[llvm] [AMDGPU] Improve codegen for uniform f16<-->i32 conversions (PR #176833)

[Syadus Sefat via llvm-commits]

mssefat wrote:


> Why are f16<-->i32 conversions legal in the first place? I don't think any subtarget has an instruction for this, does it? 
   -- I think no subtarget has it.   
> For the specific case of f16 --> u32, I think you can implement it as f16 --> u16 --> u32 which might be slightly faster.
   -- But we do not have the  "s_cvt_u16_f16" instruction.


https://github.com/llvm/llvm-project/pull/176833