Regression in 1.12: random crashes with @atomic or @atomiconce

[araujoms]

This is the worst Heisenbug I've ever seen. It all started with the tests of PermutationGroups failing in 1.12: kalmarek/PermutationGroups.jl#44 Then after it was hopefully fixed kalmarek/PermutationGroups.jl#46 it turned out that it just got harder to trigger. CI is happy on 1.12, but not on 1.13, and if I try to use the package I get random crashes. @vtjnash believes it's a bug in Julia itself, not the package, and suggested me to open a bug report here. He thinks it is related to #59559.

I tried to make a minimum working example. I failed, but at least I managed to write a test that deterministically fails on both 1.12 and 1.13. It is in this branch here: https://github.com/araujoms/PermutationGroups.jl/tree/bug . To make things more mysterious, I only get a crash if I do ]activate . and test. If I simply include runtests.jl there's no crash.

In 1.11 everything works.

[sgaure]

I can reproduce this sometimes on 1.13, only very infrequently on 1.12. Also from the command line -e 'import Pkg; Pkg.activate("."); Pkg.test()'. In #59861 it's demonstrated that the julia-compiler is non-deterministic in 1.12 and 1.13. It might be related to that. Or not.

Interestingly, the test runs fine if I redirect stderr to a file when starting julia (but not if I redirect only stdout, or stderr to stdout). Memory corruption is weird.

[araujoms]

For me it's deterministic. Even with a single thread and clean environment if fails without fail. Maybe it's architecture dependent as well? My versioninfo():

julia> versioninfo()
Julia Version 1.12.1
Commit ba1e628ee49 (2025-10-17 13:02 UTC)
Build Info:
  Official https://julialang.org release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 12 × 13th Gen Intel(R) Core(TM) i7-1355U
  WORD_SIZE: 64
  LLVM: libLLVM-18.1.7 (ORCJIT, alderlake)
  GC: Built with stock GC
Threads: 6 default, 1 interactive, 6 GC (on 12 virtual cores)
Environment:
  JULIA_NUM_THREADS = 6