They were accidentally using the 32-bit load/store instruction for 8/16-bit operations, due to incorrect patterns (8/16-bit cmpxchg and atomicrmw will be fixed in subsequent changes) llvm-svn: 270486