[Bug tree-optimization/108522] [12/13 Regression] ICE in tree-object-size when struct has VLA

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108522

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org
           Priority|P3                          |P2
   Target Milestone|---                         |12.3

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
At least for backports one option can be punt (give the constant maximum or
minimum) when variable length structures are involved, almost nobody uses them
(with the exception of Ada but then _FORTIFY_SOURCE=3 isn't on) and they are
quite a mess to support.
Otherwise, one needs to make sure to use TREE_OPERAND (t, 2) of COMPONENT_REF
if non-NULL.
On the above testcase one can see:
  _3 = &s.1_9->b{off: _1};
  _12 = __builtin_dynamic_object_size (_3, 1);
That {off: _1} in there means COMPONENT_REF's last operand isn't NULL, but is
SSA_NAME _1.  That contains something that should be used instead of
DECL_FIELD_OFFSET if specified, because in DECL_FIELD_OFFSET you'll see a
VAR_DECL that contained that value during gimplification, but isn't maintained
later on.
tree-object-size.cc currently uses byte_position, which is ok if last
COMPONENT_REF's operand is NULL, otherwise it should be
byte_from_pos (TREE_OPERAND (component_ref, 2), DECL_FIELD_BIT_OFFSET (field));
where field is TREE_OPERAND (component_ref, 1).