Problem when method returns Regexp capture group using special variable $1

[jbaiza]

Environment

Initially we found the problem using Jruby 9.2.1.0, but it's also actual in 9.1.7.0 and in the latest version.

• JRuby version (jruby -v) and command line (flags, JRUBY_OPTS, etc)

jruby 9.2.5.0 (2.5.0) 2018-12-06 6d5a228 Java HotSpot(TM) 64-Bit Server VM 25.121-b13 on 1.8.0_121-b13 +jit [darwin-x86_64]

• Operating system and platform (e.g. uname -a)

Darwin JBA-MacBook-Pro.local 18.2.0 Darwin Kernel Version 18.2.0: Fri Oct  5 19:41:49 PDT 2018; root:xnu-4903.221.2~2/RELEASE_X86_64 x86_64

Other relevant info you may wish to add:

• Installed or activated gems
  N/A
• Application/framework version (e.g. Rails, Sinatra)
  N/A
• Environment variables
  N/A

Test code:

def test_match(string)
  if string =~ /(.)/
    $1
  end
end

1000000.times { test_match("a") } 

test_match("b")

Expected Behavior

• function test_match call always returns value (as in Ruby MRI or if match is used instead of =~):

jruby-9.2.5.0 :026 > test_match("b")
 => "b"

Actual Behavior

After some iterations returned result always is nil:

jruby-9.2.5.0 :009 > test_match("b")
 => nil
jruby-9.2.5.0 :010 > test_match("b")
 => nil

Full console output:
test_script.log

[enebo]

I can get this to happen with --dev so this has to involve more than just JIT. Weirdly with -X+C I can get it to happen with no repetition in the front whereas --dev passes. So maybe 2 things wrong? This is pretty serious regression :|

[headius]

This is a problem in IR when adding frame push/pop instructions. The pop is happening before the value has been captured, so it ends up nil (or actually, it gets whatever the next-highest heap frame has for $1.

From the IR along the "then" path for the if:

  1:                pop_backref_frame
  2:                return(nth<$1>)

[headius]

Pushed a PR that should fix this. I have not written a test yet.

I do not know when this change was introduced, but it might never have worked right in the optimized "call protocol" execution modes. The problem was that frame-local implicit variables like $~ and $_ were not captured (in some circumstances) before the frame got popped, resulting in reading those values from a previous frame when they were returned. Only "binding" variables (those stored on the heap in DynamicScope instances) were safely captured before pop.

[headius]

This was introduced in 8a04861 back in 2015. We may have been moving the frame variables $~ and $_ around a bit at the time (i.e. they might have been living on the scope/binding instead), or this was just an oversight.

[headius]

@subbuss Can you review my patch? I already merged it since I'm pretty sure it's correct, but you originally removed the requireFrame and I want to make sure I didn't miss some detail.

This will add an additional Copy instruction to all returns of a non-temp or literal from a method with a heap frame. I'm not sure how much of an impact this really is, since most non-trivial data loads go immediately into a temp variable.

[subbuss]

Left a comment on the patch.