* Make sure that only wired TNs are allocated to stack locations
until the final pass.
* When using loop depth to prioritise TNs, find them by iterating over
the list of TNs, instead of reprocessing global TNs that appear in
multiple IR2 blocks.
* In the final pass that allocates leftover TNs (mostly to the stack),
make sure that any TN with a negative spill cost (i.e., the score
says it's actually useful to spill it) is allocated a stack slot.