Templated instances defined by wrap-instances are not resolved by PXDParser

[poshul]

Defining a templated class with instances specified by wrap-instances in a PXD file does not make the instanced type available for template resolution in another PXD file.

Example: in file a.pxd:

    cdef cppclass IteratorWrapper[Iterator]:
      # wrap-instances:
      #  ScoreTypeRef := IteratorWrapper[libcpp_set[ ScoreType ].iterator]
...

in file b.pxd: libcpp_vector[ ScoreTypeRef ] assigned_scores results in: b.pxd:XX:YY: unknown type in template argument

Looking at the pyx code generated by a ScoreTypeRef is properly generated, in b however the templated type is just error.

[jpfeuffer]

Hi,

can you add test files, add the test to e.g. test_code_generator and open a PR?

[jpfeuffer]

Does it work with more basic instantiation? E.g. # ScoreTypeRef := IteratorWrapper[A]

[poshul]

It looks like it doesn't work with the more basic instantiation either. Looking at our existing codebase it looks like wrap-instances are only used locally, so it may be this issue has never come up.

[jpfeuffer]

ahh hmm that might also be the cause, yes. Can you check if it works locally in the same pxd file?

[jpfeuffer]

Is a cimport for ScoreTypeRef generated in the b.pxd file?

[poshul]

It appears to not work even in the same pxd file (though it's necessarily in a different cdef extern from block on account of coming from different cpp files)

[poshul]

Yup.

[jpfeuffer]

There is a similar test for this in autowrap already. There we just explicitly specify the instantiation again. https://github.com/OpenMS/autowrap/blob/04192b9ab6ed74d9d72c032587e222dbbfdb6a38/tests/test_files/templated.pxd#L45

[jpfeuffer]

Probably only "cppclass"es can be put as template parameter. And the instantiation is already a cython "class".

[poshul]

So at least in the triply nested case libcpp_vector[ IteratorWrapper [ libcpp_set [ ScoreType ] ] ] assigned_scores this fails.

[poshul]

Probably only "cppclass"es can be put as template parameter. And the instantiation is already a cython "class".

libcpp_vector[ IteratorWrapper [ libcpp_set [ String ] ] ] assigned_scores also fails, so I think its probably how deeply nested these are allowed.

[jpfeuffer]

yes I can confirm.

self = <autowrap.Types.CppType object at 0x113478700>
typemap = {'T': <autowrap.Types.CppType object at 0x113703520>, 'T2': <autowrap.Types.CppType object at 0x113703910>, 'Templated': <autowrap.Types.CppType object at 0x113703040>, 'Templated_other': <autowrap.Types.CppType object at 0x113703700>, ...}
indent = 0

    def _transform(self, typemap, indent):

        aliased_t = typemap.get(self.base_type)
        if aliased_t is not None:
            if self.template_args is not None:
                if aliased_t.template_args is not None:
                    map_ = printable(typemap, "\n    ")
                    m = "invalid transform of %s with:\n    %s" % (self, map_)
>                   raise Exception(m)
E                   Exception: invalid transform of Templated[libcpp_vector[T]] with:
E                       T -> _T
E                       T2 -> _T2
E                       Templated -> _Templated[_T]
E                       Templated_other -> _Templated[_T2]
E                       Y -> _Y
E                       Z -> _Z

autowrap/Types.py:92: Exception

[jpfeuffer]

What if you introduce (c)typedefs for now? Let me try

[poshul]

What if you introduce (c)typedefs for now? Let me try

That at least gives us a more detailed error message: Exception: Recursion in std::vector<T> is not implemented for other STL methods and wrapped template arguments