[Mlir-commits] [mlir] [mlir][mesh, MPI] Mesh2mpi (PR #104566)

[Frank Schlimbach]

fschlimb wrote:

I realized that I had to fix a few things in mesh before this can be useful. These fixes have been merged now (#114238).
The branch got rebased to be on top these changes.
I also added a few more things here:
* lowering of  LinearIndex, MultiIndex and NeighborsIndex to MPI
* if a `memref.global` `static_mpi_rank` exists the provided (static) value will be usedas MPI_rank. This allows (static) shape propagation which is a requirement for all kinds of optimizations. This could also become a conoicalization pattern for `CommRankOp` if that's preferred.
* canonicalization fo send/recv, again to allow static shape propagation

@yaochengji @tkarna @AntonLydike @sogartar @mfrancio could you have a look (again) please?


https://github.com/llvm/llvm-project/pull/104566