[Mlir-commits] [mlir] [mlir][gpu] Pattern to promote `gpu.shuffle` to specialized AMDGPU ops (PR #137109)

[Ivan Butygin]

Hardcode84 wrote:

I was thinking about doing it as part of the `gpu.shuffle` lowering. My current motivations to do it separately:
* Easier to test when it's a small separate pattern
* I'm promoting high-level `gpu.shuffle` op to high level `amdgpu.swizzle_bitmode` op. If we decide to do it during lowering, we will need to generate `rocdl` op directly, copypasting part of the exiting shuffle/swizzle lowering. There may also be possible to run further simplifications/promotions on the resulting `amdgpu.swizzle_bitmode` (but I don't have any yet).
* It's opt-in so we won't suddenly break all AMDGPU users if something go wrong

https://github.com/llvm/llvm-project/pull/137109