Bug #15522: TestJIT#test_compile_insn_local fails on aarch64 RHEL7

• Status: Closed
• Assignee: k0kubun (Takashi Kokubun)
• Target version: -
• ruby -v: ruby 2.6.0p0 (2018-12-25 revision 66547) [aarch64-linux]
• Backport: 2.4: DONTNEED, 2.5: DONTNEED, 2.6: DONE

[vo.x (Vit Ondruch)]

Trying to build Ruby 2.6 on RHEL7, I observe the following test failure on RHEL7, but just on aarch64. The other platforms pass just fine:

  1) Failure:
TestJIT#test_compile_insn_local [/builddir/build/BUILD/ruby-2.6.0/test/ruby/test_jit.rb:64]:
Expected 3 times of JIT success, but succeeded 2 times.
script:
```
def foo
  a = 0
  [1, 2].each do |i|
    a += i
    [3, 4].each do |j|
      a *= j
    end
  end
  a
end
print foo
```
stderr:
```
JIT success (276.6ms): foo@-e:2 -> /tmp/_ruby_mjit_p20163u0.c
JIT success (347.3ms): block in foo@-e:4 -> /tmp/_ruby_mjit_p20163u1.c
MJIT warning: failure in loading code from '/tmp/_ruby_mjit_p20163u2.so': /tmp/_ruby_mjit_p20163u2.so: undefined symbol: __multi3
Successful MJIT finish
```
.
<3> expected but was
<2>.

I suspect this must be some combination of architecture/compiler (gcc (GCC) 4.8.5 20150623 (Red Hat 4.8.5-16)), because I don't observe similar issues on Fedora.

[vo.x (Vit Ondruch)]

Also, it would be nice if the JIT output used different markup, which does not collide with Redmine markup :/

[sharkcz (Dan Horák)]

IMHO either something calls __multi3 and doesn't link to libgcc or gcc emits call to __multi3 and it's not implemented in libgcc.
Would be useful to see /tmp/_ruby_mjit_p20163u1.c source and the command line used to compile/link it.

[vo.x (Vit Ondruch)]

The smaller reproducer probably is:

$ ruby --disable-gems --jit-verbose=10 --jit-min-calls=1 --jit-debug --jit-wait --jit-save-temps -e "
    begin
      def foo
        a = 0
        [1, 2].each do |i|
          a += i
          [3, 4].each do |j|
            a *= j
          end
        end
        a
      end

      print foo
    end"

[vo.x (Vit Ondruch)]

• File _ruby_mjit_p207u2.c _ruby_mjit_p207u2.c added

This should be the C code (although generated on my x86_64, because I don't have aarch64 readily available).

[k0kubun (Takashi Kokubun)]

Also, it would be nice if the JIT output used different markup, which does not collide with Redmine markup :/

Yeah. I'll change that later.

This should be the C code (although generated on my x86_64, because I don't have aarch64 readily available).

How did you get the output in the ticket description? If possible, I want you to upload rb_mjit_min_header-2.6.0.h under install directory and the exact /tmp/_ruby_mjit_p20163u1.c whose .so can't be loaded.

[k0kubun (Takashi Kokubun)]

• Assignee set to k0kubun (Takashi Kokubun)

[vo.x (Vit Ondruch)]

• Assignee deleted (k0kubun (Takashi Kokubun))

k0kubun (Takashi Kokubun) wrote:

Also, it would be nice if the JIT output used different markup, which does not collide with Redmine markup :/

Yeah. I'll change that later.

Thx

This should be the C code (although generated on my x86_64, because I don't have aarch64 readily available).

How did you get the output in the ticket description?

That is from builder which is not easily accessible. I can throw in SRPM and get resulting logs and RPMs, but I cannot easily access the intermediate results :/

If possible, I want you to upload rb_mjit_min_header-2.6.0.h under install directory and the exact /tmp/_ruby_mjit_p20163u1.c whose .so can't be loaded.

