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

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

> > This is exactly my point. There is no argument to pick one over the other. So what's stopping me from writing the reverse and add that to a canonicalization?
> 
> At the moment nothing, with this patch it would be the fact that infinite loop are explicitly prohibited in our doc :)
> We should be able to decide one way or another is my point, not punting.
> (it is expected in canonicalization that there is not always an obvious tie breaker: for example we canonicalize constants to the right for commutative operations arbitrarily)


Since you ask for a basic principle, my answer is you have to prove something is a canonical representation. Just an arbitrary choice is not justification enough. You have to prove that this is "always better representation" to have it in canonicalization. 


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