Pass -mno-avx512f flag to disable AVX512 opcodes in debs

[cottsay]

It seems that when we build packages on Xeon systems which support the AVX512 extension, the corresponding opcode is used in the resulting binaries. This leads to SIGILL on platforms which don't support the extension, so we should explicitly disable it to maintain wide compatibility.

I verified that the flag was passed through to gcc by invoking the buildfarm locally using generate_release_script.

[j-rivero]

If this is going to affect arm builds, we might have a problem:

arm-unknown-linux-gnueabihf-g++: error: unrecognized command-line option '-mno-avx512f'

https://godbolt.org/z/fEb9P1jKh

[cottsay]

If this is going to affect arm builds, we might have a problem

Unlike ros-infrastructure/ros_buildfarm#1012, this approach is arch-specific. These build files are amd64-only (the arch is specified near the end of the file).

[j-rivero]

Good to me. GCC/Clang on Focal/Jammy seems to support the option just fine.

[claraberendsen]

LGTM, as long as @clalancette approves. From my understanding from the previous discussion this is the best approach for an easy revert if necessary no?

[clalancette]

LGTM, as long as @clalancette approves. From my understanding from the previous discussion this is the best approach for an easy revert if necessary no?

I will point out that I thought our decision was to do nothing here. But given that this is easy to revert if it causes problems, we can go ahead with it.

[cottsay]

I will point out that I thought our decision was to do nothing here.

You're right - I got excited trying to find the lowest-cost solution. Since I'll be away for a while, I'll leave this PR open but convert it to a draft. If you guys feel the need to move forward with it, it should be ready to merge and deploy.

Thanks everyone for the reviews.