More on gimplify:create_tmp_var

[Richard Kenner]

    The point I was making earlier in my message was that we should be smart
    enough to know that given

      R* pr = &p.r;
      R* qr = &q.r;

Yes, but in general, you can't compute those addressing operations due to
misalignment.

Also note that we have to able to propagate this down to RTL in MEM_EXPR, so
we need to keep the chain intact.  Accurate aliasing information is critical
to insn scheduling, something we'll always being doing at the RTL level.