[cmake-developers] INTERFACE_LINK_LIBRARIES property?

[Stephen Kelly]

I've force pushed to my clone again. I haven't added tests yet, and I've 
only been smoke testing as I go along.

Brad King wrote:
> In the tll() signature policy I think LINK_PUBLIC/LINK_PRIVATE should
> be an alias for PUBLIC/PRIVATE.  The "old" signatures are only the
> one without any keywords plus the one with LINK_INTERFACE_LIBRARIES.
> The "new" signatures are the (LINK_)?(PUBLIC|PRIVATE|INTERFACE) mode.
> That is why we need a policy to make the old and new sigs exclusive.

Currently in my patch, the LINK_PUBLIC and LINK_PRIVATE are considered old 
signatures. That's easy to document and keeps most relevant existing 
documentation together without changing it.

As a consequence of being part of the old signature, the use of LINK_PRIVATE 
determines whether the old LINK_INTERFACE_LIBRARIES is populated with an 
empty value. The use of the new signature with PRIVATE determines whether to 
populate the new INTERFACE_LINK_LIBRARIES property. It might be possible to 
control whether the INTERFACE_LINK_LIBRARIES is populated with an empty 
value determined by the policy alone (and make LINK_PRIVATE an alias for 
PRIVATE as you suggest), but for anyone transitioning (and using 
EXPORT_LINK_INTERFACE_LIBRARIES) it is more things to think about, it is 
more nuances to document in the policy and in the tll() documentation and it 
is more complex to implement.

Similarly, for static libraries, we want to populate the 
INTERFACE_LINK_LIBRARIES with different stuff (genexes) when using the new 
signature. Again, that could be controlled by the policy alone, but for the 
same complexity , documentation and nuance arguments above, I don't think it 
should be.

I think a clean break is a better idea.

> Instead we should wrap the "private" link-only dependencies in an
> expression because they are the ones that do not belong in the
> interface.  Perhaps use $<LINK_PRIVATE:...> or $<LINK_ONLY:...>
> where ... can even be a list to avoid verbosity for adjacent
> link-only entries.

I've implemented LINK_ONLY now. The implementation of the genex supports 
lists, but the implementation of target_link_libraries would need to be 
refactored to emit groups instead of single libraries in order to take 
advantage of it.

Thanks,

Steve.