[Bug tree-optimization/81403] [8 Regression] wrong code at -O3

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81403

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
So this is during partial-PRE insertion where after PRE insertion of

Found partial redundancy for expression {bit_not_expr,_13} (0020)
Inserted _33 = ~_3;
 in predecessor 5 (0014)
Created phi prephitmp_34 = PHI <_33(5), _8(6)>
 in block 7 (0020)

we value-replace 0020 by prephitmp_34 and thus during partial-PRE
phi-translation
find, when translating _14 & 10393, _8 for _14 (0020) to be translated across
the edge exposing the range.

On the other edge we get from translating _14 & 10393 all the way to
translating
var_9.5_12 = var_9 as operand which leads us to var_9.1_2 = var_9 and
ultimatively
to _8 via "simplification" of ~_3 which has a leader of _33 (fine IMHO), but
unfortunately _33 has a value number of _8.

So PRE doing any insertion of a lexically equivalent expression will
necessarily
end up with a SCCVN value-number that might have range info that is not valid
there.  Thus:

            tree result = vn_nary_op_lookup_pieces (newnary->length,
                                                    newnary->opcode,
                                                    newnary->type,
                                                    &newnary->op[0],
                                                    &nary);
            if (result && is_gimple_min_invariant (result))
              return get_or_alloc_expr_for_constant (result);

            expr = pre_expr_pool.allocate ();
            expr->kind = NARY;
            expr->id = 0;
            if (nary)
              {
                PRE_EXPR_NARY (expr) = nary;
                new_val_id = nary->value_id;
                get_or_alloc_expression_id (expr);
              }

needs adjustment to push away range-info.