Binary Extension¶
Note
This content was last updated May 19, 2020. Much of the content is tested automatically to keep from getting stale, but some of the console code blocks are not. As a result, this material may be out of date. If anything does not seem correct — or even if the explanation is insufficient — please file an issue.
The bezier
Python package has optional speedups that wrap the
libbezier
library. These are incorporated into the Python interface via
Cython as a binary extension. See C ABI (libbezier) for more information
on building and installing libbezier
.
Extra (Binary) Dependencies¶
When the bezier
Python package is installed via pip, it will likely be
installed from a Python wheel. The wheels uploaded to PyPI are pre-built,
with the Fortran code compiled by GNU Fortran (gfortran
). As a
result, libbezier
will depend on libgfortran
. This can be problematic
due to version conflicts, ABI incompatibility, a desire to use a different
Fortran compiler (e.g. Intel’s ifort
) and a host of other reasons.
Some of the standard tooling for distributing wheels tries to address this. On
Linux and macOS, the tools address it by placing a copy of libgfortran
(and
potentially its dependencies) in the built wheel. (On Windows, there is no
standard tooling beyond that provided by distutils
and setuptools
.)
This means that libraries that depend on libbezier
may also need to link
against these local copies of dependencies.
Linux¶
The command line tool auditwheel adds a bezier.libs
directory to
site-packages
(i.e. it is next to bezier
) with a modified
libbezier
and all of its dependencies (e.g. libgfortran
)
>>> libs_directory
'.../site-packages/bezier.libs'
>>> print_tree(libs_directory)
bezier.libs/
libbezier-28a97ca3.so.2020.5.19
libgfortran-2e0d59d6.so.5.0.0
libquadmath-2d0c479f.so.0.0.0
libz-eb09ad1d.so.1.2.3
The bezier._speedup
module depends on this local copy of libbezier
:
$ readelf -d _speedup.cpython-38-x86_64-linux-gnu.so
Dynamic section at offset 0x447000 contains 27 entries:
Tag Type Name/Value
0x000000000000000f (RPATH) Library rpath: [$ORIGIN/../bezier.libs]
0x0000000000000001 (NEEDED) Shared library: [libbezier-28a97ca3.so.2020.5.19]
0x0000000000000001 (NEEDED) Shared library: [libpthread.so.0]
0x0000000000000001 (NEEDED) Shared library: [libc.so.6]
0x000000000000000c (INIT) 0x9d50
...
and the local copy of libbezier
depends on the other dependencies in
bezier.libs/
(both directly and indirectly):
$ readelf -d ../bezier.libs/libbezier-28a97ca3.so.2020.5.19
Dynamic section at offset 0x44dd8 contains 28 entries:
Tag Type Name/Value
0x0000000000000001 (NEEDED) Shared library: [libgfortran-2e0d59d6.so.5.0.0]
0x0000000000000001 (NEEDED) Shared library: [libm.so.6]
0x0000000000000001 (NEEDED) Shared library: [libgcc_s.so.1]
0x0000000000000001 (NEEDED) Shared library: [libc.so.6]
0x000000000000000e (SONAME) Library soname: [libbezier-28a97ca3.so.2020.5.19]
0x000000000000000c (INIT) 0x2be8
...
$ readelf -d ../bezier.libs/libgfortran-2e0d59d6.so.5.0.0
Dynamic section at offset 0x207db8 contains 31 entries:
Tag Type Name/Value
0x0000000000000001 (NEEDED) Shared library: [libquadmath-2d0c479f.so.0.0.0]
0x0000000000000001 (NEEDED) Shared library: [libz-eb09ad1d.so.1.2.3]
0x0000000000000001 (NEEDED) Shared library: [libm.so.6]
0x0000000000000001 (NEEDED) Shared library: [libgcc_s.so.1]
0x0000000000000001 (NEEDED) Shared library: [libc.so.6]
0x000000000000000e (SONAME) Library soname: [libgfortran-2e0d59d6.so.5.0.0]
0x000000000000000c (INIT) 0x19a78
...
