JIT: Make loop inversion graph based

[jakobbotsch]

Rewrite loop inversion to be graph based and to use the new loop representation.

Contributes to #107749
Contributes to #108913

Currently has large size regressions that I haven't dug into yet.

[amanasifkhalid]

Currently has large size regressions that I haven't dug into yet.

I suspect you're unblocking loop cloning in quite a few more instances -- when I prototyped this locally, I had similar size increases.

[jakobbotsch]

I suspect you're unblocking loop cloning in quite a few more instances -- when I prototyped this locally, I had similar size increases.

Doesn't appear to be loop cloning -- there are similar with loop cloning disabled in both the base/diff.

[dotnet-policy-service]

Draft Pull Request was automatically closed for 30 days of inactivity. Please let us know if you'd like to reopen it.

[amanasifkhalid]

@jakobbotsch I hope you don't mind me reviving this, but I looked at the diffs from this change on top of main to see where we stand now. With loop cloning and unrolling disabled, I see some instances where lexical loop inversion would do a transformation that fgOptimizeBranch would have done in the next phase, whereas in this PR, fgOptimizeBranch runs before graph-based loop inversion, and the latter while consider shapes that lexical inversion doesn't recognize. Thus, we can end up with extraneous cloning.

I think fgOptimizeBranch should only run right before layout where we stand to benefit most from it, since it can mess with loop recognition/opts; I can try removing the early pass as a prerequisite to this PR. However, with loop cloning/unrolling and early fgOptimizeBranch disabled, the size regressions don't come down all that much, since we're still recognizing and inverting more loops now. I'm not sure if the diffs I'm seeing are as extreme as the ones you started out with; I'll push my changes here for visibility. (I don't have permission to push to your branch; I'll open a follow-up PR). All I needed to change was add some profile maintenance code in since asserts are now enabled (a nice side effect of your implementation is loop inversion will no longer introduce profile inconsistency).