Segmentation fault on Ubuntu 16.04

[joaocandre]

I'm trying to compile the sample code on a machine with Ubuntu 16.04 LTS (GCC 5.4.0) but it's throwing a segmentation fault error at runtime. The issue seems to be Eigen-related, but I've tried with several Eigen versions and I can't understand the issue. It is strange because on an Arch Linux machine, with the latest gcc version (7.3.1), there is no issue.

Error trace from valgrind:

==21993== Memcheck, a memory error detector
==21993== Copyright (C) 2002-2015, and GNU GPL'd, by Julian Seward et al.
==21993== Using Valgrind-3.11.0 and LibVEX; rerun with -h for copyright info
==21993== Command: ./example
==21993==
==21993==
==21993== Process terminating with default action of signal 11 (SIGSEGV)
==21993== General Protection Fault
==21993== at 0x5027980: _mm256_store_pd (avxintrin.h:842)
==21993== by 0x5027980: pstore<double, __vector(4) double> (PacketMath.h:244)
==21993== by 0x5027980: pstoret<double, __vector(4) double, 32> (GenericPacketMath.h:486)
==21993== by 0x5027980: assignPacket<32, __vector(4) double> (AssignmentFunctors.h:28)
==21993== by 0x5027980: assignPacket<32, 32, __vector(4) double> (AssignEvaluator.h:598)
==21993== by 0x5027980: run (AssignEvaluator.h:386)
==21993== by 0x5027980: call_dense_assignment_loop<Eigen::Matrix<double, -1, 1>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_constant_op<double>, Eigen::Matrix<double, -1, 1> >, Eigen::internal::assign_op<double> > (AssignEvaluator.h:653)
==21993== by 0x5027980: run (AssignEvaluator.h:790)
==21993== by 0x5027980: void Eigen::internal::call_assignment_no_alias<Eigen::Matrix<double, -1, 1, 0, -1, 1>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_constant_op<double>, Eigen::Matrix<double, -1, 1, 0, -1, 1> >, Eigen::internal::assign_op<double> >(Eigen::Matrix<double, -1, 1, 0, -1, 1>&, Eigen::CwiseNullaryOp<Eigen::internal::scalar_constant_op<double>, Eigen::Matrix<double, -1, 1, 0, -1, 1> > const&, Eigen::internal::assign_op<double> const&) (AssignEvaluator.h:747)
==21993== by 0x5091FDD: call_assignment<Eigen::Matrix<double, -1, 1>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_constant_op<double>, Eigen::Matrix<double, -1, 1> >, Eigen::internal::assign_op<double> > (AssignEvaluator.h:712)
==21993== by 0x5091FDD: call_assignment<Eigen::Matrix<double, -1, 1>, Eigen::CwiseNullaryOp<Eigen::internal::scalar_constant_op<double>, Eigen::Matrix<double, -1, 1> > > (AssignEvaluator.h:693)
==21993== by 0x5091FDD: _set<Eigen::CwiseNullaryOp<Eigen::internal::scalar_constant_op<double>, Eigen::Matrix<double, -1, 1> > > (PlainObjectBase.h:682)
==21993== by 0x5091FDD: operator=<Eigen::CwiseNullaryOp<Eigen::internal::scalar_constant_op<double>, Eigen::Matrix<double, -1, 1> > > (Matrix.h:225)
==21993== by 0x5091FDD: libcmaes::CMAParameters<libcmaes::GenoPheno<libcmaes::NoBoundStrategy, libcmaes::NoScalingStrategy> >::initialize_parameters() (cmaparameters.cc:77)
==21993== by 0x5092EAD: libcmaes::CMAParameters<libcmaes::GenoPheno<libcmaes::NoBoundStrategy, libcmaes::NoScalingStrategy> >::CMAParameters(std::vector<double, std::allocator<double> > const&, double const&, int const&, unsigned long const&, libcmaes::GenoPheno<libcmaes::NoBoundStrategy, libcmaes::NoScalingStrategy> const&) (cmaparameters.cc:48)
==21993== by 0x41254D: main (in /home/**/example)
==21993==
==21993== HEAP SUMMARY:
==21993== in use at exit: 89,536 bytes in 186 blocks
==21993== total heap usage: 247 allocs, 61 frees, 127,352 bytes allocated
==21993==
==21993== LEAK SUMMARY:
==21993== definitely lost: 0 bytes in 0 blocks
==21993== indirectly lost: 0 bytes in 0 blocks
==21993== possibly lost: 0 bytes in 0 blocks
==21993== still reachable: 89,536 bytes in 186 blocks
==21993== suppressed: 0 bytes in 0 blocks
==21993== Rerun with --leak-check=full to see details of leaked memory
==21993==
==21993== For counts of detected and suppressed errors, rerun with: -v
==21993== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)
Segmentation fault (core dumped)

[joaocandre]

I've managed to install gcc-7 and attempted to compile, but the issue remains.

[beniz]

Hi, I've just ran a build on Ubuntu 16.04 and it runs fine. What exact test are you running that gives you a crash ?

[joaocandre]

@beniz The sample code on the front page (fsphere()). I straight used the same code, compiled with g ++ 'pkg-config --cflags eigen3 libcmaes' sample_code.cpp 'pkg-config --libs libcmaes'. I've come across with reports on Eigen issues with dynamic memory on StackOverflow which I suspect may be causing this.

[beniz]

Can you try the examples in tests, typically test_functions ?

[joaocandre]

Calling ./test_functions --dim 30 --lambda 100 --max_iter 120 --fname fsphere I get Illegal instruction (core dumped)

[beniz]

And calling test_functions alone ?

[joaocandre]

The same happens.

[beniz]

Works for me. Do an ldd on you test_functions to report the linked libs, though this may not report eigen stuff. I'll look at it later on more in depth. If your version on arch works you might be able to compare the ldd output and get some information there.

[beniz]

Also report your full build log in a gist

[beniz]

Tried from scratch this morning and it works for me on a 16.04lts.

[joaocandre]

Linked libraries on both systems look to be the same. Difference is solely that on Arch the library was installed/compiled from the AUR (though I assume it relies on the same compilation scripts). I'm compiling with make following the instructions, everything seems to compile ok.

Compilation log: https://gist.github.com/joaocandre/742365dcb18735b24c81eaecf2f93638

[joaocandre]

Not sure if relevant, but the libcmaes.so.0.0.0 file in Arch is 6.3 Mb, while on Ubuntu 16.04 93Mb. That's a staggering difference.

[joaocandre]

I've found that everything works when I compile with CMake, generating a 6Mb shared object. I guess there is something wrong with the Makefile!

[beniz]

See whether #183 fixes your issue as well.