Unaligned block moves and MEM_ALIGN re-broken

[law@redhat.com]

 In message <10201252000.AA13924@vlsi1.ultra.nyu.edu>, Richard Kenner writes:
 > We do indeed require that an object of an integral mode be at an alignment
 > defined by GET_MODE_ALIGNMENT.
 > 
 > The only time when this isn't true is when STRICT_ALIGNMENT is not defined.
 > but in that case, why does the alignment matter?
But STRICT_ALIGNMENT means that all objects must be on a natural
alignment boundary.  The lack of STRICT_ALIGNMENT does _NOT_ mean that
data can be access irrespective of alignment concerns.

It's perfectly legitimate and normal for some alignment to be necessary on
a target without STRICT_ALIGNMENT.  For example, a DFmode item may need to
have a 32bit alignment on a target without STRICT_ALIGNMENT.

jeff