Add parallel marking to GC

[jpsamaroo]

This is a first pass at multithreading parts of GC marking. The intended approach is simple: Store a "recruitment location" (function address) into a global, and have non-leader threads (those not running _jl_gc_collect directly) check for this global to be set to non-NULL. Threads that encounter a valid recruitment location will set their GC state to the new JL_GC_STATE_PARALLEL (name subject to change), and will call into the recruitment location. The only available recruitment function, gc_mark_loop_recruited, will put the thread to work in gc_mark_loop, assuming that the GC leader thread setup some work for this thread to do.

I'm posting this PR early because I need some help figuring out what I'm doing wrong. So far, the recruitment mechanism seems to work well, but I haven't figured out what things I need to setup for each thread before letting them enter the mark loop.

I've also got some code commented out, which would (presumably, if it worked) allow threads to steal work from each other. That's probably not required for this PR, but I started on it in case it might be useful.

N.B. I am bad at atomics, thread safety, and GC semantics, so please point out what terrible things I have done here!

@vchuravy @vtjnash

[jpsamaroo]

@yuyichao what's the logic behind the check

julia/src/gc.c

Lines 2706 to 2707 in 5118a1b

	if (__unlikely(!jl_is_datatype(vt)))
	gc_assert_datatype_fail(ptls, vt, sp);

? I'm hitting this a lot, and I'm not sure I understand the reason why it can fail.

[codecov]

Codecov Report

Merging #41760 (5118a1b) into master (2fbeef8) will increase coverage by 0.03%.
The diff coverage is 77.91%.

❗ Current head 5118a1b differs from pull request most recent head 7f4646d. Consider uploading reports for the commit 7f4646d to get more accurate results

@@            Coverage Diff             @@
##           master   #41760      +/-   ##
==========================================
+ Coverage   80.02%   80.06%   +0.03%     
==========================================
  Files         350      350              
  Lines       73943    74011      +68     
==========================================
+ Hits        59170    59254      +84     
+ Misses      14773    14757      -16     

Impacted Files	Coverage Δ
base/compiler/validation.jl	5.45% <0.00%> (-0.06%)	⬇️
base/libuv.jl	68.25% <ø> (ø)
base/set.jl	84.93% <ø> (ø)
stdlib/Distributed/src/managers.jl	44.77% <0.00%> (+10.44%)	⬆️
stdlib/InteractiveUtils/src/macros.jl	89.18% <ø> (ø)
stdlib/REPL/src/REPLCompletions.jl	25.58% <0.00%> (ø)
base/compiler/abstractinterpretation.jl	70.36% <48.14%> (-0.35%)	⬇️
base/compiler/typelattice.jl	81.33% <50.00%> (ø)
base/io.jl	88.18% <66.66%> (-0.71%)	⬇️
stdlib/Distributed/src/cluster.jl	71.94% <70.00%> (+6.20%)	⬆️
... and 69 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2fbeef8...7f4646d. Read the comment docs.

[yuyichao]

It just means that you are reading garbage somehow.

[ViralBShah]

Ok to close this?

[jpsamaroo]

@ViralBShah we have someone actively working on getting this PR up to date, who may end up pushing to this PR or another one (in which case we'll supersede this). I'd like to keep this open until that happens, just so people have something to watch for progress updates.

[jpsamaroo]

Superseded by #44643