimization/100756] [12/13 Regression] vect: Superfluous epilog created on s390x

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100756

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2022-10-21
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
            Summary|vect: Superfluous epilog    |[12/13 Regression] vect:
                   |created on s390x            |Superfluous epilog created
                   |                            |on s390x
                 CC|                            |amacleod at redhat dot com
   Target Milestone|---                         |12.3
           Keywords|                            |missed-optimization

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think the issue is that we now have

  <bb 2> [local count: 118111600]:
  _15 = n_8(D) * 4;
  if (n_8(D) > 0)
    goto <bb 5>; [89.00%]
  else
    goto <bb 7>; [11.00%]

while we probably had

  <bb 2> [local count: 118111600]:
  _15 = n_8(D) * 4;
  if (_15 > 0)
    goto <bb 5>; [89.00%]
  else
    goto <bb 7>; [11.00%]

before the change.  Loop header copying applies VN to the copied blocks:

Processing block 0: BB6
Value numbering stmt = i_9 = PHI <0(2)>
Setting value number of i_9 to 0 (changed)
Replaced redundant PHI node defining i_9 with 0
Value numbering stmt = result_14 = PHI <0(2)>
Setting value number of result_14 to 0 (changed)
Replaced redundant PHI node defining result_14 with 0
Value numbering stmt = _15 = n_8(D) * 4;
Setting value number of _15 to _15 (changed)
Making available beyond BB6 _15 for value _15
Value numbering stmt = if (_15 > i_9)
Recording on edge 6->7 _15 gt_expr 0 == true
Recording on edge 6->7 _15 le_expr 0 == false
Recording on edge 6->7 _15 ne_expr 0 == true
Recording on edge 6->7 _15 ge_expr 0 == true
Recording on edge 6->7 _15 lt_expr 0 == false
Recording on edge 6->7 _15 eq_expr 0 == false
marking outgoing edge 6 -> 7 executable
gimple_simplified to if (n_8(D) > 0)

with

  <bb 2> [local count: 118111600]:
  _15 = n_8(D) * 4;
  if (n_8(D) > 0)
    goto <bb 3>; [89.00%]
  else
    goto <bb 4>; [11.00%]

  <bb 3> [local count: 955630225]:
  # i_16 = PHI <i_13(3), 0(2)>
  # result_17 = PHI <result_12(3), 0(2)>
  _1 = (long unsigned int) i_16;
  _2 = _1 * 4;
  _3 = a_11(D) + _2;
  _4 = *_3;
  result_12 = _4 + result_17;
  i_13 = i_16 + 1;
  _5 = n_8(D) * 4;
  if (_5 > i_13)
    goto <bb 3>; [89.00%]
  else
    goto <bb 4>; [11.00%]

so it was a single use in the compare (because CSE only later introduces
more uses through DOM).

The niter code then ends up with maybe-zero as _15 <= 0 and a condition
of n_8(D) > 0 it tries to simplify with tree_simplify_using_condition
(called from simplify_using_initial_conditions).  That old machinery
would be a perfect candidate to be rewritten using path ranger, but
in a somewhat extended mode that can "skip" diamonds, aka, the path
just contains dominators of the loop entry edge on which we want to
evaluate the _15 <= 0 condition.

To make the old simplification code work we can do the following:

diff --git a/gcc/tree-ssa-loop-niter.cc b/gcc/tree-ssa-loop-niter.cc
index 1e0f609d8b6..4ffcef4f4ff 100644
--- a/gcc/tree-ssa-loop-niter.cc
+++ b/gcc/tree-ssa-loop-niter.cc
@@ -2216,6 +2216,7 @@ expand_simple_operations (tree expr, tree stop,
hash_map<tree, tree> &cache)

     case PLUS_EXPR:
     case MINUS_EXPR:
+    case MULT_EXPR:
       if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (expr))
          && TYPE_OVERFLOW_TRAPS (TREE_TYPE (expr)))
        return expr;

but that can of course have unintended side-effects elsewhere
(this function is also used by IVOPTs).