[llvm] [RISCV][WIP] Let RA do the CSR saves. (PR #90819)

[Alex Bradbury via llvm-commits]

asb wrote:

> > @mgudim @ppenzin I was interested in kicking the tires on this to see impact on recentish HEAD on SPEC. Do either of you have a recent version of this patchset that is known good? I built https://github.com/mgudim/llvm-project/tree/save_csr_in_ra3 but I'm getting compiler crashes at the initial cmake configure for llvm-test-suite with `-mllvm -riscv-save-csrs-early=true`, or when I do it with `-save-temps=obj` as well I get errors about unrecognized instruction mnemonic `BUNDLE`.
> 
> @asb I think this branch: https://github.com/mgudim/llvm-project/tree/rebased_save_csr_in_ra should be working. To enable saving csrs early don't forget to add these two flags:
> 
> ```
> -mllvm -riscv-enable-save-csr-in-ra=true 
> -mllvm -min-weight-ratio-needed-to-evict-hint=7.5
> ```
> 
> That branch only has one huge commit and its' just POC. I am still working on breaking it up into smaller commits and putting them on (https://github.com/mgudim/llvm-project/tree/save_csr_in_ra3). But that branch is in broken state right now

Thank you. I haven't looked to see what it would take to rebase on current HEAD, but comparing that branch with a baseline of commit `d7e48fbf205a01fcbc109b2555b12aa0d37845a4` (the last upstream commit before your downstream patches), and running SPEC -march=rva22u64 I get these results for dynamically executed instructions:

```
Benchmark                       Baseline            csrinra   Diff (%)
============================================================
500.perlbench_r         180245021535    174761106584     -3.04%
502.gcc_r               220872098438    204171725268     -7.56%
505.mcf_r               131589945206    132303362396      0.54%
508.namd_r              220648060756    219744519650     -0.41%
510.parest_r            291341819989    291293000597     -0.02%
511.povray_r             31914276340     27581024333    -13.58%
519.lbm_r                94166321455     96148444880      2.10%
520.omnetpp_r           138002605450    132999926345     -3.63%
523.xalancbmk_r         283566181775    269486838918     -4.97%
525.x264_r              380165034541    383766253106      0.95%
526.blender_r           660528269650    645308870803     -2.30%
531.deepsjeng_r         355058534696    357485665398      0.68%
538.imagick_r           238573640980    238240294103     -0.14%
541.leela_r             421886351049    411685513225     -2.42%
544.nab_r               415595727842    415394887715     -0.05%
557.xz_r                132548717793    133865723492      0.99%
```

There's some big wins here, especially for povray and gcc.

Based on comments in this thread there may be issues with the wider llvm-test-suite, but all SPEC benchmarks compile and run to completion. If you're planning to do a new rebase do let me know as that would be a better test to run.

https://github.com/llvm/llvm-project/pull/90819