Chairmarks segfaults on 1.12 #57018
Comments
LilithHafner
added
GC
|
With multiple threads and an assert build on 3d85309 (the assertions only seem to trigger when Julia is started with multiple threads): julia> using Chairmarks
julia> x = @be rand;
julia> typeof(x)
[115080] signal 11 (1): Segmentation fault
in expression starting at REPL[3]:1
julia: /cache/build/builder-demeter6-3/julialang/julia-master/src/gc-stock.h:437: gc_check_ptls_of_parallel_collector_thread: Assertion `__extension__ ({ __auto_type __atomic_load_ptr = (&ptls->gc_state); __typeof__ (*__atomic_load_ptr) __atomic_load_tmp; __atomic_load (__atomic_load_ptr, &__atomic_load_tmp, (memory_order_relaxed)); __atomic_load_tmp; }) == 3' failed.
julia: /cache/build/builder-demeter6-3/julialang/julia-master/src/gc-stock.h:437: gc_check_ptls_of_parallel_collector_thread: Assertion `__extension__ ({ __auto_type __atomic_load_ptr = (&ptls->gc_state); __typeof__ (*__atomic_load_ptr) __atomic_load_tmp; __atomic_load (__atomic_load_ptr, &__atomic_load_tmp, (memory_order_relaxed)); __atomic_load_tmp; }) == 3' failed.
julia: /cache/build/builder-demeter6-3/julialang/julia-master/src/gc-stock.h:437: gc_check_ptls_of_parallel_collector_thread: Assertion `__extension__ ({ __auto_type __atomic_load_ptr = (&ptls->gc_state); __typeof__ (*__atomic_load_ptr) __atomic_load_tmp; __atomic_load (__atomic_load_ptr, &__atomic_load_tmp, (memory_order_relaxed)); __atomic_load_tmp; }) == 3' failed.
julia: /cache/build/builder-demeter6-3/julialang/julia-master/src/gc-stock.h:437: gc_check_ptls_of_parallel_collector_thread: Assertion `__extension__ ({ __auto_type __atomic_load_ptr = (&ptls->gc_state); __typeof__ (*__atomic_load_ptr) __atomic_load_tmp; __atomic_load (__atomic_load_ptr, &__atomic_load_tmp, (memory_order_relaxed)); __atomic_load_tmp; }) == 3' failed.
[115080] signal 6 (-6): Aborted
in expression starting at REPL[3]:1
julia: /cache/build/builder-demeter6-3/julialang/julia-master/src/gc-stock.h:437: gc_check_ptls_of_parallel_collector_thread: Assertion `__extension__ ({ __auto_type __atomic_load_ptr = (&ptls->gc_state); __typeof__ (*__atomic_load_ptr) __atomic_load_tmp; __atomic_load (__atomic_load_ptr, &__atomic_load_tmp, (memory_order_relaxed)); __atomic_load_tmp; }) == 3' failed.
Aborted (core dumped) |
* update compat (need Julia 1.8 for allequal) * Stop testing on nightly due to JuliaLang/julia#57018
|
I don't need threads. It looks like the return value, which should be: is instead getting a pointer to the raw memory allocation instead of the Vector wrapper of the Memory wrapper of it (happens to be a malloc object), which segfaults in the GC later |
|
There appears to be an ABI disagreement. The caller (which is |
|
Could this be an artifact of mine and @oscardssmith PR? |
|
For those playing along at home, the ABI issue comes from the abstract type precompile https://github.com/LilithHafner/Chairmarks.jl/blob/80e84d0ba90fcac61959d1f992e034f5445bf4cd/src/precompile.jl#L7, which made Julia think that the function returned a boxed Tuple, instead of an unboxed Tuple |
On 1.11, Chairmarks works fine. On 1.12, it segfaults. The strings
"@inbounds"and"unsafe"do not appear in thesrcdirectory and I don't do anything particularly internal that I know of. IMO, the funkeyest code is here: https://github.com/LilithHafner/Chairmarks.jl/blob/80e84d0ba90fcac61959d1f992e034f5445bf4cd/src/benchmarking.jl#L115-L165.Tagging GC at a guess because it appears in the segfault stack trace, but it could certianly be due to a non-gc issue.
MWE:
The text was updated successfully, but these errors were encountered: