[llvm] [AMDGPU][TTI] Refine gfx9 packed FP32 SLP costs for pair formation and shuffles (PR #208572)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm commented:

The cost model test changes appear incidental. We should have a dedicated shuffle cost model test. 

Gating this on the specific element type doesn't sound right. You don't know just from the type the use context. This also is implicitly assuming the legal 2 element case, directly folding into the small set of pk instructions. 

Looking at the 2 element shuffle in isolation, 3 of them are a cheap, and it's only the high-to-low shuffle that's implemented with v_pk_mov_b32 



https://github.com/llvm/llvm-project/pull/208572