I will see what I can do about it. Hopefully, @sharkcz will be able to help :)

[k0kubun (Takashi Kokubun)]

• Status changed from Open to Feedback

Thank you. I'll wait for that first.
Also having logs with --jit-verbose=2 version of https://bugs.ruby-lang.org/issues/15522#note-3 would be helpful.

P.S. The output format is changed in r66781.

[vo.x (Vit Ondruch)]

• File rb_mjit_min_header-2.6.0.h rb_mjit_min_header-2.6.0.h added
• File mjit_multi3.tgz mjit_multi3.tgz added
• Status changed from Feedback to Open

So here is the output:

$ ruby --disable-gems --jit-verbose=2 --jit-min-calls=1 --jit-debug --jit-wait --jit-save-temps -e "
    begin
      def foo
        a = 0
        [1, 2].each do |i|
          a += i
          [3, 4].each do |j|
            a *= j
          end
        end
        a
      end

      print foo
    end"
MJIT: CC defaults to /usr/bin/gcc
MJIT: tmp_dir is /tmp
Creating precompiled header
Starting process: /usr/bin/gcc /usr/bin/gcc -w -Wfatal-errors -fPIC -shared -w -pipe -ggdb3 -o /tmp/_ruby_mjit_hp111u0.h.gch /opt/rh/rh-ruby26/root/usr/include/rb_mjit_min_header-2.6.0.h
start compilation: foo@-e:3 -> /tmp/_ruby_mjit_p111u0.c
Starting process: /usr/bin/gcc /usr/bin/gcc -w -Wfatal-errors -fPIC -shared -w -pipe -ggdb3 -o /tmp/_ruby_mjit_p111u0.o /tmp/_ruby_mjit_p111u0.c -c -Wl,-z,relro -nostartfiles -nodefaultlibs -nostdlib
Starting process: /usr/bin/gcc /usr/bin/gcc -shared -Wfatal-errors -fPIC -shared -w -pipe -ggdb3 -o /tmp/_ruby_mjit_p111u0.so /tmp/_ruby_mjit_p111u0.o -Wl,-z,relro -nostartfiles -nodefaultlibs -nostdlib
MJIT warning: failure in loading code from '/tmp/_ruby_mjit_p111u0.so': /tmp/_ruby_mjit_p111u0.so: undefined symbol: __multi3
start compilation: block in foo@-e:5 -> /tmp/_ruby_mjit_p111u1.c
Starting process: /usr/bin/gcc /usr/bin/gcc -w -Wfatal-errors -fPIC -shared -w -pipe -ggdb3 -o /tmp/_ruby_mjit_p111u1.o /tmp/_ruby_mjit_p111u1.c -c -Wl,-z,relro -nostartfiles -nodefaultlibs -nostdlib
Starting process: /usr/bin/gcc /usr/bin/gcc -shared -Wfatal-errors -fPIC -shared -w -pipe -ggdb3 -o /tmp/_ruby_mjit_p111u1.so /tmp/_ruby_mjit_p111u1.o -Wl,-z,relro -nostartfiles -nodefaultlibs -nostdlib
MJIT warning: failure in loading code from '/tmp/_ruby_mjit_p111u1.so': /tmp/_ruby_mjit_p111u1.so: undefined symbol: __multi3
start compilation: block (2 levels) in foo@-e:7 -> /tmp/_ruby_mjit_p111u2.c
Starting process: /usr/bin/gcc /usr/bin/gcc -w -Wfatal-errors -fPIC -shared -w -pipe -ggdb3 -o /tmp/_ruby_mjit_p111u2.o /tmp/_ruby_mjit_p111u2.c -c -Wl,-z,relro -nostartfiles -nodefaultlibs -nostdlib
Starting process: /usr/bin/gcc /usr/bin/gcc -shared -Wfatal-errors -fPIC -shared -w -pipe -ggdb3 -o /tmp/_ruby_mjit_p111u2.so /tmp/_ruby_mjit_p111u2.o -Wl,-z,relro -nostartfiles -nodefaultlibs -nostdlib
MJIT warning: failure in loading code from '/tmp/_ruby_mjit_p111u2.so': /tmp/_ruby_mjit_p111u2.so: undefined symbol: __multi3
168Stopping worker thread
Successful MJIT finish

Header and the content of /tmp directory (except the .gch, which was huge :/) are attached.

k0kubun (Takashi Kokubun) wrote:

P.S. The output format is changed in r66781.

