[llvm] [AMDGPU] Improve register allocation to reduce MFMA hazard NOPs (PR #156943)

[Sirish Pande via llvm-commits]

https://github.com/srpande commented:

Since this change is actually introduding anti-hints in general in RA, I would rephrase the original summary to include anti-hint. 
The key part of this work is introduction of anti-hint and using that in allocation order to improve RA. 
Can you please reword that?

https://github.com/llvm/llvm-project/pull/156943