[llvm] [AMDGPU][True16] added Pre-RA hint to improve copy elimination (PR #103366)

[Joe Nash via llvm-commits]

Sisyph wrote:

I don't think CostPerUse is the right solution, because we want to take use instruction context into account to decide whether a value should go into v.l or v.h. CostPerUse would tell you to always use .l over .h. We want to use .h whenever possible, because that's the way to reduce register pressure, unless the value in the register would later be forced into a VGPR32 (by being used in a 32-bit alu) 

https://github.com/llvm/llvm-project/pull/103366