Ephemerons: wait on all domains to be done marking before asking for another round

[gasche]

Ephemeron marking must be done after normal marking, and may in turn reveal more normal marking work, which in turn requires a new round of ephemeron marking (marking all "TODO" ephemerons again to check if their keys have been marked), etc., until a fixpoint is reached as no additional normal-marking happens anymore.

On a single-domain system there is an obvious good strategy: wait until all normal-marking work is done, and then do ephemeron marking, and repeat. On multi-domain systems we only know when we (the current domain) are done marking. The trunk code asks for a new ephemeron round (on all domains) whenever a domain is done marking; this can lead to useless repetition of ephemeron-marking work: if two domains finish marking one after the other, each will ask all domains to do a round of ephemeron-marking.

This PR changes the end-of-normal-marking logic to ensure that we ask for a new round of epehmeron marking when all domains are done marking, rather than whenever the current domain is done marking.

@OlivierNicole and @damiendoligez reviewed an earlier version of this branch yesterday and we found a concurrency bug together. The new version uses a CAS loop to avoid concurrency issues, as we discussed.

[eutro]

The CAS loop looks sound to me, and I see why it is necessary instead of just checking caml_atomic_counter_decr(&num_domains_to_mark) == 0 (although we miss a debug assert that domains_still_marking > 0).

[eutro]

It sounds like this comment is no longer true?

[gasche]

Thanks! I fixed the comment.