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

[Shreeyash Pandey via llvm-commits]

bojle wrote:

Hey @hstk30-hw.

Sorry for the delayed reply. I've measured the perf difference between patched and base llvm on xz. I notice
a 0.8% increase in dynamic instruction count with the patch. Greatest delta seems to be in the function called
lzma_lzma_optimum_normal. Here's the llvm IR for the translation unit containing this function along with the assembly
annotation for patched and base. I am looking to reduce the test further to pin-point exactly what's causing the regressions.

>From the brief look I had at it, I can see that there are extra `mov <reg>, <constant>` instructions compared to base, indicating
remats (expected from this patch). 

[lzma_optimum_normal_review_artifacts.zip](https://github.com/user-attachments/files/31304581/lzma_optimum_normal_review_artifacts.zip)


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