[llvm] [AMDGPU] misched: avoid subregister dependencies (PR #140255)

[Robert Imschweiler via llvm-commits]

ro-i wrote:

But aren't the true 16 instructions dedicated instructions? Here, the instructions that deal with 16-bit values, for example, *do* take 32-bit operands and then select the relevant parts (which may be one or both parts) according to op_sel/op_sel_hi. So, the instruction using only the lower half of their Nth operand, the instruction using only the upper half, and the instruction using both halves are essentially the same instruction, just with different op_sel/op_sel_hi values

https://github.com/llvm/llvm-project/pull/140255