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

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

> > I think we have the same discussion many times. Canonicalizations should be narrow.
> 
> Let’s not have the discussion repeatedly and keep it grounded in first principles then please? “Narrow” does not mean anything to me!
> 
> Back to first principles: canonicalization is about finding a single form for two equivalent programs.

There might not be a single form. The two equivalent programs are equivalent and there is nothing to chose between them without additional context that only exist in the use case.

> 
> If you have an expand shape in the input IR: would you do the opposite transformation?

No. IMO that is an opinionated decision. Should be left to the downstream users to pick what works for their context.



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