YJIT: Add --yjit-pause and RubyVM::YJIT.resume

[maximecb]

This allows booting YJIT in a suspended state. We chose to add a new command line option as opposed to simply allowing YJIT.resume to work without any command line option because it allows for combining with YJIT tuning command line options. It also simpifies implementation.

Co-authored by Alan, Kokubun & Maxime

---

Benchmark results (locally on M1/Rosetta):

./run_benchmarks.rb -e "normal::ruby --yjit" -e "paused::ruby --yjit-pause" --rss liquid-render railsbench hexapdf liquid-c activerecord mail psych-load ruby-lsp sequel

normal: ruby 3.3.0dev (2023-03-28T16:04:45Z runtime-enable-yjit bee3d1410a) +YJIT dev [x86_64-darwin22]
paused: ruby 3.3.0dev (2023-03-28T16:04:45Z runtime-enable-yjit bee3d1410a) +YJIT dev [x86_64-darwin22]

-------------  -----------  ----------  ---------  -----------  ----------  ---------  -------------  --------------
bench          normal (ms)  stddev (%)  RSS (MiB)  paused (ms)  stddev (%)  RSS (MiB)  normal/paused  paused 1st itr
activerecord   47.4         1.8         82.8       44.8         1.8         60.8       1.06           0.44          
hexapdf        2485.5       3.9         840.1      2483.7       1.9         862.6      1.00           1.01          
liquid-c       63.6         3.4         66.3       63.1         2.5         55.4       1.01           1.08          
liquid-render  137.4        3.4         60.6       136.3        2.2         48.6       1.01           0.95          
mail           120.8        1.9         103.0      118.7        1.3         89.6       1.02           1.03          
psych-load     1956.3       0.2         46.5       1982.8       0.5         43.0       0.99           0.99          
railsbench     1679.0       1.3         213.1      1653.4       1.2         169.8      1.02           1.17          
ruby-lsp       142.0        63.6        296.6      130.5        77.1        279.0      1.09           1.08          
sequel         71.0         1.5         64.9       68.4         1.2         45.2       1.04           0.41          
-------------  -----------  ----------  ---------  -----------  ----------  ---------  -------------  --------------

Seems worth upstreaming as this will facilitate further testing (e.g. on SFR) and this change is not hard to revert. The RSS seems to be almost always significantly less, and performance a little bit better. We should ignore the ruby-lsp performance data point since this benchmark is so noisy.

[k0kubun]

I think this is one of the options we want users to use. Can we update the help in ruby.c?

[k0kubun]

Benchmark results (locally on M1/Rosetta):

I got a bit confused, but given that it's getting faster while paused on boot, I guess you put RubyVM::YJIT.resume at the beginning of run_benchmark?

[maximecb]

I think this is one of the options we want users to use. Can we update the help in ruby.c?

I was thinking we could leave it undocumented for a few weeks, until we've had time to test it on SFR/Core. WDYT?

I got a bit confused, but given that it's getting faster while paused on boot, I guess you put RubyVM::YJIT.resume at the beginning of run_benchmark?

I did

[k0kubun]

I was thinking we could leave it undocumented for a few weeks, until we've had time to test it on SFR/Core. WDYT?

That makes sense too 👍