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

[Mehdi Amini]

joker-eph wrote:

> Since you ask for a basic principle

You're providing a vague statement: this isn't a set of principles upon which we can build anything right now. 
Also: when I ask you to provide principles, just not doing it and trying to escape with returning it as "you should prove that..." isn't productive.

For example my principles for canonicalization:
- We shouldn't lose information
- The canonicalization can be "undone" without heroics (ideally analyses should be able to pattern-match easily in the target form)

> my answer is you have to prove something is a canonical representation. 

Sorry but I can't follow you: are you saying that the normal standard for canonicalization in MLIR is that we have to **prove** anything? I don't believe we ever hold any canonicalization to this standard (please prove to me that canonicalizing constant to the right is canonical...).




You probably should re-read https://sunfishcode.github.io/blog/2018/10/22/Canonicalization.html which is the underlying reference in our documentation: https://mlir.llvm.org/docs/Canonicalization/



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