Apparently the %import directive is meant to be used with .i files
rather than .h files.
No, %import is meant to be used with SWIG interface files, just like
%include is. In many cases, regular header files are valid SWIG
interface files, and that's the case with IMP. By %import'ing all of
IMP, you're telling SWIG about a whole bunch of classes it doesn't need
to care about for impEM. So this change is unnecessary - unless it does
actually fix a problem you're having (in which case, add a testcase
which crashes with the current setup and works with your patch).
In addition, we are not properly sharing the swig runtime info between
modules. Apparently we need to add " -DSWIG_TYPE_TABLE=IMP" to each
compilation of a _wrap.cc so that the information is properly shared.
On the contrary, you only need to define that if you _don't_ want to
share the information except with other modules that have the same type
table. So we don't need this change either - by default everything is
shared (and since we use namespaces properly we shouldn't conflict with
other SWIG modules). See http://www.swig.org/Doc1.3/Modules.html#Modules_nn2Admittedly that documentation is a little misleading, so again, if you
need it to fix a problem, add a testcase.
I would also propose changing IMP_exceptions.i as follows in order to
allow other libraries to reuse the exception handling function by
%including IMP_exceptions.i rather than redefine a new one:
Sounds good to me. There's been no need to do that thus far because the
impEM guys seem to be allergic to exceptions. ;)