https://github.com/dcaballe commented: This is cool :) I was wondering if, in addition to the tensor vs memref semantics, we should take into account the speculatability of the nested operations? https://github.com/llvm/llvm-project/pull/108032