[llvm] [AMDGPU] Handle True16 high-half 16-bit packing idiom with uniform operand (PR #211102)

[Jay Foad via llvm-commits]

jayfoad wrote:

> It does work, but the codegen seems worse for the test cases I provide:

That's a shame. But maybe it would make sense to fix the correctness problem first, and then follow up with a separate PR to improve the generated code?

https://github.com/llvm/llvm-project/pull/211102