[Mlir-commits] [mlir] [mlir][tensor] Add a PadOp::FoldReifiedShape canonicalization (PR #145732)

[Mehdi Amini]

joker-eph 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

I don't understand what you mean actually, I feel you're pointing at something and driving a vastly overreaching conclusion, but I'm not sure I follow your reasoning here (nor the actually conclusion: what does it mean to be "moot"?).

> upstream example, we push constants and some operations back into gpu kernels to undo CSE

Undoing canonicalization does not invalidate the canonicalization (neither the particular one, nor the concept of canonicalization) : this seems to be mistaking "canonicalization" with "optimization" here IMO.

> On the technical concept, I think we have a strong alignment: specific properties of the IR are desired over a generic "canonicalization". This seems to be a dominant viewpoint by the way, just nobody took the time to share it.

I don't know what this means: in the meantime count me strongly against any "alignment" around this please.

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