[Bug middle-end/57748] [4.7/4.8/4.9 Regression] ICE when expanding assignment to unaligned zero-sized array

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57748

--- Comment #33 from Richard Biener <rguenth at gcc dot gnu.org> ---
My point is that

> -      mode = TYPE_MODE (TREE_TYPE (tem));
> -      if (TREE_CODE (tem) == MEM_REF
> -         && mode != BLKmode
> -         && ((align = get_object_alignment (tem))
> -             < GET_MODE_ALIGNMENT (mode))
> -         && ((icode = optab_handler (movmisalign_optab, mode))

is wrong because it asks if we can do a V2DFmode (the mode of the
whole struct!) unaligned store which the backend says, yes!, and
it builds a V2DFmode store for us to use.

But the _access_ is V1DFmode!  The access is not to 'tem' but to 'to'!
Using the mode of 'tem' is just wrong to use for movmisalign.