[mlpack] make test failure on OS X

[Kevin Locke]

On 03/24/2014 09:22 AM, Ryan Curtin wrote:
>>> On Wed, Mar 19, 2014 at 01:40:59PM -0600, Kevin Locke wrote:
>>>> $ ./bin/mlpack_test -t LoadSaveTest/SaveArmaASCIITest
>>>> Running 1 test case...
>>>> [WARN ] Unable to detect type of 'test_file.txt'; load failed. Incorrect extension?
>>>> /path/to/mlpack-1.0.8/src/mlpack/tests/load_save_test.cpp:202: fatal error in "SaveArmaASCIITest": critical check data::Load("test_file.txt", test) == true failed
>>>>
>>>> *** 1 failure detected in test suite "MLPACKTest"
>>>>
>>>>
>>>> So, is this an error that I should be concerned about?  It is
>>>> something related to my machine or build configuration?  Is there more
>>>> that I can do to help diagnose the issue?  Any input/help is greatly
>>>> appreciated.
>
> I'm sorry for the slow response -- I was out of town last week and had
> spotty internet access.  Li Dong, in another thread, suggested that the
> use of -std=c++11 causes the bus error issue, and I am curious if it
> also causes the test failure.  Could you try compiling without
> -std=c++11 and seeing if LoadSaveTest/SaveArmaASCIITest still fails?  If
> it still fails, can you attach the file "test_file.txt" which should be
> generated in the directory you run mlpack_test in?
>
> I am trying to get a hold of an OS X system (or get my VM running again)
> so that I can dig into this more thoroughly.

Not a problem at all.  Thanks for continuing to investigate.

I can confirm that removing -std=c++11 has no effect.  However,
recompiling without -std=c++11 and without -stdlib=libc++ (which
included recompiling boost and armadillo without -stdlib=libc++) was
successful in avoiding the error, which suggests that the issue is a
difference between libc++ and libstdc++.

For reference, I am on OS X 10.8 and my understanding is that in OS X
10.9 (Mavericks) clang++ links against libc++ by default.  So testing
differences may require -stdlib=libstdc++ on OS X 10.9 and later.

Kevin