Hey hey, comrades!

I have a fix for some (most?) of the crashes we were seeing while
running multi-threaded code such as (guix build compile), and,
presumably, the grafting code mentioned at the beginning of this bug
report, although I haven’t checked yet.

So, ‘scm_i_vm_mark_stack’ marks the stack precisely, but contrary to
what I suspected, precise marking is not at fault.

Instead, the problem has to do with the fact that some VM instructions
change the frame pointer (vp->fp) before they have set up the dynamic
link for that new frame.

As a consequence, if a stop-the-world GC is triggered after vp->fp has
been changed and before its dynamic link has been set, the stack-walking
loop in ‘scm_i_vm_mark_stack’ could stop very early, leaving a lot of
objects unmarked.

The patch below fixes the problem for me.  \o/

I’m thinking we could perhaps add a compiler barrier before ‘vp->fp = new_fp’
statements, but in practice it’s not necessary here (x86_64, gcc 7).

Thoughts?

I’d like to push this real soon.  I’ll also do more testing on real
workloads from Guix, and then I’d like to release 2.2.4, hopefully
within a few days.

Thank you and thanks Andy for the discussions on IRC!

Ludo’, who’s going to party all night long.  :-)