Can linalg represent the loop of the reduced domain?

Recently, I read the relevant code and design concept documents of linalg dialect, and linalg cannot express padding. But there is still a question: How should linalg represent multiple nested loops, in other words, how to represent a combination of multiple linalg.generic? Judging from the current code, this is only the case for Softmax in linalg. It seems that there is no formal solution for this situation yet, or is it still under discussion? As for me, my current practical need is to treat this situation as a whole operator and tile and fuse this whole in the producer-consumer link. For specific requirements, see: How to extend linalg ops without modifying MLIR code?