On Nov 26, 2008, at 1:38 PM, Ben Webb wrote:
> Daniel Russel wrote: >> I agree that it is a bit controversial, but everything I have read >> prefers it to using LD_LIBRARY_PATH (since that, among other ill >> effects, changes the behavior of 3rd party libraries). > > I will look at the patch. But rpath will probably have to be > optional - > defaulting to true on OS X. The Modeller binary packages on OS X and > Linux are built with rpath for precisely the reason you mention - to > avoid the bundled copies of Fortran runtime libraries, HDF5, libpython > etc. from conflicting with system copies. But the Unix and Windows > builds do not use rpath. From what I understand, one shouldn't be using rpath on mac os and should be using install_name_tool and the linker to make sure the correct paths are saved in the library itself. Unfortunately, scons doesn't seem to set library ids automatically for some reason, so we will, at some point, have to patch installation ourselves.
>> > > Windows looks for dynamic libraries in the same directory as the > parent > binary (executable, or Python extension) and then falls back to > looking > in PATH. DLL and their paths can also be registered in some random > place > in the registry, IIRC. Maybe that was what I was thinking of. Anyway, the current setup doesn't seem to do much for most windows users anyway, so no great loss :-)