[Mlir-commits] [mlir] Remove Pure attribute from Linalg::IndexOp. (PR #68894)

[llvmlistbot at llvm.org]

erick-xanadu wrote:

Hi, I am pinging this since CSE was brought up again internally. It looks like there are several different things discussed here:

1. Nesting `linalg` operations is not the preferred way of representing programs. And as a result neither does having a function call inside a `linalg` operation. It is not ideal to disallow this via the verifier because

> it could be a way of expressing programs, but it's not the "preferred way" afaik

So, no change there?

2. Having `linalg.index` being pure is not sound. But having CSE work on `linalg.index` is something that is wanted. It looks like the best solution is to model `linalg.index` as reading from a loop index resource that is written to by `linalg.generic`?

Happy to work on this, but I just would like a little bit more direction before starting to avoid working and having my work not being used at all. Thanks!

https://github.com/llvm/llvm-project/pull/68894