jitlayers: Use GlobalISel on AArch64 at -O0/-O1

[IanButterworth]

#60339 was auto-closed when #59950 merged and can't be reopened.

[IanButterworth]

@xal-0 I think this fix is in-line with the reported issue so I'll merge. Post merge review would still be appreciated. Thanks

[gbaraldi]

I just wanted to ask what does clang do and that we should mostly mirror it. If rust does this then I guess it's also fine

[IanButterworth]

Claude:

---

Looking at LLVM's AArch64TargetMachine.cpp, the relevant knob is:

static cl::opt<int> EnableGlobalISelAtO(
    "aarch64-enable-global-isel-at-O", cl::Hidden,
    cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"),
    cl::init(0));

with the gate optLevel <= EnableGlobalISelAtO, so Clang enables GlobalISel on AArch64 only at -O0 by default (and falls back to SelectionDAG above that). This PR enables it at both -O0 and -O1, which is slightly more aggressive than Clang.

Given the fallback mode (GlobalISelAbortMode::Disable) and the Float16 atomic regression test added here, I think it's reasonable, but it would be conservative to narrow the gate to opt_level < 1 to exactly mirror Clang.

[IanButterworth]

It would be great if we could run an aarch64 pkgeval before and after this change, but I understand that's quite manual and time consuming

[oscardssmith]

I don't think we actually have a way to run aarch64 pkgeval.

[IanButterworth]

We have done it once at least. @maleadt did it. Not sure where the results are

[IanButterworth]

I think it's this? https://github.com/JuliaCI/NanosoldierReports/blob/9b7d04883c4e1ea8f4c4b9953fadc81e7362bcb7/pkgeval/by_hash/5e9a32e_vs_5e9a32e/report.md

Not sure how to get github to render the md

[IanButterworth]

It renders as a gist for me https://gist.github.com/IanButterworth/85ba4e9d67e44cb3260de7b316291007

[maleadt]

Yeah it's a pain. Also, the machine in question is used for Base CI now, so it's not available anymore for a full-system run (and running under qemu introduces plenty of issues with several packages).