[Mlir-commits] [mlir] [mlir] [linalg] Add output shape verifier for linalg ops. (PR #91673)

[llvmlistbot at llvm.org]

https://github.com/MaheshRavishankar commented:

Something similar came up recently (https://discourse.llvm.org/t/rfc-remove-arith-math-ops-on-tensors/74357/63?u=maheshravishankar)  Along those lines, it might not be a good idea to change the verifier here cause it might actually make it harder to lower to linalg from front end dialects.

Could we repurpose what you wanted to do more as a canonicalizer. If a particular dimension accessed using an expression is static but another using the same expression is dynamic, a canonicalization of linalg op could make the dynamic dimensions static (and introduce a `tensor.cast` on the operand). That might fix the issue you are having.

Sorry for the delay in the response, and apologies for the conflicting suggestions... I hadnt fully thought this through.

https://github.com/llvm/llvm-project/pull/91673