Bug #14949: test_jit.rb test_unload_units fails on 32-bit

• Status: Closed
• Assignee: k0kubun (Takashi Kokubun)
• Target version: -
• Backport: 2.3: UNKNOWN, 2.4: UNKNOWN, 2.5: UNKNOWN

[normalperson (Eric Wong)]

I get multiple definition warnings from this test. It only
happens on my 32-bit systems (both VM and real HW) on Debian 9.
Attached log was made with below patch:

--- a/test/ruby/test_jit.rb
+++ b/test/ruby/test_jit.rb
@@ -541,32 +541,33 @@ def test_unload_units
     Dir.mktmpdir("jit_test_unload_units_") do |dir|
       # MIN_CACHE_SIZE is 10
       out, err = eval_with_jit({"TMPDIR"=>dir}, "#{<<~"begin;"}\n#{<<~'end;'}", verbose: 1, min_calls: 1, max_cache: 10)
       begin;
         10.times do |i|
           eval(<<-EOS)
             def mjit#{i}
               print #{i}
             end
             mjit#{i}
           EOS
         end
       end;
       assert_equal('0123456789', out)
       compactions, errs = err.lines.partition do |l|
         l.match?(/\AJIT compaction \(\d+\.\dms\): Compacted \d+ methods ->/)
       end
       assert_match(/\A#{JIT_SUCCESS_PREFIX}: block in <main>@-e:/, errs[0])
       9.times do |i|
         assert_match(/\A#{JIT_SUCCESS_PREFIX}: mjit#{i}@\(eval\):/, errs[i + 1])
       end
+      warn "errs: #{errs.join}"
       assert_equal("Too many JIT code -- 1 units unloaded\n", errs[10])
       assert_match(/\A#{JIT_SUCCESS_PREFIX}: mjit9@\(eval\):/, errs[11])
 
       # On --jit-wait, when the number of JIT-ed code reaches --jit-max-cache,
       # it should trigger compaction.
       assert_equal(2, compactions.size)
 
       # verify .o files are deleted on unload_units
       assert_send([Dir, :empty?, dir])
     end
   end

[k0kubun (Takashi Kokubun)]

Eric, thanks to notify this. To fix it, I want to know the content of MJIT header files. I could see debian i686-linux on rubyci was failing in the same reason, but it looks a little hard to know the path of MJIT header from test_jit.rb for debugging.

If it's reproductive on your local environment, could you attach the rb_mjit_header.h in the build directory and/or rb_mjit_min_header-2.6.0.h in somewhere?

[normalperson (Eric Wong)]

• File rb_mjit_header.h rb_mjit_header.h added

Attached is rb_mjit_header.h from a 32-bit userspace on 64-bit kernel/HW
and mjit_header_i686_native.h from a 32-bit machine

[normalperson (Eric Wong)]

• File mjit_header_i686_native.h mjit_header_i686_native.h added

Header from native 32-bit machine (can't attach multiple files
w/o JS, it seems)

[MSP-Greg (Greg L)]

I'm also seeing a lot of error text, 395 items in the errs array. Asserts pass up to the assert_equal at line 562. The matches on errs[10] & errs[11] exist, but they're at 201 & 202 respectively.

I've got rb_mjit_min_header-2.6.0.h, I can attach it if you'd like, it's about 5 MB in size.

Also, to get it to run, I changed the current loop to what's below.

Thanks, Greg

begin;
  1.times do |i|
    eval 'def mjit0 ; print 0 end ; mjit0'
    eval 'def mjit1 ; print 1 end ; mjit1'
    eval 'def mjit2 ; print 2 end ; mjit2'
    eval 'def mjit3 ; print 3 end ; mjit3'
    eval 'def mjit4 ; print 4 end ; mjit4'
    eval 'def mjit5 ; print 5 end ; mjit5'
    eval 'def mjit6 ; print 6 end ; mjit6'
    eval 'def mjit7 ; print 7 end ; mjit7'
    eval 'def mjit8 ; print 8 end ; mjit8'
    eval 'def mjit9 ; print 9 end ; mjit9'
  end
end;

[k0kubun (Takashi Kokubun)]

• Status changed from Open to Closed

Applied in changeset trunk|r64138.

---

transform_mjit_header.rb: accept having attribute

between a type of return value and function identifier. Like:

rb_control_frame_t *
__attribute__ ((__fastcall__)) rb_vm_opt_struct_aref(rb_execution_context_t *ec, rb_control_frame_t *reg_cfp)
{
    ...
}

[Bug #14949]

[k0kubun (Takashi Kokubun)]

Eric, thanks to the header information, I could fix at least the RubyCI's Debian 7.11 i686.