[llvm] [AMDGPU] Convert ds_bpermute/wave_shuffle XOR patterns to DPP row_xmask and permlanex16 (PR #194565)

[Pankaj Dwivedi via llvm-commits]

PankajDwivedi-25 wrote:

> > > Do we really want to be implementing these optimizations twice, once for bpermute and once for wave_shuffle? Maybe we should canonicalize bpermute to the corresponding wave_shuffle?
> 
> > 
> 
> > Yes, that would be much better. In that case, other PR:#192246 will be all that is required?
> 
> 
> 
> As Jay has mentioned in PR:#192246, there's currently a few PRs in flight that all do similar things around adding optimizations for wave_shuffle and ds_bpermute (including one I had opened that has some overlap with yours as well). I think we only need a single PR to get all of these optimizations on. That one you linked seems to currently cover the most cases and includes the consolidation of both intrinsics to one optimization path. I'd be in favor of focusing our efforts on it if you and Jay are as well.

I agree, That's much better.

It's better to continue wave_shuffle which is umbrella interface to all these.

https://github.com/llvm/llvm-project/pull/194565