runtime_intrinsics: Fix f16 ABI fallback

[Keno]

The x86 psABI specifies that f16 goes on the stack (but is returned in xmm0). Our emulation for compilers that do not support _Float16 did not handle this properly. Fixes #61072.

I think this fixes it at least - I'm too tired to actually go check at this point, so I'll just let CI try it ;). My remaining concern is that I don't quite understand why seeing this behavior is AVX512 dependent.

[DilumAluthge]

Okay, CI passed on Windows, but neither of the Windows jobs ran on rhea, so I don't know if we would have caught the bug.

[gbaraldi]

I suspect it's dependent on it because it changes if the cpu has native support or not

[DilumAluthge]

I suppose someone with access to rhea will need to run these tests manually on the machine, to confirm the fix? @gbaraldi Any chance you have access?

[giordano]

$ julia +nightly~x86 --cpu-target=pentium4 -E 'Float16(3.0) * 2'
Float16(4.0)
$ julia +nightly~x86 -E 'Float16(3.0) * 2'
Float16(6.0)
$ ./julia-46c8113b1a/bin/julia --cpu-target=pentium4 -E 'Float16(3.0) * 2'
Float16(6.0)
$ ./julia-46c8113b1a/bin/julia -E 'Float16(3.0) * 2'
Float16(6.0)
$ julia +nightly~x86 -e 'using InteractiveUtils; versioninfo()'
Julia Version 1.14.0-DEV.1761
Commit 9aff288d6de (2026-02-20 13:53 UTC)
Build Info:
  Official https://julialang.org release
Platform Info:
  OS: Linux (i686-linux-gnu)
  CPU: 224 × Intel(R) Xeon(R) CPU Max 9480
  WORD_SIZE: 32
  LLVM: libLLVM-20.1.8 (ORCJIT, sapphirerapids)
  GC: Built with stock GC
Threads: 1 default, 1 interactive, 1 GC (on 224 virtual cores)
$ ./julia-46c8113b1a/bin/julia -e 'using InteractiveUtils; versioninfo()'
Julia Version 1.14.0-DEV.1760
Commit 46c8113b1a4 (2026-02-20 07:23 UTC)
Build Info:
  Official https://julialang.org release
Platform Info:
  OS: Linux (i686-linux-gnu)
  CPU: 224 × Intel(R) Xeon(R) CPU Max 9480
  WORD_SIZE: 32
  LLVM: libLLVM-20.1.8 (ORCJIT, sapphirerapids)
  GC: Built with stock GC
Threads: 1 default, 1 interactive, 1 GC (on 224 virtual cores)

Seems to work for me (not on rhea, no clue of what machine is that)

[Keno]

I suspect it's dependent on it because it changes if the cpu has native support or not

I don't understand this. On demeter (no AVX512FP16):

keno@demeter6:~$ julia +nightly~x86 --cpu-target=pentium4 -E 'Float16(3.0) * 2'
Float16(4.0)

So the bug at least is not dependent on having support.

[giordano]

What's the microarchitecture of demeter? I couldn't reproduce on a bunch of machines I tried (including znver4, which has avx512, but not avx512-fp16), only on sapphirerapids.

[DilumAluthge]

This needs backporting to all active release branches, I assume?

[Keno]

What's the microarchitecture of demeter?

znver4

[giordano]

Ah, checking the history of the znver4 machine I tested this on, I had tried release~x86, not nightly~x86. I can confirm I reproduce the bug with nightly~x86 (and 1.13-nightly~x86) on znver4, but not release~x86.