task: Add missing write-barrier in save_stack

[topolarity]

Caught by JuliaCI/julia-buildkite#541.

Reproduces readily with JULIA_COPY_STACKS=1 in an MMTK_PLAN=StickyImmix build:

$ JULIA_COPY_STACKS=1 ./julia --startup-file=no -e "fib(x) = @sync begin; function f(x); x in (0, 1) && return x; a = Threads.@spawn f(x - 2); b = Threads.@spawn f(x - 1); fetch(a) + fetch(b); end; f(x); end; println(fib(25))"

[2035119] signal 11 (1): Segmentation fault

This doesn't affect the stock GC which has a persistent remset, but it does matter for MMTK.

Investigated from core dump by Claude Fable 🤖 (w/ some human prodding)

[qinsoon]

LGTM. Thanks for the fix.

[topolarity]

I'm a bit surprised StickyImmix was ever working, given this bug (threads.jl hits it very reliably)

Is it just that we never tested it on the full test suite?

[qinsoon]

I'm a bit surprised StickyImmix was ever working, given this bug (threads.jl hits it very reliably)

Is it just that we never tested it on the full test suite?

We used to run the test suite in the binding CI (before moving the binding code into the Julia repo). We run the tests with our scripts, but mostly we run make test-xx. The last run was from last month, and threads.jl tests passed.

Sticky immix was also tested in the original PR to buildkite (JuliaCI/julia-buildkite#517) in Feb. But after that, we haven't ever run the entire test suite for sticky immix on Julia's CI.

[topolarity]

I'm glad I asked - Thanks for helping me reconstruct the history.

That means this regressed recently then. Looks like https://github.com/JuliaLang/julia/pull/62224/changes#diff-1ac89047bd5ddf49f271321da0d86793efdc534d0edd56548b3ea0dcd2f37c04R464
which moved the barrier before the store but also accidentally across a safepoint.