Feature #18440: YJIT is enabled if any YJIT tuning options are set

• Status: Feedback
• Assignee: -
• Target version: -

[georgeclaghorn (George Claghorn)]

I was testing YJIT in a Rails app with RUBYOPT="--yjit --yjit-exec-mem-size=32". I saw some weird issues around Rails view caching, so I attempted to temporarily disable YJIT by removing --yjit and leaving the tuning options in place (RUBYOPT="--yjit-exec-mem-size=32"). However, YJIT remained enabled until I also removed --yjit-exec-mem-size=32.

$ ruby -e 'puts RubyVM::YJIT.runtime_stats.inspect'
nil

$ ruby --yjit-exec-mem-size -e 'puts RubyVM::YJIT.runtime_stats.inspect'
{:inline_code_size=>378495, :outlined_code_size=>311079}

I expected removing --yjit to disable YJIT and was surprised by this.

[georgeclaghorn (George Claghorn)]

• ruby -v changed from 3.1.0 to ruby 3.1.0p0 (2021-12-25 revision fb4df44d16) [arm64-darwin21]

[georgeclaghorn (George Claghorn)]

Sorry, there's a small typo in the example:

$ ruby --yjit-exec-mem-size=32 -e 'puts RubyVM::YJIT.runtime_stats.inspect'
{:inline_code_size=>67775, :outlined_code_size=>52667}

[georgeclaghorn (George Claghorn)]

From the code, this appears to be intentional. setup_yjit_options returns true if any of the --yjit-* options are provided:

            else if (strcmp("yjit", s) == 0 || setup_yjit_options(s, &opt->yjit)) {
#if USE_MJIT
                FEATURE_SET(opt->features, FEATURE_BIT(yjit));
#else
                rb_warn("Ruby was built without JIT support");
#endif
            }

[k0kubun (Takashi Kokubun)]

• Tracker changed from Bug to Feature
• ruby -v deleted (ruby 3.1.0p0 (2021-12-25 revision fb4df44d16) [arm64-darwin21])
• Backport deleted (2.6: UNKNOWN, 2.7: UNKNOWN, 3.0: UNKNOWN, 3.1: UNKNOWN)

From the code, this appears to be intentional.

I think so. At least for MJIT in 2.6~3.0, we used to show:

  --jit           enable JIT with default options (experimental)
  --jit-[option]  enable JIT with an option (experimental)

but we ended up removing it in 3.1 for MJIT and YJIT. Maybe we should add such descriptions again for clarity.

[Eregon (Benoit Daloze)]

I think only enabling on explicit --jit/--yjit/--mjit is better because it's more intuitive.
If e.g. --yjit-exec-mem-size=32 is in RUBYOPT it seems unexpected that it would enable YJIT without any --yjit anywhere.

[k0kubun (Takashi Kokubun)]

• Status changed from Open to Feedback

[byroot (Jean Boussier)]

For what it's worth, I would expect --yjit-* options not to turn on YJIT. I agree that it's surprising, I would expect only --yjit or RUBY_YJIT_ENABLE=1 to actually enable it.