Differentiate between word and subword memory operations as they take different amount of cycles to complete. This just adds a basic model of the subword latency to the scheduler. llvm-svn: 266898
