Blackfin: dpmc: optimize hibernate/resume path

Blackfin: dpmc: optimize hibernate/resume path

The current save logic used in hibernation is to do a MMR load (base +
offset) into a register, and then push that onto the stack.  Then when
restoring, pop off the stack into a register followed by a MMR store
(base + offset).  These use plenty of 32bit insns rather than 16bit,
are pretty long winded, and full of pipeline bubbles.

So, by taking advantage of MMRs that are contiguous, the multi-register
push/pop insn, and register abuse, we can shrink this code considerably.

When saving, the new logic does a lot of loads into the data and pointer
registers before executing a single multi-register push insn.  Then when
restoring, we do a single multi-register pop insn followed by a lot of
stores.  Overall, this allows us to cut the insn count by ~30%, the code
size by ~45%, and drastically reduce the register hazards that trigger
bubbles in the pipeline.

Signed-off-by: Mike Frysinger <vapier@gentoo.org>