[Bug middle-end/80342] New: useless outermost conversions not fully elided by genmatch generated code

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80342

            Bug ID: 80342
           Summary: useless outermost conversions not fully elided by
                    genmatch generated code
           Product: gcc
           Version: 7.0.1
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: rguenth at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

if you write

(simplify
 (plus @1 @2)
 (convert (plus @1 @2)))

then genmatch generates

    *res_code = NOP_EXPR;
    {
      tree ops1[2], res;
      ops1[0] = captures[0];
      ops1[1] = captures[1];
      code_helper tem_code = PLUS_EXPR;
      tree tem_ops[3] = { ops1[0], ops1[1] };
      gimple_resimplify2 (lseq, &tem_code, TREE_TYPE (ops1[0]), tem_ops,
valueize);
      res = maybe_push_res_to_seq (tem_code, TREE_TYPE (ops1[0]), tem_ops,
lseq);
      if (!res) return false;
      res_ops[0] = res;
    }
    gimple_resimplify1 (lseq, res_code, type, res_ops, valueize);

which means the (plus @1 @2) is pushed to seq even if the conversion ends up
being simplified away.  This results in forwprop changes like

+  _1 = x_6(D) - _10;
+  _7 = _1;

and also in missed optimizations in SCCVN which uses seq == NULL and thus
won't survive this.

Individual patterns can be adjusted to provide special-casing for the
case of useless conversions but a genmatch solution is prefered.  Note
it won't come cheap (it will duplicate the transform code).