erf performance regression on Windows

[johnnychen94]

We recently upgraded our Julia from 1.10.8 to 1.10.9 and noticed the performance difference of erf.

using BenchmarkTools
using SpecialFunctions

@btime erf.(-5.0:1e-6:5.0);

The benchmark is running on my laptop windows and WSL(ubuntu) by switching different OpenLibm_jll versions.
The SpecialFunctions I use is the master branch version (738435a)

Julia Version 1.10.9
Commit 5595d20a28 (2025-03-10 12:51 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 20 × 12th Gen Intel(R) Core(TM) i7-12700H
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, alderlake)
Threads: 1 default, 0 interactive, 1 GC (on 20 virtual cores)

julia version	OpenLibm_jll version	OS	time (ms)
1.10.9	0.8.5+0	Windows	375.325
1.10.9	0.8.5+0	Linux	195.182
1.10.9	0.8.4+1	Windows	362.406
1.10.9	0.8.4+1	Linux	198.031
1.10.9	0.8.4+0	Windows	368.287
1.10.9	0.8.4+0	Linux	193.683
1.10.9	0.8.1+4	Windows	363.095
1.10.9	0.8.1+4	Linux	197.910
1.10.9	0.8.1+3	Windows	187.219
1.10.9	0.8.1+3	Linux	190.244
1.10.9	0.8.1+2	Windows	189.346
1.10.9	0.8.1+2	Linux	204.383

The erf's performance on windows has dropped 2x since OpenLibm_jll v0.8.1+4.

P.S. Julia 1.10.8 ships OpenLibm_jll v0.8.1+2, and Julia 1.10.9 ships OpenLibm_jll v0.8.1+4

[inkydragon]

The stack protection was enabled in OpenLibm_jll v0.8.1+4, and the performance slowdown was somewhat of an expected impact.

ref:

• v0.8.1+4 diff: JuliaBinaryWrappers/OpenLibm_jll.jl@c4c6215
• Yggdrasil c48529abde: https://github.com/JuliaPackaging/Yggdrasil/blob/c48529abde37f64e6cecd2d3f52c360267ccd085/O/OpenLibm/OpenLibm%400.8.1/build_tarballs.jl
• Yggdrasil PR: https://github.com/JuliaPackaging/Yggdrasil/pull/10447/files

[giordano]

The stack protection was enabled in OpenLibm_jll v0.8.1+4, and the performance slowdown was somewhat of an expected impact.

How's that change relevant for Windows? The stack protection should be applied only to ELF files, no?