Conversion from Float64 to Float16 is incorrect

[zsoerenm]

Today I observed something unusual:

julia> Float16(4.0)
Float16(0.0)

julia> Float16(4.0f0)
Float16(4.0)

julia> versioninfo()
Julia Version 1.7.3
Commit 742b9abb4d (2022-05-06 12:58 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: AMD Ryzen 7 PRO 5850U with Radeon Graphics
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, znver3)
Environment:
  JULIA_EDITOR = code
  JULIA_NUM_THREADS = 

This might be related:
#45433
#44829

[oscardssmith]

How did you download Julia?

[oscardssmith]

wait, I'm pretty sure this is just a dup of #45433

[zsoerenm]

Well, that's embarrassing. I used https://aur.archlinux.org/packages/julia-bin from arch linux AUR.
I now downloaded Julia from the website and it does seem to work fine.

[oscardssmith]

For some reason, most linux distros don't listen to us when we say that they shouldn't link against system LLVM, which leads to a lot of people ending up with broken versions of Julia.

[zsoerenm]

I'm sorry, it was not an error from the AUR package.
I had to LD_PRELOAD=/usr/lib64/libstdc++.so.6 /bin/julia because of this issue:
JuliaGL/GLFW.jl#198
With this pre load Float16 returns wrong results even with the downloaded julia binary.

[oscardssmith]

It looks like that issue has been resolved since Julia 1.6. You shouldn't need the LD_PRELOAD anymore.

[zsoerenm]

Unfortunately this was not the case.
It looks like it is necessary each time Julias version and System version are out of sync: JuliaGL/GLFW.jl#198 (comment)

[zsoerenm]

I upgraded to Julia v1.8.0beta3 and the error is gone even with LD_PRELOAD of libstdc++.so.6.