[Bug tree-optimization/101641] Bogus redundant store removal

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101641

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-07-27
             Status|UNCONFIRMED                 |ASSIGNED
           Keywords|                            |wrong-code

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
We're already going great lengths trying to preserve stores that change the
dynamic type in a way relevant for downstream users (which we actually do not
see).  In particular we've settled to

          alias_set_type set = ao_ref_alias_set (&lhs_ref);
          alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref);
          if ((vnresult->set != set
               && ! alias_set_subset_of (set, vnresult->set))
              || (vnresult->base_set != base_set
                  && ! alias_set_subset_of (base_set, vnresult->base_set)))
            resultsame = false;

but this case is so degenerate that all the alias sets and base alias sets
of both pu->y (vnresult) and pu->x (lhs_ref) are 2 since we use the alias
set of the union type for all accesses that have the union in their path.

The only remaining chance we have here is to look at the actual types
and reject punning attempts.  We are really missing a way to preserve
the TBAA effect of a store but not the store itself.