Fix assertion failure from #12712

[jmid]

This little PR fixes the systhreads assertion failure from #12712.

By good old printf-debugging:tm: I found out that the failure happens
on a boundary condition when 4/5 minor heap pointers align:

 young_start: 4029677568
 young_limit: 4029677568
 young_trigger: 4029677568
 young_ptr: 4029677568
 young_end: 4029685760

Multiple reruns confirmed that the condition is the same (the
addresses naturally vary) every time the assertion fails.

Here's my current understanding:

• The minor heap uses a young_ptr bump allocator starting
  from young_end and working towards young_start.
• young_trigger typically marks the mid-point between the two,
  and hitting it triggers a major gc slice.
• Other domains may set young_limit to UINTNAT_MAX to notify
  a domain (irrelevant in the present systhread test with only 1 domain)

With the above in mind, AFAICS, the above condition may happen when

• a previous young_ptr allocation just fit perfectly
• young_trigger is set to young_start by caml_poll_gc_work

This also explains why running with a reduced minor heap
OCAMLRUNPARAM="s=1024" helps to reproduce the failure.

The PR simply proposes to adjust the assertion to account for the
boundary condition.

With the fix, the systhread test from #12712 runs fine across 20 iterations.
To confirm it further, I've also locally run

• the ocaml compiler test suite
• the multicoretests test suite both with and without the debug runtime

[gasche]

(cc @gadmm if you have the time)

[gadmm]

Yes I think this is the right thing to do. The intent might have been that if young_ptr == young_limit then all (non-zero) allocations will fail anyway (but not pure polling which is a zero-size allocation). But the edge case where young_limit == young_ptr becomes important for memprof where young_limit gets other intermediate values (correctly, there is nothing to poll in this case, but all non-zero allocations must fail). So I think changing the assertion is the correct move.

[gasche]

Approved on @gadmm's behalf. Thanks both!

[gasche]

(I assumed the lack of a Changes entry was intentional, but please do feel free to write one.)

[jmid]

Thanks for the prompt review! I've added a Changes entry.

[kayceesrk]

The change looks good to me. The reasoning

With the above in mind, AFAICS, the above condition may happen when

• a previous young_ptr allocation just fit perfectly
• young_trigger is set to young_start by caml_poll_gc_work

also makes sense to me.

[dra27]

@Octachron - @jmid is continuously testing the released compiler, so there is use in this being back-ported to 5.1.1. May I do that?

[Octachron]

Since this is a small assertion fix, ok for backporting to 5.1.1 .