Subject: [Boost-bugs] [Boost C++ Libraries] #10799: Several problems building the example/tutorial project
From: Boost C++ Libraries (noreply_at_[hidden])
Date: 2014-11-17 22:21:29
#10799: Several problems building the example/tutorial project
------------------------------+-------------------------
Reporter: m@… | Owner: rwgk
Type: Bugs | Status: new
Milestone: To Be Determined | Component: Python
Version: Boost 1.57.0 | Severity: Showstopper
Keywords: |
------------------------------+-------------------------
-- There are a number of problems reported here, so you may want to break
this down into pieces for separate tracking. However, I would consider
this item to be fixed only when I can build the example/tutorial and it
works.
I had a hard time building the example/tutorial project. Some of this is,
I believe, due to bugs in the distribution, and some is probably my lack
of understanding.
Some of this I posted to the Boost interest group before I could join this
(Python C++) SIG, hoping it would get forwarded on here, so there may be a
duplication.
1. Calling bjam in the example/tutorial project failed to even start up
the build system. This was because the example/bootstrap.jam pointed to
the wrong path for the build system root. Instead of
../../../tools/build/v2, it should be ../../../tools/build/src/kernel.
When I changed this, bjam now got past the build system startup.
2. Building the project not only compiles hello.cpp, but it also builds a
private copy of the Boost/Python library. So I needed to supply the
properties needed to correctly build this library (i.e., link=shared,
address-mode=64). Of course, I needed to supply those same properties
anyway as part of creating the extension.
There's probably a way to change something so that the extension uses the
library built in the Boost.Python's own project, or if I have obtained the
libraries without having to build them, it would use these. I don't know
if you intended to have the tutorial example make its own copies, but it
seems a waste of resources to do so.
3. The link for debug mode failed, saying that the .pdb file was in the
wrong format (LNK1207). This is a bug, due to an option in the link
command '/IMPORTLIB:...hello_ext.pdb'. So the linker is creating the .pdb
file as an import library, then complaining that it's not a valid pdb
file. I changed '.pdb' to '.lib'. I could also have removed this option
entirely, since hello_ext.pyd doesn't have anything to export anyway.
4. Before figuring out that the link was the problem, I changed the /Z7
argument to /Zi in the compile command for hello.cpp. I don't know if
this was necessary, or if it was necessary to leave it in place. For now,
I just wanted to get it to build. Without /Z7, the debug symbols go into
example/tutorial/vc120.pdb. I don't know if the linker found these or
not. When I try stepping into the extension, I'll know for sure.
Microsoft prefers that .pdb files be used for debug symbols rather than
embedding them in the .obj files, so this might be the only real reason to
make the change.
5. The link for both release mode failed with two undefined symbols,
__imp_DecodePointer and __imp_EncodePointer, which are in kernel32.lib. I
tried adding kernel32.lib to the link's inputs. But then it warned that
there could be static constructors/destructors not getting called. After
much research on this topic, I found that the source of the problem was
the /NOENTRY argument in the link command, which has the consequence that
the automatic CRT initialization of the DLL doesn't occur. So I remove
the /NOENTRY and got not warnings, and I didn't need to add kernerl32.lib
either.
6. A minor point. The MACHINE:X64 is redundant. The linker knows it's
X64 because of the architecture of the input files and libraries. Nothing
wrong with it being there, but it's just clutter in the code.
7. Now bjam was successful in building whatever it wanted to build. It
said that hello.test passed. Sounds great, I thought. But I then went
into Python and tried 'import hello_ext' and that failed. So I have an
issue with the test program passing, while the extension didn't actually
work.
8. The problem was that bjam didn't put the hello_ext.pyd file in my
Python's lib/site-packages folder. It built the .pyd and .pdb files in
the example/tutorial/bin/... staging area, and copied (ONLY) the .pyd file
to example/tutorial. So not only did the .pyd file get put in the wrong
place, but the .pdb was left out. If I am going to debug my extension
(such as with Visual Studio's Python Tools), the .pdb file also needs to
be in the lib/site-packages folder. Without the .pdb file, the PTVS
debugger will not set breakpoints or step into the extension code.
9. I spent another few hours figuring out why the import statement in
Python didn't work. Python could not load the hello_ext.pyd because it
has references to BOOST_PYTHON_....DLL. If I run Python from the
example/tutorial directory itself, it works because this DLL had been
build there. That explains why the hello.test passed (as bjam ran it from
that same directory). So to fix this problem, I copied the DLL (and its
accompanying PDB) from the place where it was built into a directory in my
PATH. It would also work to have these files placed in the Python lib
/site-packages folder, since the main DLL's folder is one of the places
that Windows looks for imported DLLs.
To summarize, the fixes I made in order to build the example and use it in
Python were:
* Change the build system path in examples/bootstrap.jam.
* Change the /IMPORTLIB:....pdb to ....lib in the link command. I
could have removed it altogether.
* Remove /NOENTRY from the link command.
* Manually copy the example/tutorial/bin/.../hello_ext.(pyd,pdb) files
to Python's lib/site-packages.
* Manually copy the Boost.Python DLL and PDB files to a location in my
%PATH% (or to lib/site-packages)
These last two steps are not necessary if I run Python from the
example/tutorial directory itself, as Python will find the PYD there and
it will find the Boost.Python DLL there as well. So if the other problems
are fixed, then the tutorial would be telling the truth that the extension
can be imported and works correctly PROVIDED that you don't change
directories after doing the build.
Maybe there's something later in the tutorial about modifying the jamfile
to specify where you want the files installed so that Python can import
them no matter what directory Python was called from. If this is the
case, that's OK with me, as long as the tutorial says '''clearly''' where
you first build the example that at this point, Python '''must be run from
the example/tutorial directory'''.
This cost me about a whole day's time, and I'm pretty resourceful. And I
actually want to build my own extensions in VS instead of bjam, linking to
the Boost.Python DLL. I started of by trying this, and my extension
crashed. So that's why I went about building the example with bjam, to at
least see what a successful deployment looked like. I now understand that
my original attempt crashed because linking to the Boost.Python DLL is not
enough to be sure that the target PYD will load.
If I had understood that in the first place, I would have taken care to
put the Boost.Python DLL on my PATH, and not tried building the example in
the manner specified in the tutorial. But then I would not have
discovered those build bugs, and they would have gone unreported.
I suppose that a lot of people would have given up building the tutorial
example and then given up on Boost.Python altogether.
One more comment. You should make sure, as part of the QA process for
releasing Boost that you be sure that all the examples can be built and
work properly, particularly on Windows with all the MSVC toolsets. Item
(1) would apply to all platforms, so I would guess that the
example/tutorial was not tested at all.
Michael Rolle
-- Ticket URL: <https://svn.boost.org/trac/boost/ticket/10799> Boost C++ Libraries <http://www.boost.org/> Boost provides free peer-reviewed portable C++ source libraries.
This archive was generated by hypermail 2.1.7 : 2017-02-16 18:50:17 UTC