[llvm] [AMDGPU][GISel] Do not scalarize uniform v2f16/v2bf16 operations (PR #222359)

[Matt Arsenault via llvm-commits]

================

----------------
arsenm wrote:

Can you pre-fix these tests using void returns? They are storing to memory and using VGPR results contrary to the name. This is using a v_and_b32, which in this context makes sense since it has a single VGPR use. I don't think we actually implemented that optimization though? 

https://github.com/llvm/llvm-project/pull/222359