[Mlir-commits] [mlir] [mlir] Introduction of LocalEffectsOpInterface (PR #130341)

[Uday Bondhugula]

bondhugula wrote:

> I don't quite understand why we need a new interface and why we can't express this through the existing side-effect interface?

The meaning of “local effect” is that the underlying resource can possibly only be accessed by this op and only this op. The current design in MLIR doesn’t capture this info. Side effect interface and the `Resource` definition doesn't distinguish between local or non-local. 

One can potentially extend `Resource` to add a flag (specify the default as well), but that's just another way of saying "local effects" and it would force ops to declare and specify an abstract default resource and an abstract effect(s). Instead, we could implicitly assume that the default resource/effects for all non-pure ops that don't implement memory effect interface is "non-local" effects. They can implement the "local effects" interface to say otherwise.

https://github.com/llvm/llvm-project/pull/130341