[Mlir-commits] [mlir] [mlir][linalg] Extend `FuseElementwiseOps` pattern to work with named ops (PR #144922)

[llvmlistbot at llvm.org]

srcarroll wrote:

> I dont have a major issue with the change itself, but I think this is different from the direction I think we had agreed on

I wasn't involved in the discussion so I dont know exactly what was agreed upon but, as far as I know, this doesn't preclude any op design changes in linalg.

>  I think we need to remove regions from the elementwise/contract/fill ops.

Not sure why we would do that. If we want these ops to have generic counterparts, then I dont think it makes sense to remove regions. But again, since I wasn't involved in the discussion I'll concede to those of you who know more about this. Nevertheless, I dont think these ops should be excluded from fusion if that's what you're getting at.

> At the very least I think a lot of complexity introduced here are due to map operations. If we fix that then Id expect almost no changes to the fusion logic here. Id rather do that than import that complexity here.

Agree. discussed [here](https://github.com/llvm/llvm-project/pull/144922/files/f7e164beebdc8195b6781c2a8ce8bc1bea7757cd#r2158519239)

https://github.com/llvm/llvm-project/pull/144922