1.0.0 crash with Google perf tools

[juliensimon]

Hi,

1.0.0 dumps core at loading time when built with Google perf tools.

I'm using a freshly booted Deep Learning AMI, Ubuntu Linux - 2.4_Oct2017 (ami-dca37ea5).

Python 2.7.12 (default, Nov 20 2017, 18:23:56)
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.
import mxnet
src/tcmalloc.cc:278] Attempt to free invalid pointer 0x1d9cc60
Aborted (core dumped)

Steps to reproduce:

• sudo apt-get install libgoogle-perftools-dev
• clone a fresh MXNet repo
• set USE_GPERFTOOLS = 1
• build and install

The same build without libgoogle-perftools-dev installed works fine.

[eric-haibin-lin]

@cjolivier01

[cjolivier01]

@eric-haibin-lin

[eric-haibin-lin]

@bhavinthaker @mbaijal should we also cover this in CI build tests?

[cjolivier01]

It may be that it's not gperftools that's the problem, but that there's a bug and the normal malloc library doesn't make any issue of trying to free what it sees as an invalid pointer

[bhavinthaker]

Hi Haibin: Yes, we should cover these tests in the CI. However I would request you or any team member to help contribute to the unit tests or create a backlog item (issue/Jira) so that any of the community members can work on it.

Hi Chris: Your theory could be valid. I remember Solaris being permissive but HPUX being strict for the same code in C. Such problems were typically found with strict checks using compiler flags/tools. Could you please investigate the problem and let us know your findings?

Do we have a unit-test in CI with Gperf enabled? If not, please add one. This should have been enforced by code-reviewers of the PR that enabled gperftools.

[bhavinthaker]

Let's try a coverity run too: @rahul003

[juliensimon]

Hello,

FWIW, I ran into this issue again a couple of days when building master.

[KellenSunderland]

Maybe we could build and test with gperf, clang, and a few sanitizers?

[rahul003]

Related issues
#9096
#9098