[Mlir-commits] [mlir] [AMDGPU] fold `memref.subview` into `amdgpu.gather_to_lds` (PR #149851)

[Alan Li]

lialan wrote:

> I disagree. MemRef transforms should not depend on target specific dialects like NVGPU and AMDGPU. Repeating the mistake for AMD doesn't help improve the poor state of dependency management.

I actually think you have a point that this is a design thing that folding source into an op requires the deps of the op. Isolating it to only AMDGPU avoid polluting other target pipelines. 

@krzysz00 @kuhar we need a consensus here.

https://github.com/llvm/llvm-project/pull/149851