Note
The runtime path (RPATH
) uses $ORIGIN
to specify a path
relative to the directory where the extension module (.so
file) is.
macOS¶
The command line tool delocate adds a bezier/.dylibs
directory
with copies of libbezier
, libgfortran
, libquadmath
and
libgcc_s
:
>>> dylibs_directory
'.../site-packages/bezier/.dylibs'
>>> print_tree(dylibs_directory)
.dylibs/
libbezier.2020.5.19.dylib
libgcc_s.1.dylib
libgfortran.5.dylib
libquadmath.0.dylib
The bezier._speedup
module depends on the local copy
of libbezier
:
$ otool -L _speedup.cpython-38-darwin.so
_speedup.cpython-38-darwin.so:
@loader_path/.dylibs/libbezier.2020.5.19.dylib (...)
/usr/lib/libSystem.B.dylib (...)
Though the Python extension module (.so
file) only depends on libbezier
it indirectly depends on libgfortran
, libquadmath
and libgcc_s
:
$ otool -L .dylibs/libbezier.2020.5.19.dylib
.dylibs/libbezier.2020.5.19.dylib:
/DLC/bezier/libbezier.2020.5.19.dylib (...)
@loader_path/libgfortran.5.dylib (...)
/usr/lib/libSystem.B.dylib (...)
@loader_path/libgcc_s.1.dylib (...)
@loader_path/libquadmath.0.dylib (...)
Note
To allow the package to be relocatable, the libbezier
dependency is
relative to the @loader_path
(i.e. the path where the Python extension
module is loaded) instead of being an absolute path within the file
system.
Notice also that delocate
uses the nonexistent root /DLC
for
the install_name
of libbezier
to avoid accidentally pointing
to an existing file on the target system.
Windows¶
A single Windows shared library (DLL) is provided: bezier.dll
.
The Python extension module (.pyd
file) depends directly on this library:
> dumpbin /dependents _speedup.cp38-win_amd64.pyd
Microsoft (R) COFF/PE Dumper Version ...
Copyright (C) Microsoft Corporation. All rights reserved.
Dump of file _speedup.cp38-win_amd64.pyd
File Type: DLL
Image has the following dependencies:
bezier-e5dbb97a.dll
python38.dll
KERNEL32.dll
VCRUNTIME140.dll
api-ms-win-crt-stdio-l1-1-0.dll
api-ms-win-crt-heap-l1-1-0.dll
api-ms-win-crt-runtime-l1-1-0.dll
...
For built wheels, the dependency will be renamed from bezier.dll
to a
unique name containing the first 8 characters of the SHA256 hash of the DLL
file (to avoid a name collision) and placed in a directory within the
bezier
package: for example extra-dll/bezier-e5dbb97a.dll
.
In order to ensure this DLL can be found, the bezier.__config__
module adds the extra-dll
directory to the DLL search path on import.
(%PATH%
is used on Windows as part of the DLL search path. For Python
versions starting with 3.8, modifying os.environ["PATH"]
after Python
startup no longer works; instead the os.add_dll_directory()
function
achieves the same goal in a more official capacity.)
The libbezier
DLL has no external dependencies, but does have
a corresponding import library — usr/lib/bezier.lib
— which is
provided to specify the symbols in the DLL.
On Windows, building Python extensions is a bit more constrained. Each
official Python is built with a particular version of MSVC and
Python extension modules must be built with the same compiler. This
is primarily because the C runtime (provided by Microsoft) changes from
Python version to Python version. To see why the same C runtime must be used,
consider the following example. If an extension uses malloc
from
MSVCRT.dll
to allocate memory for an object and the Python interpreter
tries to free that memory with free
from MSVCR90.dll
, bad things
can happen:
Python’s linked CRT, which is
msvcr90.dll
for Python 2.7,msvcr100.dll
for Python 3.4, and severalapi-ms-win-crt
DLLs (forwarded toucrtbase.dll
) for Python 3.5 … Additionally each CRT uses its own heap for malloc and free (wrapping WindowsHeapAlloc
andHeapFree
), so allocating memory with one and freeing with another is an error.
