JIT mode execution does not call trace functions

This is a known issue....and it won't be in RC2, but it MUST be done for 1.0. I had left tracing out of the compiled code when the compiler was not enabled by default, but within the past few weeks we found the compiler was working well enough to enable permanently. It's not a difficult thing to add, and I'll add it.

I committed a fix for this that basically reverts JITted methods to interpreted mode, effectively allowing them to trace correctly. Since we're already written off trace-based debugging as inefficient, I believe this is an acceptable fix for 1.0 to work. When we implement JRuby-debug, I think it may be worth looking at a better way to do line-by-line tracing than checking for a trace at every line (yuck). Perhaps someone familiar with the built-in Java debugging hooks would be able to help?

Anyway, I'm marking this resolved, since it is.

They seems to be fixed. set_trace_func seems to be called. But when you try the simplest sample:

$ echo "puts 'Hello'" > tester.rb 
$ jruby -rdebug tester.rb
Debug.rb
Emacs support available.

Hello

When you try to watch the parameter passed to the trace function, file and line are empty (or zero).

$ cat a.rb
require 'stf.rb'
a = 5
b = 4
a+b
$ cat stf.rb 
set_trace_func proc { |_, file, line, _, _, _|
  puts "#{file}:#{line}"
}
$ jruby a.rb
:0
:0

So it is not usable with this regression. Not sure whether to file a new one or to reopen this one... reopening.

I think this is a problem with the top-level script getting compiled. Since there's no trace function set when it starts, JRuby compiles it. Since we're still in the middle of executing it when the trace func is enabled, it stops working.

Martin, wouldn't it just be possible for you to have JRuby running from the debugger disable the JIT completely? It's just one system property to set: jruby.jit.enabled=false. I think that's probably going to be the best bet here. I don't want to hack up the JIT any more than I have to to get this working.

I've replied globally since I'm not only one interested. More people needs to know this (e.g. RDT guys). So yup, works for me/us. But not for everybody else... more on dev ML.

I gave it a quick try and debugging with JRuby works very decently now (with the switch). Actually I think the best in last 4 months or so! There have to be more fixes which made it works this way. Nice work, thanks.

I have committed a change such that if you specify -rdebug it will not try to compile anything (basically, it's the same check as in the JIT, but it does it an additional time before compiling the startup (main) script). I think that will resolve this issue for cases where you want to run -rdebug, but it doesn't resolve the script you showed above (Martin). For that you need to turn off the JIT.

Is this acceptable?

Yes. That's quite good for us. I'll use the JIT switch (jruby.jit.enabled=false) always in debug-commons-java so I depend on it to be working. Is that OK for the near future to depend on it?
Also CLI debugger now works again. So for me it is fixed enough. Thanks.

Yes Martin, jruby.jit.enabled=false will be a supported mode of execution into the future. Even our unit test suite runs twice, once with the JIT disabled. We will support fully-interpreted execution well into the future, which means you can count on that continuing to work.

Marking as fixed, since Martin's happy.