jaskij

@kraj can you direct me on how to dig into it?

Looks like it's picking up the GNU linker. ``` root@yocto:~/yocto/build/tmp-glibc/work/cortexa53-crypto-my_distro-linux/openmp/14.0.3-r0/build# aarch64-my_distro-linux-ld --version GNU ld (GNU Binutils) 2.38.20220708 Copyright (C) 2022 Free Software Foundation, Inc. This program is free software; you...

You set me on the right track. Knowing that it's the wrong linker, I tested some stuff and adding `-fuse-ld=lld` to OpenMP recipe worked. ```bitbake LDFLAGS:append = "-fuse-ld=lld" ``` Now,...

Sorry for the comment spam, but I seem to have found it - using lld distrowide is hidden behind a separate distro features, `ld-is-lld`. Which is not documented anywhere -...

@kraj `LDFLAGS:append` and documentation both against master and then cherrypick the append to Kirkstone? Or how do you want those patches?

Just tested the workaround with 1.75, and it seems to work. Granted, my build is from x86-64 to x86-64, so there may have been some contamination Yocto didn't catch.

Just saw this issue and, at least according to GCC, NEON shouldn't be used by default on ARMv7-A regardless of the support, because it's not IEEE-754 compliant. To quote the...

@galmasi I'm not one of the maintainers, just a person also using `meta-rust`. But yes, I do agree with you that disabling it by default is probably the right choice....

Found the authoritative source: > Additionally, NEON instructions always treats denormals as zero. From [ARM Cortex-A Series Programmer's Guide for ARMv7-A](https://developer.arm.com/documentation/den0013/d/Introducing-NEON/NEON-architecture-overview/Data-types), at the very bottom.

Doing some more reading, it seems that LLVM will handle at least the rounding situation correctly.