If we didn't do this, then the CSE generator would have to anticipate
that the CSE compute instruction would need to depend on more inames
(those unused hw axes) than actually declared--because previously
innocuous-seeming inames could suddenly turn out to be hardware axes.