Assembler:
- Improve pseudo_li implementation. Load from previous
  instances (if available and close enough), otherwise
  emit at end of current extended basic block. (End ==
  unconditional branch.) Break the block if the distance
  for the furthest load would get too large.
- Peephole? Could e.g. turn "ldr lr,[sp,#OFF]; mov pc,lr"
  into "ldr pc,[sp#OFF]", but then the LR save slot must
  be in the caller's frame not the callee's.

hipe_arm:
- Handle more non-trivial immediates in mk_li/mk_load/mk_store.

Floating point:
- Drop no_inline_fp. Implement FP ops as calls to C or ASM
  primops. All FP values passed by reference in memory.
  This should at least reduce consing costs.

Linear scan:
- Do not hardcode temp1/temp2/temp3. Instead just take three
  regs from (All\Fixed)\Params. (Ditto in PowerPC.)
