[Bug c++/68727] [5/6/7 regression] invalid offsetof expressions accepted

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68727

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This changed with r215070.
The problem with the move is that build_class_member_access_expr when
diagnosing this saw object_type of B, while finish_offsetof sees
((struct A *) (struct B *) 0B + (sizetype) *(long int *) (((struct B *)
0B)->_vptr.B + 18446744073709551592))->i
as expr, where object is INDIRECT_REF with type A.
But expr is unfortunately a result of folding, so trying to discover the
original type used in __builtin_offsetof (such as looking for NOP_EXPR of
null_pointer_node and using type from that) might be too unreliable.
Consider
struct A { int i[2]; };
struct B: virtual A { };
struct C: virtual A { };
__SIZE_TYPE__ a[] = { __builtin_offsetof (B, A::i[(__SIZE_TYPE__)(&((C
*)__null)->A::i)]) };
where there is both (B *) 0B and (C *) 0B in the expression.
Another option is to pass the original object (i.e. (B *) 0B in this case) to
finish_offsetof in addition to expr, but then we need to stick it into
OFFSETOF_EXPR as well.