This works around a bug in the branch prediction logic of the
CPU which can cause a hard hang. The existing fix-loongson2f-jump
works around the same bug, but by 'fixing' the jump target to be within the
same 256MB region as the kernel. This will (silently) cause a jump
to the wrong address if the jump is intentionally to some other
region (e.g. a call to pmon).
fix-loongson2f-btb instead works around the issue by adding an explicit flush
of the branch target buffer (via a write to a cop0 register) before every
jump register instruction.
