[llvm] PHIElimination: add target hook to control reuse. (PR #163604)

[Junjie Gu via llvm-commits]

jgu222 wrote:

> Is this the problem solved by SIOptimizeVGPRLiveRange?
> 
> You generally want to work in terms of common subclasses than exact class checks

Based on my understanding of AMD's SIOptimizeVGPRLiveRange, it adds phi to make live ranges precise for AMD GPU code.  Intel's control-flow is different, it uses goto/join, just like a normal CPU goto. but with execution mask taking into consideration.  No, this  issue isn't solved by SIOptimizeVGRPLiveRange.

https://github.com/llvm/llvm-project/pull/163604