https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88963
--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue RTL expansion spills is the inner MEM_REF is expanded to
(mem:BLK (reg/v/f:DI 88 [ a ]) [1 *a_7(D)+0 S64 A256])
and somehow
if (MEM_P (to_rtx))
{
/* If the field is at offset zero, we could have been given the
DECL_RTX of the parent struct. Don't munge it. */
to_rtx = shallow_copy_rtx (to_rtx);
set_mem_attributes_minus_bitpos (to_rtx, to, 0, bitpos);
if (volatilep)
doesn't transfer the smaller mode of the BIT_FIELD_REF to to_rtx. It might
be that we can use TYPE_MODE (TREE_TYPE (to)) in place of mode1 here:
gcc_checking_assert (known_ge (bitpos, 0));
if (optimize_bitfield_assignment_op (bitsize, bitpos,
bitregion_start, bitregion_end,
mode1, to_rtx, to, from,
reversep))
result = NULL;
else
result = store_field (to_rtx, bitsize, bitpos,
bitregion_start, bitregion_end,
mode1, from, get_alias_set (to),
nontemporal, reversep);
similar to how expand_expr tries to do that for bitfield extraction. Otherwise
the code has no idea it should use vector stores when later running into
store_bit_field_1 with a VOIDmode fieldmode where it ends up doing
store_integral_bit_field.
