[Mlir-commits] [mlir] [mlir][tensor] Make tensor::PadOp a ReifyRankedShapedTypeOpInterface and add a PadOp::FoldReifiedShape canonicalization (PR #145732)

[Oleksandr Alex Zinenko]

ftynse wrote:

I generally think the canonical form in MLIR is moot, given that we can add arbitrary operations and run transformations in vastly different paradigms (trivial example, LICM/CSE/GVN was harming affine transformations so much they had to undo it in Polly: https://dl.acm.org/doi/abs/10.1145/3168815, but these are generally considered "useful simplifications" for regular SSA; upstream example, we push constants and some operations back into gpu kernels to undo CSE) that may have different pre-requirements. We should rather move to a set of composable "normal forms" that have well-defined, programmatically verifiable invariants, and passes can declare which of the "normal forms" they expect and produce.

> Anyway, not trying to start an RFC in a PR or an ODM but a discussion on "is making shapes more static" a canonicalization and "creating IR that can be reverted" would be useful principles to have a general agreement on.

I think this is precisely a discussion that should happen at an ODM. This is about design, and we need high-bandwidth discussion. ODM is not a seminar were people present finished work, or rather, it should be that. It is also worth surfacing this discussion to the forum, there was a related topic recently. I feel like I've written the text from above at least twice in the past month or so.

https://github.com/llvm/llvm-project/pull/145732