I added a pessimization in r217102 to prevent miscompiles when the
incremented induction variable was used in a comparison; it would be
poison.
Try to use the incremented induction variable more often when we can be
sure that the increment won't end in poison.