[Mlir-commits] [mlir] [mlir] Add loaded URI attribute type (PR #67276)

[Mehdi Amini]

joker-eph wrote:

> Not sure I agree, I see this as no more specific than DenseElementsAttr.

There is history behind builtin, and it's pretty much admitted that we likely wouldn't have most of the things "builtin" if they were added in today's MLIR land. (we even looked into how to move tensor and memref to their own dialects!).

So the existence of something in builtin isn't a reason in itself to add more :)

Right now this looks like a nice convenient tool for specific cases (looks almost like it could be an "example" of using resources somehow), but too far from something "core enough" to be there.
Overall it's not clear to me that this is pulling its weight as an attribute in itself, instead of a more general mechanism (seems almost closer to an interface? But how do we handling the registrations of various loading mechanisms?). 

While landing code incrementally is preferred, I'd also want to understand the end picture more clearly (how we'll handle multiple schemes, etc.). Basically this raises more design questions to me than anything.



https://github.com/llvm/llvm-project/pull/67276