For a source:
|cdef extern from "<foo>":
|    cdef void foo_cpp() nogil except +
|
|cdef call_foo() nogil:
|    foo_cpp()

We used to generate something like this to actually call foo_cpp() in call_foo():
|try{foo_cpp();}
|catch(...) {
|  Py_BLOCK_THREADS; __Pyx_CppExn2PyErr(); Py_UNBLOCK_THREADS
|  `code.error_goto(self.pos))`
|}

where Py_BLOCK_THREADS expands to "PyEval_RestoreThread(_save);".
__Pyx_CppExn2PyErr() (and alternatives, see SimpleCallNode) calls CPython A API
methods so it needs to be guarded in a nogil environment.

One problem is that "PyThreadState *_save" is only declared by "with nogil:"
block, so a .cpp file that doesn't compile is generated for the above code.

However, I think the real issue is that Py_(UN)BLOCK_THREADS is inappropriate
here, as it isn't allowed to be called recursively and is valid only directly
in a Py_BEGIN_ALLOW_THREADS ... Py_END_ALLOW_THREADS. IMO PyGILState_Ensure()
and PyGILState_Release() (through `code.put_ensure_gil()` and a friend) is the
correct thing to call here as it is allowed to be called recursively and
actually ensures the current thread can call CPython C API.

This patch does exactly this (and it breaks the generated code to multiple
lines as it would be way too long otherwise), plus it extends the
cpp_exceptions_nogil.pyx test with above example that doesn't compile with this
fix not applied.

Note that we explicitly pass declare_gilstate=True to put_ensure_gil(), as
PyGILState_Ensure() docs state that recursive calls to it must not share the
PyGILState_STATE. C++-style declaring a variable inside a block should be
no-problem here, as try{} .. catch() is obviously valid only in a C++ code.

--HG--
extra : transplant_source : %AA%F3%BDk%FE%F9%01%7F%8C%A4n%5E%DA4%97%A5%D9%AF%D60

--HG--
extra : transplant_source : %8A%F1%17%5C%A7%CA%1C%B1%17%A1%C5t%06O%5BI%F2%19Vq

A code like
|cdef double[:] foo():
|    raise AttributeError('dummy')

generated C code like this:
|  __pyx_L1_error:;
|  (...)
|  __pyx_r.memview = NULL;
|  __Pyx_AddTraceback("view_return_errors.foo", __pyx_clineno, __pyx_lineno, __pyx_filename);
|  __pyx_L0:;
|  if (unlikely(!__pyx_r.memview)) {
|    PyErr_SetString(PyExc_TypeError,"Memoryview return value is not initialized");
|  }
|  __Pyx_RefNannyFinishContext();
|  return __pyx_r;
|}

Which is incorrect in error case, because we set __pyx_r.memview to NULL and
then we test it, so that the PyErr_SetString() is always called in the error
case, which swallows previously-set error. (it also swallowed the traceback,
which I don't understand)

A fix is to jump to return_from_error_cleanup label in case return type is
memoryview, as it is currently done for the case when buffers are present.

A testcase that fauils w/out this fix applied is included, too.

v2: fix test under Python 3 by not using StandardError

--HG--
extra : transplant_source : G%B5%99Og%D1%81%25k%8F%1F%7B%02V%3E%B9%A4y%FF%EA

--HG--
extra : transplant_source : %7B%A4%F1%C4/%E4l%2C_%BFF%5B%B7%9C%9F%E0_%2B%15%3D

move exception class into shadow function as we may not want to export it under the cython.* namespace at this point

removed broken minivect related code from branch since minivect will not be released as part of 0.18

If '' is in sys.path and a module is found the package_path
is relative and breaks the build process.

--HG--
extra : transplant_source : 9%EA%CC%A6%3D%1B9R%EF%0DmM%CFZ%18%F3%EC%06%3B%B7

Use OS-dependent directory separator - a / on windows is interpreted by LINK as a command line switch

--HG--
extra : transplant_source : %A8%F23%AF%26%BC%82y1%86S%1Ac%D3%40%089o%DCQ

--HG--
extra : transplant_source : %E6%CA%F8%11%E8%81u%B9%95%3D%27%C1%0F%F3O%8A%12%3Cnl

--HG--
extra : transplant_source : U%15%0B%8E%81%02%F2kE%AA%07u%EF%82%3D14%F1C%86

--HG--
extra : transplant_source : H%91%CF%08t%B1%908%AE%26%81%1B%F9%2C%9A%3Fh%ECWK

BF: replace show-ref with rev-parse --verify for determining current treeish location