[Mlir-commits] [mlir] [mlir][python] allow upstream dialect registration (PR #74252)

[Stella Laurenzo]

stellaraccident wrote:

> As a vague design suggestion, I would find it convenient to have `from mlir.dialects import foo` to have `foo` loaded in whatever context I create afterwards and `from mlir.dialects.foo import parser` to have `foo` registered (but not loaded) in whatever context I create afterwards + have a mechanism to override this behavior should we need a clean context for some reason. But maybe it's too magic for other folks.

I think the issue is that we use this in various jit and multi-client flows where "after" has no defined meaning -- it is all concurrent. I'm certainly not against doing something more ergonomic for this class of things but would like to avoid non-deterministic behavior in concurrent situations.

https://github.com/llvm/llvm-project/pull/74252