http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55882
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> 2013-01-08 15:47:10 UTC --- Ok with respect to alignment the issue is that when we expand_assignment for to == dmfe[i_1].use_alt_rd_dqs we fall into the old trap for STRICT_ALIGNMENT targets to use the mode alignment. That's because we fall into misalignp = false; to_rtx = expand_expr (tem, NULL_RTX, VOIDmode, EXPAND_WRITE); which creates (mem/c:BLK (plus:SI (reg/f:SI 189 virtual-stack-vars) (const_int 4 [0x4])) [0 dmfe+0 S992 A32]) as dmfe is 4-byte aligned. With offset = (sizetype) i_1 * 124 + 2 and bitregion_start = 0, bitregion_end = 63, bitpos = 48 and after adjusting the offset we feed (mem/c:BLK (plus:SI (reg/f:SI 206) (const_int 6 [0x6])) [0 dmfe A16]) into set_mem_attributes_minus_bitpos (with bitpos == 48) and get_object_alignment returns 32. But we fail to consider bitpos when using that alignment. So ... does the following patch fix it for you? Index: gcc/emit-rtl.c =================================================================== --- gcc/emit-rtl.c (revision 195014) +++ gcc/emit-rtl.c (working copy) @@ -1839,7 +1839,13 @@ set_mem_attributes_minus_bitpos (rtx ref if (!align_computed) { - unsigned int obj_align = get_object_alignment (t); + unsigned int obj_align; + unsigned HOST_WIDE_INT obj_bitpos; + get_object_alignment_1 (t, &obj_align, &obj_bitpos); + if (apply_bitpos) + obj_bitpos += apply_bitpos; + if (obj_bitpos != 0) + obj_align = (obj_bitpos & -obj_bitpos); attrs.align = MAX (attrs.align, obj_align); } }