[Mlir-commits] [mlir] mlir::mesh::shardingOp adding shard-size control (PR #98145)

[Frank Schlimbach]

fschlimb wrote:

@sogartar You're probably right and we can go without the the `force` attribute. I am pretty sure I can handle all update_halo cases without that if I move the generation of `mesh.update_halo` into the `spmdize` method for inplace operations like `insert_slice`. An extra optimization pass should be able to get rid of extra `update_halo` ops, e.g. when the output of one is the input of the other.
I removed `force` from `mesh.shard` and adapted related code/tests.

I also added a check in a symbolverifier to disallow sharded_dims_sizes on dynamic meshes.

As far as I can tell I addressed all your concerns and suggestions.

https://github.com/llvm/llvm-project/pull/98145