[clang] [llvm] [RegisterCoalescer] Improve register allocation for return values by limiting rematerialization (PR #163047)

[Kai Luo via llvm-commits]

bzEq wrote:

>  This happens with optimized IR too:

The pasted IR can be optimized by `opt` further, so I'm curious about the original source code.

https://llvm.godbolt.org/z/911fvaTxv

https://github.com/llvm/llvm-project/pull/163047