Build fails on Linux with GCC 9, 10, and 12 about tensorstore

[laramiel]

I think that our current build machine uses gcc 12.2.0, as does my workspace (gcc (Debian 12.2.0-3) 12.2.0).
Can you run the following in each build environment and tell me what it says:

./bazelisk.py build tensorstore/...
`./bazelisk.py info output_base`/external/local_config_cc/cc_wrapper.sh --version

These errors were on pip install tensorstore? What's your os, other environment like, because we should have prebuilt packages from github CI.

[wookayin]

The cc_wrapper.sh --version prints the same one as gcc --version on the $PATH.

(on the master branch)

GCC 10.3 (The system g++ installed on this Linux machine):

❯❯❯ which -a gcc g++
gcc: aliased to nocorrect gcc
/bin/gcc
/usr/bin/gcc
/bin/g++
/usr/bin/g++

❯❯❯ $(./bazelisk.py info output_base)/external/local_config_cc/cc_wrapper.sh --version
gcc (Ubuntu 10.3.0-1ubuntu1~20.04) 10.3.0
Copyright (C) 2020 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

GCC 12.2 (conda gcc/g++)

❯❯❯ which g++
$HOME/.miniforge3/bin/g++         # g++ (conda-forge gcc 12.2.0-19) 12.2.0

❯❯❯ $(./bazelisk.py info output_base)/external/local_config_cc/cc_wrapper.sh --version 

x86_64-conda-linux-gnu-cc (conda-forge gcc 12.2.0-19) 12.2.0
Copyright (C) 2022 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

With those bazel commands I somehow run into a different error:

ERROR: .../tensorstore/tensorstore/serialization/BUILD:64:20: Linking tensorstore/serialization/function_test failed: (Exit 1): x86_64-conda-linux-gnu-cc failed: error executing command $CONDA_PREFIX/bin/x86_64-conda-linux-gnu-cc @bazel-out/k8-fastbuild/bin/tensorstore/serialization/function_test-2.params

Use --sandbox_debug to see verbose messages from the sandbox
bazel-out/k8-fastbuild/bin/_solib_k8/libexternal_Scom_Ugoogle_Uabsl_Sabsl_Stime_Slibtime.so: error: undefined reference to 'clock_gettime'

This particular error has something to do with glibc 2.17+ not requiring -lrt flag.

One note is that the error messages in the OP was during installation via pip install . (which invokes bazel like python -u bazelisk.py build -c opt //python/tensorstore:_tensorstore__shared_objects --verbose_failures --copt=-fvisibility=hidden). In this case the same gcc, g++ on the $PATH was used.

[jbms]

The gcc 9 failures are known and expected --- we just need to change the documentation.

The other issues ideally can be solved. Which Python version?

[wookayin]

I see, documentation could be updated. Having GCC 12 as a standard, recommended version sounds good, but due to GCC's incompatibility with CUDA (https://stackoverflow.com/questions/6622454/cuda-incompatible-with-my-gcc-version) I hope that at least GCC 10.x could be supported. I think conda-shipped gcc 12.2 should work, but the linux system I'm particularly using has a bit old libstdc++ and glibc.

Python -- fails both on Python 3.10 and 3.11. I was actually trying to build tensorstore for python 3.11 because there is no prebuilt wheel available (it'd be also great if an official py311 support can be added!), but build failure is also there on python 3.10 as well with the same environment. In this issue, I more look to C++ build commands and/or documentations could be improved so bazel build can work with different environment configurations.

[laramiel]

The clock_getime failure is from Abseil; maybe we should raise a bug there. I don't see any -lrt conditional flags in the build, and I haven't looked into which gcc / glibc version require it.

I should try and setup an image to repro this. uname -a? You're running ubuntu and installed gcc how?

[wookayin]

Yes, my env is Ubuntu 20.04 LTS (focal), which ships with GCC 10. (Click to expand)

$ uname -a
Linux HOSTNAME 5.4.0-131-generic #147-Ubuntu SMP Fri Oct 14 17:07:22 UTC 2022 x86_64 x86_64 x86_64 GNU/Linux

$ cat /etc/lsb-release                                                                                                                  
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=20.04
DISTRIB_CODENAME=focal
DISTRIB_DESCRIPTION="Ubuntu 20.04.5 LTS"

$ sudo apt install gcc-10 g++-10
# (Optional)
$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 10
$ sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 10

[sameeul]

Just wanted to report that, with GCC 10.4 + and GCC 11, the build process (both python setup.py develop and building executables from the example fail due to failure to build boringssl. I can build cleanly with GCC 10.2 and GCC 10.3. I know this is not exactly a tensorstore issue rather an issue coming from one of its dependencies. I was just curious if any of you came across this except the author of issue (The GCC 12.X log).

[jbms]

@sameeul Are you also using gcc from conda?

It appears that it lacks proper C11 support, which is required by boringssl. For example, after installing conda-forge:

echo -e "#include <assert.h>\nstatic_assert(1);" > test.c
gcc -std=c11 -c test.cc

Fails with

t.c:2:15: error: expected declaration specifiers or '...' before numeric constant
    2 | static_assert(1);
      |               ^

In particular, if you look at mambaforge/x86_64-conda-linux-gnu/sysroot/usr/include/assert.h you will see that it is a very old version that lacks a #define static_assert line.

[jbms]

I filed upstream issue: conda-forge/linux-sysroot-feedstock#44

[sameeul]

@jbms : yes, I was using conda gcc in cases where my build was failing. The cases where it was working were compilers that came directly from debian-bullseye repo. Thanks for looking into it.