[Mlir-commits] [mlir] [vector][mlir] Canonicalize to shape_cast where possible (PR #140583)

[Andrzej Warzyński]

banach-space wrote:

Returning after some OOO + sick leave. I wanted to clarify my earlier feedback.

> I'm sorry, but this response is a bit confusing to me. I'm very aware of the previous prs by James, because James and I are aligned on what we are working towards. 

**On context:** @Groverkss ,I wasn’t suggesting _you_ weren’t aware of previous PRs. My intent was to provide context for _other reviewers_ (including, but not limited to you) so they could see that this work is part of a longer effort, not an isolated change. And where my support was coming from.

> In fact, i'm more confused that we talk about context here, given that the same change was reverted in the past and I distinctly remember we never came to conclusion what is the right form here: https://github.com/llvm/llvm-project/pull/72918 . I do not see that we resolved any of the concerns in that discussion, so I don't think concluding this is beneficial is correct, without previous context.

**On the 2023 concerns:** The PR you linked was ~2 years ago. It’s not clear to me whether the issues raised there still apply. If they do, could someone summarize them for us here? The earlier discussions:
*  https://github.com/llvm/llvm-project/pull/72105, 
* https://github.com/llvm/llvm-project/pull/72918
* https://discourse.llvm.org/t/improving-handling-of-unit-dimensions-in-the-vector-dialect?

involved many people and spanned months, but there’s been little follow-up since. That suggests (to me) that the concerns may no longer be relevant, especially given the changes in the ecosystem that Diego pointed out.

> The RFC referenced above is also a good example of what can happen when we pull too many voices into a discussion.
> (...)
>  I know not everyone agrees with this but the Tensor WG was built with the intent to improve on this situation and help build certain level of consensus before having a broader discussion.

**On process:** I share the worry that RFCs can stall progress. We’ve seen this before (e.g. the [0-D vectors RFC](https://discourse.llvm.org/t/rfc-should-we-restrict-the-usage-of-0-d-vectors-in-the-vector-dialect) that took 6 months and required an ODM). That said, I’m not against RFCs, just cautious that we don’t end up in the same situation again. The Tensor WG seems like a good venue to first align and then escalate if needed.

**Next step:** If anyone believes the 2023 concerns still block this direction, could you restate them here? That would help us decide whether an RFC is warranted, or whether we can move forward with this canonicalization.

In the meantime, I added this to the Tensor WG agenda.

Thank you,
-Andrzej

https://github.com/llvm/llvm-project/pull/140583