Build failure with GLIBCXX_ASSERTIONS

[nalimilan]

The Fedora RPM nightlies have started to fail between April 10th and April 15th EDIT: actually that error was present since January in Fedora rawhide. They use USE_SYSTEM_LLVM=1, and it happens both with LLVM 3.9 and LLVM 6 (with all Julia patches included). EDIT2: this only happens when building with GLIBCXX_ASSERTIONS enabled, and is unrelated to USE_SYSTEM_LLVM.

Any idea what might be going on? What can I do to debug this?

Precompile: 92.720144 seconds 17.0215%
signal (6): Aborted
in expression starting at no file:0
gsignal at /lib64/libc.so.6 (unknown line)
abort at /lib64/libc.so.6 (unknown line)
__replacement_assert at /usr/include/c++/8/x86_64-redhat-linux/bits/c++config.h:2389
operator[] at /usr/include/c++/8/bits/stl_vector.h:932
operator() at /builddir/build/BUILD/julia/src/processor.cpp:405 [inlined]
serialize_target_data at (null):0
serialize_target_data at /builddir/build/BUILD/julia/src/processor.cpp:415
serialize_target_data<9> at /builddir/build/BUILD/julia/src/processor.cpp:425
jl_get_llvm_clone_targets at /builddir/build/BUILD/julia/src/processor_x86.cpp:905
CloneCtx at /builddir/build/BUILD/julia/src/llvm-multiversioning.cpp:348
runOnModule at /builddir/build/BUILD/julia/src/llvm-multiversioning.cpp:1031
_ZN4llvm6legacy15PassManagerImpl3runERNS_6ModuleE at /usr/lib64/llvm6/lib/libLLVM-6.0.so (unknown line)
operator() at /builddir/build/BUILD/julia/src/jitlayers.cpp:1177 [inlined]
jl_dump_native at (null):0
jl_dump_native at /builddir/build/BUILD/julia/src/jitlayers.cpp:1186
jl_write_compiler_output at /builddir/build/BUILD/julia/src/precompile.c:84
jl_atexit_hook at /builddir/build/BUILD/julia/src/init.c:233
main at /builddir/build/BUILD/julia/ui/repl.c:241
__libc_start_main at /lib64/libc.so.6 (unknown line)
_start at /builddir/build/BUILD/julia/build/usr/bin/julia-debug (unknown line)
Allocations: 240731622 (Pool: 240661198; Big: 70424); GC: 643
/bin/sh: line 1: 13399 Aborted                 (core dumped) /builddir/build/BUILD/julia/build/usr/bin/julia-debug -O0 -C "x86-64" --output-o /builddir/build/BUILD/julia/build/usr/lib64/julia/sys-debug-o.a --startup-file=no --warn-overwrite=yes --sysimage /builddir/build/BUILD/julia/build/usr/lib64/julia/basecompiler.ji 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]: *** [Makefile:204: /builddir/build/BUILD/julia/build/usr/lib64/julia/sys-debug-o.a] Error 1
make: *** [Makefile:81: julia-sysimg-debug] Error 2

https://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/fedora-28-x86_64/00754012-julia/build.log.gz

[nalimilan]

Actually it also happens without USE_SYSTEM_LLVM=1:
https://copr-be.cloud.fedoraproject.org/results/nalimilan/julia-nightlies/fedora-28-x86_64/00754876-julia/build.log.gz

It's probably due to the use of gcc 8.1.1, as the error doesn't happen on older distros which use gcc 7 or lower, and it appeared in January, i.e. when gcc 8 was introduced in Fedora rawhide.

[CodeLenz]

I have been using gcc 8.1.1 (gcc version 8.1.1 20180502 (Red Hat 8.1.1-1) (GCC)) with Julia master since the release of fedora 28. At first, I had to clean everything and build all the dependencies from scratch.

My Make.user is

OPENBLAS_TARGET_ARCH=HASWELL
MARCH=native
LLVM_VER=6.0.0

[nalimilan]

After rebuilding Julia and its deps dozens of times, it turns out that all that's needed to reproduce the failure is to build LLVM with GLIBCXX_ASSERTIONS:

make -C deps install-llvm CFLAGS="-Wp,-D_GLIBCXX_ASSERTIONS" CXXFLAGS="-Wp,-D_GLIBCXX_ASSERTIONS"
make

Apparently the error may have revealed a real bug, since that flag enables checks for array indices. See the similar error found with MariaDB.

[ViralBShah]

Closing as we have moved to newer LLVM versions. Perhaps best to file a new issue if seen again.