[llvm] [AMDGPU] Add IR LiveReg type-based optimization (PR #66838)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

Decoupling from "[AMDGPU]: Accept constant zero bytes in v_perm OrCombine" as that is taking longer than expected, and this has priority.

As a result, in the exotic cases (e.g. v3i8), we may produce suboptimal codegen, but, for the normal case, codegen is much improved.

https://github.com/llvm/llvm-project/pull/66838