[Mlir-commits] [mlir] [mlir] load dialects for non-namespaced attrs (PR #94838)

[Oleksandr Alex Zinenko]

ftynse wrote:

Yeah, we shouldn't just load all available dialects. This defies the point of separating registration from loading. 

It is unclear to me why `#attr = #test.nested_polynomial<<1 + x**2>>` is the accepted syntax. Looks ambiguous. From looking at it in isolation, it I expect that this is just a `NestedPolynomialAttr` from the test dialect with `<1 + x**2>` being its syntax. IMO, we shouldn't elide the attribute mnemonic nor the dialect prefix here at all (why/where do we?) for the nested attribute. We could elide the dialect prefix if the attribute is immediately nested in another attribute from the same dialect, but the rest sounds ambiguous to me.

https://github.com/llvm/llvm-project/pull/94838