[llvm-branch-commits] [llvm] [AMDGPU] Add true16 patterns for build_vector (vgpr, 0) (PR #192147)

[Stanislav Mekhanoshin via llvm-branch-commits]

rampitec wrote:

FYI: I have tried to use pattern with `and 0xffff` like we have for fake16 and older targets, and I've got mixed results, so I didn't like resulting code. In some cases it allows that `and` to become a dual instruction, but in much more cases it has produced just longer VOP3. With that I can technically switch to `v_cvt_u32_u16` for gfx11+ on fake16 too, but it makes a little sense as all of that now going into true16.

https://github.com/llvm/llvm-project/pull/192147