Arch Linux

Description:
It's not possible to use libtensorflow_cc with C++17.
I'm not sure what is the best way out of it, but probably it worth considering just building it with C++17? (but that will break linkage with pre-C++17 C++ versions)
Another option would be patching absl/base/config.h as described below.

The reason for it that there are ABSL headers bundled with tensorflow, and tensorflow uses absl::string_view in their API.
absl::string_view:
- is a typedef to std::string_view if it detects C++17,
- has a custom implementation otherwise.

libtensorflow_cc.so is built with C++11 so it uses the custom implementation
However when those headers are read with C++17, functions start to pass std::string_view as parameter instead.

To work around it, it's possible to comment out the line:
#define ABSL_HAVE_STD_STRING_VIEW 1
in /usr/include/tensorflow/absl/base/config.h

Additional info:
* package version: 2.1.0-4 (it has always been like that though)

Steps to reproduce:

1. Type this code and save to test.cc:
=== cut ===
#include <tensorflow/cc/ops/standard_ops.h>

using namespace tensorflow;
using namespace tensorflow::ops;

int main() {
auto scope = Scope::NewRootScope();
auto input = Const(scope, Tensor(DataType::DT_INT32, {}));
auto filter = Const(scope, Tensor(DataType::DT_INT32, {}));
Conv2D(scope, input, filter, {1, 1, 1, 1}, "SAME");
}
=== cut ===
2. Compile it this way (c++11), it compiles fine:
$ g++ test.cc -I/usr/include/tensorflow /usr/lib/libtensorflow_cc.so /usr/lib/libtensorflow_framework.so

3. Compile it with C++17, get a linker error:
$ g++ test.cc -I/usr/include/tensorflow /usr/lib/libtensorflow_cc.so /usr/lib/libtensorflow_framework.so --std=c++17
/usr/bin/ld: /tmp/ccTk1GBe.o: in function `main':
test.cc:(.text+0x2b6): undefined reference to `tensorflow::ops::Conv2D::Conv2D(tensorflow::Scope const&, tensorflow::Input, tensorflow::Input, absl::Span<int const> const&, std::basic_string_view<char, std::char_traits<char> >)'
collect2: error: ld returned 1 exit status