[Mlir-commits] [mlir] [mlir] Add optional layout attribute to VectorType (PR #71916)

[Nicolas Vasilache]

nicolasvasilache wrote:

This needs its own RFC, I too would prefer to see something less intrusive with room for evolution.

One possible path forward is to start from something like a FragmentedVectorType that is not automatically a builtin VectorType and interfaces that define the semantics of such fragments. The existing VectorType is a particular limit case with a single fragment.

Over time, ops, patterns and transformations can be ported to `VectorTypeOrFragmentedVectorType`.
At some point in the future, if we get to full convergence great, if not that's fine too: it is unclear that everything makes sense to have as a fragmented vector.

One big question is whether the number of fragments and the data paths to convert between different such "layouts" are statically known but the system should be designed such that we can ask this question later (e.g. if `?` quantities in vectors ever become a thing).

https://github.com/llvm/llvm-project/pull/71916