[llvm] [AMDGPU] [SIFrameLowering] Use LiveRegUnits instead of LivePhysRegs (PR #65962)

[Pranav Taneja via llvm-commits]

pranavTanejaAMD wrote:

> > > Whole stack LGTM. Thanks for working on this! Have you measured any compile-time difference from this change?
> > 
> > 
> > [compileTimeDiff.log](https://github.com/llvm/llvm-project/files/12595643/compileTimeDiff.log)
> > Here's the difference between compile time from both the original and modified builds on compiling tests from llvm-test-suite
> 
> So didn't really do anything for x86, which is expected. Ideally we would have GPU compute compile time tracking but we don't

If it serves as a proxy for AMDGPU compilation time, here's the difference in time it takes to complete testing the lit tests in "llvm/test/CodeGen/AMDGPU", with the modified branch with LiveRegUnits taking about a second less in multiple trials. 
[litTestingTime.patch](https://github.com/llvm/llvm-project/files/12616379/litTestingTime.patch)


https://github.com/llvm/llvm-project/pull/65962