[llvm] [SystemZ] Add a SystemZ specific pre-RA scheduling strategy. (PR #135076)

[Jonas Paulsson via llvm-commits]

JonPsson1 wrote:

I have done a rebase (Dec 4) and compared three recent versions of this strategy, and on average across benchmarks, it seems that they are fairly close:

```
(A) With "tiny regions" limit of 10                                           98.219 %
(B) Current HEAD, with HighSUs instead of "tiny regions"                      98.526 %
(C) TopCycles=6, no HasDistToTop                                              98.537 %

```
```
(A)
- Improvements:
0.858: f538.imagick_r
0.882: f507.cactuBSSN_r
0.974: f526.blender_r
0.975: f544.nab_r
0.990: i557.xz_r
- Regressions:
1.014: i500.perlbench_r
```

```
(B)
- Improvements:
0.867: f538.imagick_r
0.886: f507.cactuBSSN_r
0.975: f526.blender_r
0.982: i525.x264_r
- Regressions:
1.014: f511.povray_r
1.012: i500.perlbench_r
```

```
(C)
- Improvements:
0.869: f538.imagick_r
0.886: f507.cactuBSSN_r
0.977: f526.blender_r
0.987: i525.x264_r
0.988: f519.lbm_r
- Regressions:
(none)
```

Conclusion: With this particular version (there were no regressions with A or B last time), the "tiny regions" version is still hard to beat on average, but the "no HasDistToTop" is doing quite well still with no regressions, and maybe this would actually be ideal (with a simplification of tryCandidate() as discussed).


https://github.com/llvm/llvm-project/pull/135076