This problem has been largely fixed in newer versions of Python but is still worth knowing.
Unfortunately, there is no Fortran compiler provided by MSVC. The
MinGW-w64 suite of tools is a port of the GNU Compiler Collection (gcc
)
for Windows. In particular, MinGW includes gfortran
. However, mixing the
two compiler families (MSVC and MinGW) can be problematic because MinGW uses
a fixed version of the C runtime (MSVCRT.dll
) and this dependency cannot
be easily dropped or changed.
A Windows shared library (DLL) can be created after compiling each of the Fortran submodules:
$ gfortran \
> -shared \
> -o bezier.dll \
> ${OBJ_FILES} \
> -Wl,--output-def,bezier.def
Note
Invoking gfortran
can be done from the Windows command prompt (e.g.
it works just fine on AppVeyor), but it is easier to do from a shell that
explicitly supports MinGW, such as MSYS2.
By default, the created shared library will depend on gcc
libraries
provided by MinGW:
> dumpbin /dependents ...\bezier.dll
...
Image has the following dependencies:
KERNEL32.dll
msvcrt.dll
libgcc_s_seh-1.dll
libgfortran-3.dll
Unlike Linux and macOS, on Windows relocating and copying any dependencies
on MinGW (at either compile, link or run time) is explicitly avoided. By adding
the -static
flag
$ gfortran \
> -static \
> -shared \
> -o bezier.dll \
> ${OBJ_FILES} \
> -Wl,--output-def,bezier.def
all the symbols used from libgfortran
or libgcc_s
are statically
included and the resulting shared library bezier.dll
has no dependency
on MinGW:
> dumpbin /dependents extra-dll\bezier-e5dbb97a.dll
Microsoft (R) COFF/PE Dumper Version ...
Copyright (C) Microsoft Corporation. All rights reserved.
Dump of file extra-dll\bezier-e5dbb97a.dll
File Type: DLL
Image has the following dependencies:
KERNEL32.dll
msvcrt.dll
USER32.dll
...
Note
Although msvcrt.dll
is a dependency of bezier.dll
, it is not
a problem. Any values returned from Fortran (as intent(out)
) will
have already been allocated by the caller (e.g. the Python process).
This won’t necessarily be true for generic Fortran subroutines, but
subroutines marked with bind(c)
(i.e. marked as part of the C ABI
of libbezier
) will not be allowed to use allocatable
or
deferred-shape output variables. Any memory allocated in Fortran will be
isolated within the Fortran code.
However, the dependency on msvcrt.dll
can still be avoided if desired.
The MinGW gfortran
default “specs file” can be captured:
$ gfortran -dumpspecs > ${SPECS_FILENAME}
and modified to replace instances of -lmsvcrt
with a substitute, e.g.
-lmsvcr90
. Then gfortran
can be invoked with the flag
-specs=${SPECS_FILENAME}
to use the custom spec. (Some
other dependencies may also indirectly depend on msvcrt.dll
,
such as -lmoldname
. Removing dependencies is not an easy process.)
From there, an import library must be created
> lib /def:.\bezier.def /out:.\lib\bezier.lib /machine:${ARCH}
Note
lib.exe
is used from the same version of MSVC that compiled the
target Python. Luckily distutils
enables this without difficulty.
Source¶
For code that depends on libgfortran
, it may be problematic to also
depend on the local copy distributed with the bezier
wheels.
The bezier
Python package can be built from source if it is not feasible to
link with these libraries, if a different Fortran compiler is required or
“just because”.
The Python extension module can be built from source via:
$ # One of
$ BEZIER_INSTALL_PREFIX=.../usr/ python -m pip wheel .
$ BEZIER_INSTALL_PREFIX=.../usr/ python -m pip install .
$ BEZIER_INSTALL_PREFIX=.../usr/ python setup.py build_ext
$ BEZIER_INSTALL_PREFIX=.../usr/ python setup.py build_ext --inplace