Build failure with LLVM 3.4

[nalimilan]

I've made some progress in building an RPM package on RHEL6, but I'm getting a segfault at the end of the construction of the system image. Does that sound familiar to anybody? I've installed devtoolset-2, so it's using GCC 4.8, and LLVM is 3.4.2.

precompile.jl
signal (11): Segmentation fault
jl_gen_llvm_gv_array at /builddir/build/BUILD/julia/src/cgutils.cpp:235
_ZNKSs7_M_dataEv at /opt/rh/devtoolset-2/root/usr/include/c++/4.8.2/bits/basic_string.h:293
julia_trampoline at /builddir/build/BUILD/julia/src/init.c:1034
unknown function (ip: 4199629)
__libc_start_main at /lib64/libc.so.6 (unknown line)
unknown function (ip: 4199685)
/bin/sh: line 3: 17626 Segmentation fault      /builddir/build/BUILD/julia/build/usr/bin/julia --build /builddir/build/BUILD/julia/build/usr/lib64/julia/sys -J/builddir/build/BUILD/julia/build/usr/lib64/julia/$([ -e /builddir/build/BUILD/julia/build/usr/lib64/julia/sys.ji ] && echo sys.ji || echo sys0.ji) -f sysimg.jl
*** This error is usually fixed by running 'make clean'. If the error persists, try 'make cleanall'. ***
make[1]: Leaving directory `/builddir/build/BUILD/julia'
make[1]: *** [/builddir/build/BUILD/julia/build/usr/lib64/julia/sys.o] Error 1
make: *** [release] Error 2

http:https://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/epel-6-x86_64/julia-0.4.0dev-0.20140930.fc20/build.log

Cc: @crayxt

[crayxt]

Same issue with LLVM_SVN. You see there was rather big merge #8420

[crayxt]

On centos7, and cento6 with stock compiler and LLVM 3.3 compilation succeeds.

[JeffBezanson]

LLVM 3.4 could certainly be the issue here, or more accurately, our level of compatibility with it.

[crayxt]

Here's what git bisect says

c67837ca8e69505bf684cd71ac11d69a178c86c9 is the first bad commit
commit c67837ca8e69505bf684cd71ac11d69a178c86c9
Author: Jeff Bezanson <[email protected]>
Date:   Fri Sep 19 10:16:28 2014 -0400

    get dates tests passing with checked signed<->unsigned conversion

:040000 040000 e82473eb2a61e8d2c4b940699b97d309211efc15 510ecb2cf580a67c61ceaf63d06db4bc7045c0f8 M      base
:040000 040000 19c3a64cf2deb31286fd6c50ffeefb83221cfeb3 0d1121276be3750f7923b94ca851986ba3199a1d M      src
:040000 040000 9e0b3d75b77f5acdf4894ba495463132fdf81b39 de9ad7a6c3332d0c5ae89e355ced99fe48143a39 M      test

[staticfloat]

I see segfaults like this intermittently. If you build this over and over again, does the segfault reliably happen?

[Keno]

I have seen this too on LLVM 3.6.

[nalimilan]

@staticfloat I've run it three times, and it failed every time.

[nalimilan]

I also see this on Fedora 20 (again with LLVM 3.4):

precompile.jl
fatal: error thrown and no exception handler available.
MemoryError()
*** This error is usually fixed by running 'make clean'. If the error persists, try 'make cleanall'. ***
make[1]: Leaving directory `/builddir/build/BUILD/julia'
make[1]: *** [/builddir/build/BUILD/julia/build/usr/lib64/julia/sys.o] Error 1
make: *** [release] Error 2

http:https://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/fedora-20-x86_64/julia-0.4.0dev-0.20141002.fc20/build.log

The error is different but it seems to appear at the same step.

[Keno]

Something is really broken with MCJIT versions of julia at the moment. This really needs to be fixed.

[nalimilan]

Today some of the builds passed, and only failed in the first test:

/builddir/build/BUILD/julia/build/usr/bin/julia --check-bounds=yes -f ./runtests.jl core
symbol could not be found jl_sysimg_gvars (-1): /builddir/build/BUILD/julia/build/usr/bin/../lib64/julia/sys.so: undefined symbol: jl_sysimg_gvars
symbol could not be found jl_globalUnique (-1): /builddir/build/BUILD/julia/build/usr/bin/../lib64/julia/sys.so: undefined symbol: jl_globalUnique
/bin/sh: line 1: 10184 Segmentation fault      /builddir/build/BUILD/julia/build/usr/bin/julia --check-bounds=yes -f ./runtests.jl core
make: *** [core] Error 139

http:https://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/fedora-20-x86_64/julia-0.4.0dev-0.20141003.fc20/build.log

But another one is still the same, so maybe there's indeed a random component:
http:https://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/fedora-rawhide-x86_64/julia-0.4.0dev-0.20141003.fc20/build.log

[Keno]

Yep, seen that too. For what it's worth, building with assertions gives:

julia: /home/kfischer/julia-master/deps/llvm-svn/include/llvm/ADT/StringMap.h:320: bool llvm::StringMap<llvm::Value *, llvm::MallocAllocator>::insert(MapEntryTy *) [ValueTy = llvm::Value *, AllocatorTy = llvm::MallocAllocator]: Assertion `NumItems + NumTombstones <= NumBuckets' failed.

though I think memory corruption may be more likely.

[Keno]

Actually, I think I may know what the issue is. Hold on.

[nalimilan]

@Keno Looks good now, thanks! Now I need to find out why tests fail on 32-bit only.