This is a special case because one node maps to two different G_
instructions, and the operand order is changed.
This mostly enables G_FCMP for AMDPGPU. G_ICMP is still manually
selected for now since it has the SALU and VALU complication to deal
with.
llvm-svn: 370280