[llvm] [AMDGPU][True16] Improve 16-bit argument handling (PR #217811)

[Shilei Tian via llvm-commits]

shiltian wrote:

Hmm, after giving it some more thought, I'm wondering if we could do it right in the first place: we need to copy to/from the lower 16 bits of the corresponding physical register. We can still leave the calling convention unchanged, which means that even two adjacent 16-bit arguments would still use two VGPRs.

https://github.com/llvm/llvm-project/pull/217811