When I worked on GPU compiler backends the middle and backend passes basically were a collection of more basic passes that did one thing, in sequence. I wouldn't call this an alternative architecture. I think it's just an implementation detail of the dominant architecture.