Nice, thanks!

[sharkcz (Dan Horák)]

I suppose "-nostartfiles -nodefaultlibs -nostdlib" are the reason that libgcc isn't linked into the _ruby_mjit_p111u0.so

__multi3 is provided in /lib64/libgcc_s-4.8.5-20150702.so.1

I guess -nodefaultlibs should be omitted and/or -static-libgcc added. Skipping libgcc seems dangerous in general on any arch.

[k0kubun (Takashi Kokubun)]

• Status changed from Open to Assigned
• Assignee set to k0kubun (Takashi Kokubun)

Header and the content of /tmp directory (except the .gch, which was huge :/) are attached.

Thanks! Much appreciated.

I suppose "-nostartfiles -nodefaultlibs -nostdlib" are the reason that libgcc isn't linked into the _ruby_mjit_p111u0.so

__multi3 is provided in /lib64/libgcc_s-4.8.5-20150702.so.1

Possibly. It's causing issues on #15513 as well... I'll try to fix it.

[vo.x (Vit Ondruch)]

sharkcz (Dan Horák) wrote:

I guess -nodefaultlibs should be omitted and/or -static-libgcc added. Skipping libgcc seems dangerous in general on any arch.

I removed -nodefaultlibs and it changed nothing. But reading GCC options 1 manual, the -nostdlib description, it says:

In other words, when you specify -nostdlib or -nodefaultlibs you should usually specify -lgcc as well.

and indeed after adding -lgcc, the build passed. Also, the -lgcc is already used in some cases 2, maybe it should be used in all cases ...

[k0kubun (Takashi Kokubun)]

• Status changed from Assigned to Closed

Applied in changeset trunk|r66812.

---

mjit_worker.c: pass -lgcc to GCC platforms

using -nodefaultlibs -nostdlib.

I assume libgcc is needed when we use -nostdlib, and it's linked on some
platforms but not linked on some platforms (like aarch64, and possibly
AIX as well) as said in https://wiki.osdev.org/Libgcc :

You can link with libgcc by passing -lgcc when linking your kernel
with your compiler. You don't need to do this unless you pass the
-nodefaultlibs option (implied by -nostdlib)

Also note that -nostdlib is not strictly needed (rather implied
-nodefaultlibs is problematic for Gentoo like Bug#15513, which will be
approached later) but helpful for performance. So I want to keep it for
now.

[Bug #15522]

I'm not trying to add -nodefaultlibs -nostdlib for AIX in this commit
because AIX RubyCI is dead right now, but I'll try to add them again
once RubyCI is fixed.

[k0kubun (Takashi Kokubun)]

@vo.x Could you check if r66812 works?

[k0kubun (Takashi Kokubun)]

• Backport changed from 2.4: UNKNOWN, 2.5: UNKNOWN, 2.6: UNKNOWN to 2.4: DONTNEED, 2.5: DONTNEED, 2.6: REQUIRED

[vo.x (Vit Ondruch)]

k0kubun (Takashi Kokubun) wrote:

@vo.x Could you check if r66812 works?

Actually it is r66811 + r66812. Applying these two patches, the test suite passes on RHEL7 on all supported architectures and it keeps passing on Fedora Rawhide on all supported architectures. Thx for the fix.

[naruse (Yui NARUSE)]

• Backport changed from 2.4: DONTNEED, 2.5: DONTNEED, 2.6: REQUIRED to 2.4: DONTNEED, 2.5: DONTNEED, 2.6: DONE

ruby_2_6 r66849 merged revision(s) 66811,66812,66816.