[Bug ipa/113520] ICE with mismatched types with LTO (tree check: expected array_type, have integer_type in array_ref_low_bound)

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113520

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2024-01-22
          Component|tree-optimization           |ipa
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |jamborm at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
           Keywords|                            |lto

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hmm, we are supposed to "handle" this during in-streaming.  This seems to work
but then for some reason it gets fiddled with again...

Ah, so this is reading of IPA CP summaries where it seems that we do not
apply these "tricks", gimple-streamer-out.cc has

          /* Wrap all uses of non-automatic variables inside MEM_REFs
             so that we do not have to deal with type mismatches on
             merged symbols during IL read in.  The first operand
             of GIMPLE_DEBUG must be a decl, not MEM_REF, though.  */
          if (!flag_wpa && op && (i || !is_gimple_debug (stmt)))
            {
              basep = &op;
              if (TREE_CODE (*basep) == ADDR_EXPR)
                basep = &TREE_OPERAND (*basep, 0);
              while (handled_component_p (*basep))
                basep = &TREE_OPERAND (*basep, 0);
              if (VAR_P (*basep)
                  && !auto_var_in_fn_p (*basep, fn->decl)
                  && !DECL_REGISTER (*basep)) 
                {
                  bool volatilep = TREE_THIS_VOLATILE (*basep);
                  tree ptrtype = build_pointer_type (TREE_TYPE (*basep));
                  *basep = build2 (MEM_REF, TREE_TYPE (*basep),
                                   build1 (ADDR_EXPR, ptrtype, *basep),
                                   build_int_cst (ptrtype, 0));
                  TREE_THIS_VOLATILE (*basep) = volatilep;
...

and gimple-streamer-in.cc undoes this when the prevailing decls are compatible:

          /* At LTO output time we wrap all global decls in MEM_REFs to
             allow seamless replacement with prevailing decls.  Undo this
             here if the prevailing decl allows for this.
             ???  Maybe we should simply fold all stmts.  */
          if (TREE_CODE (*opp) == MEM_REF
              && TREE_CODE (TREE_OPERAND (*opp, 0)) == ADDR_EXPR
              && integer_zerop (TREE_OPERAND (*opp, 1))
              && (TREE_THIS_VOLATILE (*opp)
                  == TREE_THIS_VOLATILE
                       (TREE_OPERAND (TREE_OPERAND (*opp, 0), 0)))
              && !TYPE_REF_CAN_ALIAS_ALL (TREE_TYPE (TREE_OPERAND (*opp, 1)))
              && (TREE_TYPE (*opp)
                  == TREE_TYPE (TREE_TYPE (TREE_OPERAND (*opp, 1))))
              && (TREE_TYPE (*opp)
                  == TREE_TYPE (TREE_OPERAND (TREE_OPERAND (*opp, 0), 0))))
            *opp = TREE_OPERAND (TREE_OPERAND (*opp, 0), 0);

I suppose we might want to split these out so summary streaming can apply
this to streamed trees as well?