[llvm] [AArch64][RegAlloc] Enable callee-saved register optimization for AArch64 (PR #214773)

[Florian Hahn via llvm-commits]

https://github.com/fhahn commented:

> At scale 100, the wins are concentrated in the call-heavy benchmarks: perlbench_r (+2.1%), gcc_r (+1.3%), and x264_r (+1.2%), while omnetpp_r and mcf_r are essentially flat. The regressions at this scale are deepsjeng_r (−0.9%) and xz_r (−1.6%); both quite small, closer to run-to-run noise. 

Did you get a chance to analyze the root causes of the regressions? It seems at least 2 improvements would also be quite close to run-to-run noise.

https://github.com/llvm/llvm-project/pull/214773