Using LDP/STP when accessing m_ppc_state lets us load/store two
registers at once. We previously opportunistically used STP, but this
new analysis lets us move loads earlier and move stores later to make
use of LDP/STP in more situations. This reduces code size and time spent
on m_ppc_state accesses, possibly with exceptions when under heavy
register pressure.
This commit adds the new bitsets load_pairs and store_pairs to
BlockRegStats, which indicate which registers should be treated as a
pair when loading and storing to m_ppc_state. The commits after this one
will add code that reads these bitsets to determine when to use LDP/STP.
x64 doesn't have instructions that can load/store a pair of registers,
so the new PPCAnalyst code is ifdef'd for AArch64.
