[Mlir-commits] [mlir] [mlir] Handle arith.const expr in dispatchIndexOpFoldResult func (PR #122432)

[llvmlistbot at llvm.org]

rutkoor wrote:

Hi @matthias-springer @MaheshRavishankar 
Thanks for the comments. Just to give you some background, I have encountered an issue with a test case from a third-party library when lowering to an MLIR module. Specifically, the stride information is not being populated when creating the memref.reinterpret_cast instruction. After investigation, I have traced the problem to the dispatchIndexOpFoldResult function, which is invoked from the third-party library.

Given my six months of experience with MLIR, I am uncertain whether these changes should be made in the MLIR repository or within the third-party library. The builder is currently failing, resulting in 16 test cases failing. I have managed to resolve 15 of these test cases by adjusting the static shape and stride information.

I believe that this change will not only benefit the third-party library but also improve some MLIR tests by enabling vector code generation through static stride information. Additionally, I am aware of issues with the verifier, which is legalizing certain operations incorrectly.

I would appreciate feedback on whether it is advisable to proceed with this PR. Your guidance will be invaluable.

Thank you.

cc: @javedabsar1 

https://github.com/llvm/llvm-project/pull/122432