[Mlir-commits] [mlir] [MLIR][Tensor] Canonicalize fully covering slice insertions into tensors with unit prefixes (PR #92912)

[Stella Laurenzo]

stellaraccident wrote:

> (I'm happy to be told I'm wrong on any of these statements, but I have seen a rebuttal in the thread so far!). We're pretty far from anything "excessive" here!

Minor correction: this thread is full of rebuttals from three different contributors in good standing to this dialect, but they are not passing your bar for listening and letting the conversation transpire. 

I jumped in here as a circuit breaker to the form the discussion was taking because I receive strong feedback from the community that when it turns this way, it is incredibly damaging, and that outcome is the opposite of what we all want. I've been guilty of this as many times as anyone -- just trying to make us better. 

Once we got out of rapid fire mode, I got the sentiment in back channels that it seems past time to write down the canonicalization approach for this dialect. There is a logic to it, but it kind of comes out of history and precedence (and not everyone can produce a fully formed, first principles argument in one go on a topic like this). Mahesh's original response was kind of extending this common law to provide guidance. If the discussion had continued at that point, versus diverting into a more rhetorical form, the other contributors would have had time to see it and likely have come to a similar consensus that writing down more formal guidelines is needed. 

The points and insights you are bringing up would enrich such a thing but are really hard to litigate on a very low level point like this specific one. 

The one thing I would ask: this is a volunteer project, and engaging to produce a well reasoned approach to something like this is real work. It takes time and can't always be formulated on demand in an adversarial/debate like forum on a single issue like this in real time. The choice to step back and consider is something that the contributors in good standing to the component have to be able to make, and that is what I'm intervening to make sure can happen here.

https://github.com/llvm/llvm-project/pull/92912