[llvm] [AMDGPU][GlobalISel] Fold G_TRUNC(G_LSHR(x, 16)) into hi16 subregister copy in True16 mode (PR #196028)

[Harrison Hao via llvm-commits]

harrisonGPU wrote:

> This should be done in tablegen (and I'd kind of expect we would have such a pattern already that works in sdag)

Yes, sdag already supported this pattern, when I  implemented this PR: https://github.com/llvm/llvm-project/pull/192904, I noticed this issue.

https://github.com/llvm/llvm-project/pull/196028