Different behavior for Array#reject in required file

[Maumagnaguagno]

A few days ago I noticed a behavior change related to Array#reject using GitHub Actions and jruby-head.
Today I was able to obtain the following minimal reproducible example using Linux and RVM with two files:

test_jruby.rb

puts RUBY_DESCRIPTION
puts 'test1'
[[1,2]].reject {|e,| p e}
require_relative 'test_jruby2'

test_jruby2.rb

puts 'test2'
[[1,2]].reject {|e,| p e}

The expected output contains the description strings and the first element, e, of each Array.

Output of Ruby MRI 3.3.1 (expected)

$ ruby test_jruby.rb 
ruby 3.3.1 (2024-04-23 revision c56cd86388) [x86_64-linux]
test1
1
test2
1

Output of JRuby 9.4.7.0 (expected)

$ ruby test_jruby.rb 
jruby 9.4.7.0 (3.1.4) 2024-04-29 597ff08ac1 OpenJDK 64-Bit Server VM 25.402-b06 on 1.8.0_402-b06 +jit [x86_64-linux]
test1
1
test2
1

Output of JRuby head (unexpected output in test2)

$ ruby test_jruby.rb 
jruby 9.4.8.0-SNAPSHOT (3.1.4) 2024-05-05 d7358ea8c5 OpenJDK 64-Bit Server VM 25.402-b06 on 1.8.0_402-b06 +jit [x86_64-linux]
test1
1
test2
[1, 2]

Note that using underscore, [[1,2]].reject {|e,_| p e}, is enough to obtain the correct behavior.
This used to work without the underscore and the difference between main and required files is surprising.

[enebo]

This can just be reduced to:

puts 'test2'
[[1,2]].reject {|e,| p e}

If run by default command-line we unconditionally compile the main file and if compiled this works as expected. If I run it with interpreter (jruby --dev reject1.rb) then we see the wrong binding.

[enebo]

So this is specific to how we dispatch from some enumerable methods (and possibly others) but only when calling into the interpreter. This partially explains why we didn't see this failing. If I just call a proc with |e,| from interp or JIT it works as expected.

[enebo]

In the end one proc parameter method was not updated for the arity "fix" for , procs.

[headius]

Is there a test we can add for this somewhere? I know these JIT-Int-JIT transitions are hard to test. 🫤

[enebo]

@headius this is not a JIT-int-JIT thing but merely it did not work from native -> int for yieldSpecific. I added a regression test which will run as interp. Should be covered.