Fix using roundeven intrinsic on darwin

[straight-shoota]

Resolves #10475

Based on #10476 to make sure it actually works (and continues to work).

[asterite]

Thank you!

Do you know why that function isn't there on Mac? And what's the benefit of using that over the alternative?

[straight-shoota]

I have no idea why it's missing on macos. But it should be available in LLVM 11 (https://releases.llvm.org/11.0.0/docs/LangRef.html#llvm-roundeven-intrinsic) and it does on linux.
Maybe we shouldn't use it altogether, or just for linux explicitly?

As far as I understand, roundeven was introduced for slightly improved semantics (it's about specific error conditions IIRC). rint might be removed at some point, but might not. It definitely feels better to use the more recent roundeven when it's available.

[asterite]

Well, the docs say it's not supported on all target. Maybe this is one case.

[straight-shoota]

Could be. But when a major platform doesn't support even the main float types, I'd expect a more clear message that it's only supported on some targets at all.
Btw. the same is said about rint so you could assume it should be unavailable on similar targets.

[asterite]

Is there a performance difference or something else? I think that if they are equivalent but one works in more platforms, we should use that alternative.

[straight-shoota]

Well roundeven is supposed to habe the better, standard-conformant semantics. Not sure if that really matters for our use case, but on platforms that support it, I don't think there's any reason to use rint over roundeven.

Since we don't know about other platforms, I'd limit usage to LLVM >= 11 on linux. That should be good.

[bcardiff]

Since llvm@11 support still has it's gotchas (ref: #10359). I think this is fine as is. The official releases does not use llvm@11 and we can even fallback to llvm@9 in homebrew. Meanwhile this can be merged.

Worst-case, if roundeven_f32 is missing on some platform, the method can easily be monkey-patched and a PR submitted.

[straight-shoota]

I actually think we should use rint for now on every platform. That seems to be more widely available. As Ary said there's no real downside to using rint over roundeven. So we should just go with that and not block ourselves for improving LLVM 11 support.
We can introduce roundeven later when it has more support.

[asterite]

@straight-shoota Thank you! And just to be clear, it might have looked like I'm totally against using roundeven but that's not the case. I'm sure eventually we'll want to use it, it just seems that it's poorly supported right now.

[SlayerShadow]

@straight-shoota, thank you! It fixes compiler at least on Alpine and with latest LLVM versions. Compiler compiled with --release flag still doesn't work but it seems unrelated issue.