[Bug tree-optimization/78189] [5/6/7 Regression] movaps generated for unaligned store in aligned struct, when struct is referenced via unaligned member.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78189

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2016-11-03
      Known to work|                            |4.9.4
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
   Target Milestone|---                         |5.5
            Summary|movaps generated for        |[5/6/7 Regression] movaps
                   |unaligned store in aligned  |generated for unaligned
                   |struct, when struct is      |store in aligned struct,
                   |referenced via unaligned    |when struct is referenced
                   |member.                     |via unaligned member.
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Mine.  Looks like we are confused by the MEM_REF offset:

MEM[(struct B *)misalignedPtr_1(D) + -8B].a.a

DR_BASE_ADDRESS is misalignedPtr_1, DR_INIT (used as 'misalign') is 16(OVF)
and DR_ALIGNED_TO is 128.  That looks good but then we go:

  /* To look at alignment of the base we have to preserve an inner MEM_REF
     as that carries alignment information of the actual access.  */
  base = ref;
  while (handled_component_p (base))
    base = TREE_OPERAND (base, 0);
  if (TREE_CODE (base) == MEM_REF)
    base = build2 (MEM_REF, TREE_TYPE (base), base_addr,
                   build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)), 0));
  unsigned int base_alignment = get_object_alignment (base);

and that computes an aligned base (but doesn't factor in the misalignment
from the MEM_REF offset we just stripped).