From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2021-01-20 08:52:50
On 1/20/21 11:48 AM, Andrey Semashev wrote:
> On 1/20/21 6:23 AM, Stefan Seefeld via Boost wrote:
>> I'm unfortunately not familiar enough with the platform (Windows) or
>> this compiler (MSVC) to help, but I would like to add some notes in
>> case someone else on this list has ideas:
>> Compiling a trivial C++ applet based on code such as
>> on Linux simply requires a command such as
>> g++ -I /usr/include/python3.8 -o np_test np_test.cpp -lboost_numpy38
>> -lboost_python38 -lpython3.8
>> I find particularly confusing that the MSVC linker would generate a
>> missing symbol error for the `ndarray::get_data()` function, given
>> that this function (as defined in
>> ought to be inlined, and not generate a symbol at all.
>> Does anyone familiar with this platform have an idea under what
>> circumstances this may result in a link error ?
> The class is marked with BOOST_NUMPY_DECL, which, I assume, is
> __declspec(dllexport)/(dllimport) on Windows. This means that when
> compiling the user's code the compiler may look for the imported
> function from the dll when inlining does not happen, which is usual in
> debug builds, for example.
> The fact that there is no such function probably means that the class is
> never included when Boost.Python is compiled (i.e. when the class is
> marked as dllexport).
Actually, no, it must be indirectly included in src/numpy/ndarray.cpp.
But that file is a part of boost_numpy library, not boost_python. Does
it get exported from boost_numpy?
> Marking the whole class as dllexport/dllimport has other consequences,
> such as exporting type information for the class. If this is not needed,
> I would suggest marking individual members to export/import and not the
> whole class.
>> (See the exact command-line in the previous post).
> What previous post?
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk