YJIT: Ruby 4.0 YJIT is slower than Ruby 3.4 YJIT on chunky-png, liquid-c, liquid-render, and mail

[k0kubun]

liquid-c, liquid-render, and mail benchmarks seem to show a significant regression in Ruby 4.0 YJIT. chunky-png also seems to be slower with Ruby 4.0 YJIT.

@jeremyevans mentioned this when I talked with him at SFRuby, and I didn't have a good explanation. We should investigate which change triggered this, and ideally, fix it before the Ruby 4.0 release.

[k0kubun]

For liquid-render, looking at the "Time" tab of https://rubybench.github.io/benchmarks/ruby-bench.html#liquid-render, 2025-11-01 seems a bit suspicious ruby/ruby@17e95cb...17a7b4e

[k0kubun]

For liquid-c https://rubybench.github.io/benchmarks/ruby-bench.html#liquid-c, 2025-10-30 ~ 2025-11-01 seems like a problematic range ruby/ruby@6e618a2...17a7b4e

[k0kubun]

mail is kind of weird.

On https://rubybench.github.io/benchmarks/ruby-bench.html#mail, it feels like this regression was introduced yesterday ruby/ruby@167c3db...aa9e15c.

But on speed.ruby-lang.org https://speed.ruby-lang.org/stats-timeline.html#yjit_speedup+mail, I'm not sure if we can conclude anything out of the unstable results:

[k0kubun]

liquid-c: https://speed.ruby-lang.org/benchmark_timeline.html#liquid-c

I feel like the interpreter also had a regression and it's just copied to YJIT results.

[k0kubun]

liquid-render: https://speed.ruby-lang.org/benchmark_timeline.html#liquid-render

probably "the interpreter also had a regression and it's just copied to YJIT results" as well?

[k0kubun]

mail: https://speed.ruby-lang.org/benchmark_timeline.html#mail

This looks pretty bad. Given that the Ruby 3.4 results were very stable, the unstable results are probably not the benchmark or the benchmark server's fault, but Ruby 4.0 interpreter's.

[k0kubun]

chunky-png: https://speed.ruby-lang.org/benchmark_timeline.html#chunky-png

Something weird is going on at the Oct 9th point, but the latest results seem to be in the same situation.

[k0kubun]

So I think it's probably an interpreter regression, not YJIT's. Since speed.ruby-lang.org shows only Auguest-or-later results and rubybench.github.io shows only July-or-later results, I'm not sure when it started.

[tekknolagi]

Is this write barrier related potentially? ruby@9b87a0b

[k0kubun]

For the #885 (comment) point, it could be. But a bigger issue is that we had an interpreter regression long ago, which was at least before Aug 5th. I think we want to figure that out first.

[jhawthorn]

Is this write barrier related potentially? ruby@9b87a0b

I don't think so. I'm expecting that only to be called on classes with multiple Ruby::Box

[tekknolagi]

@jhawthorn can you link the relevant commits/PRs? And then can we close?