Thanks for summarizing (and organizing) that round table, @rengolin !
Some more details on the named ops topic: One question was could we retain the semantics/idiom of named linalg ops that lower into a graph of linalg.generic ops, e.g. linalg.softmax.
One possible solution might be to encapsulate the lowered generic ops within the body of some isolated-from-above op, making it hard for e.g. cse to break the idiom these ops form while facilitating moving/lowering/optimizing them as a whole (kind of a local, ad-hoc alternative for function outlining). Here’s how such an op might look if added to the scf dialect.
Other uses for ad-hoc grouping of ops might be limiting transformation scopes (as proposed by the PR Renato mentioned) and as the result of partitioning.