Don't try to use AVX512 without OS support

[silby]

AVX512 instructions require OS support as well as CPU support. Bits 5, 6, and 7 of XCR0 should be enabled to run with AVX512 support. This fixes SIGILL on OpenBSD, for example.

OpenBSD ports commit with identical patch

Required alongside the similar fix to simdutf pulled in by #564 for text to work right on OpenBSD when the processor flags AVX512 support.

[Bodigrim]

@silby is it possible to setup a regression CI job? Which OS do not support AVX512?

[silby]

Well the only one I know about is OpenBSD, which is where we discovered the problem. (Whole saga here.) I'm not an expert on really anything involved here I just happened to try to use pandoc on OpenBSD-current on my 11th-gen Intel laptop and banged my head against it until I found enough clues. So I think realistic regression-testing would require specific hardware for the test runner, since it would have to have a processor indicating AVX512 support.

Edit: It looks like to run your Cirrus-CI OpenBSD build with AVX512 you would have to run it on self-managed machines so that you could pick a compute instance type with the required architecture, e.g. Google Compute Engine N2 guarantees at least Cascade Lake. You can't get that granular without provisioning it yourself.

[Bodigrim]

@blackgnezdo could you please advise?

[blackgnezdo]

Since @silby figured this out and tested the fix, I'm pretty comfortable with suggesting this gets merged.

I don't have any insights about where we can host a CI with such special CPU configuration.

[Bodigrim]

Could anyone check that the patch does not disable AVX-512 on platforms capable of it? I don't have access to modern Intel CPU at the moment.

[Lysxia]

I don't have access to a AVX-512 capable CPU either.

[Bodigrim]

I tried this on macOS x86_64 with AVX-512 instructions and, unfortunately, the patch just disabled their usage. I think this is a peculiarity of macOS (see https://stackoverflow.com/questions/72522885/are-the-xgetbv-and-cpuid-checks-sufficient-to-guarantee-avx2-support and golang/go#43089), but I'm not keen to break macOS for the sake of OpenBSD.

I guess we can guard xgetbv check by #ifndef __APPLE__, but we still need an evidence that xgetbv does not disable AVX-512 on Windows / Linux.

[Bodigrim]

Converting to draft until someone fixes the behaviour on MacOS and tests it on Windows / Linux machines which have AVX-512.

[falsifian]

Another problem is that this causes a SIGILL on CPUs that don't have the xgetbv instruction. See here. Checking for the osxsave CPU bit as is done here and avoiding the xgetbv if not set would probably fix it.

(Incidentally, AVX-512 support was just added to OpenBSD-current, so hopefully the original problem no longer affects OpenBSD.)