[cmake-developers] [patch] Export template instantiations with GenerateExportHeader

[Roger Leigh]

On 30/07/2015 18:34, Brad King wrote:
> On 07/30/2015 11:23 AM, Roger Leigh wrote:
>> foo_EXPORT_TEMPLATE template class foo_EXPORT std::allocator<char>;
> [snip]
>>   #    ifdef @EXPORT_IMPORT_CONDITION@
>>           /* We are building this library */
>>   #      define @EXPORT_MACRO_NAME@ @DEFINE_EXPORT@
>> +#      define @EXPORT_MACRO_NAME at _TEMPLATE
>>   #    else
>>           /* We are using this library */
>>   #      define @EXPORT_MACRO_NAME@ @DEFINE_IMPORT@
>> +#      define @EXPORT_MACRO_NAME at _TEMPLATE extern
>>   #    endif
> [snip]
>> Note that I'm by no means a Windows C++ or DLL expert--I'm primarily a
>> C++ UNIX developer--this is just from reading the docs and experimenting.
>
> The problem with this approach is that the non-extern explicit
> template instantiation should not be done in the header file
> because if more than one source file in a single library includes
> the header then each of those object files will provide the
> explicit instantiation.  Only one source should provide the symbols.
>
> So, one actually wants the "extern" to appear in the header
> file in all cases except when included in *one* specific
> source file within the library.  For this there needs to
> be an identifying macro associated with the source file.
> Therefore it does not make sense for this macro to be defined
> as part of GenerateExportHeader because that generates macros
> that switch off of library-level conditions.

This is where my knowledge of Windows linking falls short.

Are the duplicated template exports here at the level of the translation 
unit not elided when linking the DLL?  In the dlltest git repo, I 
specifically create DLLs with duplicate template exports in different 
translation units to make sure it doesn't error out, and it appears to 
cope.  I haven't checked on whether it removed duplicates from the 
linked objects though--is this even possible?  I thought all current 
toolchains were able to eliminate duplicate instantiations, or else 
you'd have massive explosions of duplicate templates common across all 
translation units.  Given the backwardness of the Windows linker, it 
wouldn't surprise me if it were true though...


Regards,
Roger