Use pkg-config autoconf macros to detect flags for Modules/Setup

[tiran]

BPO	45573
Nosy	@Yhg1s, @brettcannon, @tiran, @ned-deily, @erlend-aasland
PRs	bpo-45573: Use pkg-config for Modules/Setup #29164 bpo-45573: Detect stdlib extension modules in configure #29534 bpo-45573: Use Makefile's dependencies in setup.py (GH-29559) #29559 bpo-45573: Introduce extension module flags in Makefile (GH-29594) #29594 bpo-45573: check for ossaudiodev in configure (GH-29614) #29614 bpo-45573: Add Modules/Setup.stdlib with conditional modules (GH-29615) #29615 bpo-45573: Move mandatory core modules to Modules/Setup.bootstrap (GH-29616) #29616
Superseder	bpo-45774: Detect SQLite in configure.ac

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/tiran'
closed_at = None
created_at = <Date 2021-10-22.15:22:43.638>
labels = ['type-feature', 'build', '3.11']
title = 'Use pkg-config autoconf macros to detect flags for Modules/Setup'
updated_at = <Date 2021-11-19.15:41:03.253>
user = 'https://github.com/tiran'

bugs.python.org fields:

activity = <Date 2021-11-19.15:41:03.253>
actor = 'christian.heimes'
assignee = 'christian.heimes'
closed = False
closed_date = None
closer = None
components = ['Build']
creation = <Date 2021-10-22.15:22:43.638>
creator = 'christian.heimes'
dependencies = []
files = []
hgrepos = []
issue_num = 45573
keywords = ['patch']
message_count = 14.0
messages = ['404781', '404783', '404784', '406084', '406135', '406138', '406139', '406331', '406375', '406520', '406527', '406535', '406586', '406591']
nosy_count = 5.0
nosy_names = ['twouters', 'brett.cannon', 'christian.heimes', 'ned.deily', 'erlendaasland']
pr_nums = ['29164', '29534', '29559', '29594', '29614', '29615', '29616']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = '45774'
type = 'enhancement'
url = 'https://bugs.python.org/issue45573'
versions = ['Python 3.11']

[tiran]

pkg-config [1] is a standard tool on Linux and other platforms to detect presence of dependencies as well as to figure out which compiler and linker flags they require. Development packages provide a .pc file, e.g. ncurses provides a ncursesw.pc file.

$ pkg-config --libs --cflags ncursesw
-D_DEFAULT_SOURCE -D_XOPEN_SOURCE=600 -lncursesw -ltinfo

I propose to use a modified version of pkg-config's PKG_HAVE_DEFINE_WITH_MODULES macro in our configure script. On succss the modified macro defines:

• HAVE_FOO=1
• FOO_CFLAGS="some compile flags"
• FOO_LIBS="some library flags for linker"

On error, it sets nothing and does not cause configure to stop with an error.

The macro also allows users to override flags by setting FOO_CFLAGS and FOO_LIBS env vars. HAVE_FOO is added to pyconfig.h. The FOO_CFLAGS/LIBS are added to Makefile, from where it can be consumed by Modules/Setup.

Eventually Python could use the flags in setup.py, too. For now I would like to start with Modules/Setup. It is only used by some power users and has less risk of breaking the setup of beginners.

[1] https://www.freedesktop.org/wiki/Software/pkg-config/

[tiran]

I added some whitespace to Modules/Setup. All lines starting with #([a-z_]) compile cleanly on Fedora 34 with all dependencies available. dbm, tkinter, and sqlite are missing or untested.

$ sed -E -i 's/^#([a-z_])/\1/g' Modules/Setup
$ make
...
The following modules found by detect_modules() in setup.py, have been
built by the Makefile instead, as configured by the Setup files:
_abc               _asyncio           _bisect         
_blake2            _bz2               _codecs_cn      
_codecs_hk         _codecs_iso2022    _codecs_jp      
_codecs_kr         _codecs_tw         _contextvars    
_crypt             _csv               _curses         
_curses_panel      _datetime          _elementtree    
_hashlib           _heapq             _json           
_lsprof            _lzma              _md5            
_multibytecodec    _opcode            _pickle         
_posixsubprocess   _queue             _random         
_sha1              _sha256            _sha3           
_sha512            _socket            _ssl            
_statistics        _struct            _testbuffer     
_testimportmultiple   _testinternalcapi   _testmultiphase 
_typing            _xxsubinterpreters   _xxtestfuzz     
_zoneinfo          array              audioop         
binascii           cmath              fcntl           
grp                math               mmap            
nis                ossaudiodev        pwd             
pyexpat            readline           resource        
select             spwd               syslog          
termios            time               unicodedata     
xxlimited          xxlimited_35       zlib   

$  ./python -c "import sys; print(sys.builtin_module_names)"
('_abc', '_ast', '_asyncio', '_bisect', '_blake2', '_bz2', '_codecs', '_codecs_cn', '_codecs_hk', '_codecs_iso2022', '_codecs_jp', '_codecs_kr', '_codecs_tw', '_collections', '_contextvars', '_crypt', '_csv', '_curses', '_curses_panel', '_datetime', '_elementtree', '_functools', '_hashlib', '_heapq', '_imp', '_io', '_json', '_locale', '_lsprof', '_lzma', '_md5', '_multibytecodec', '_opcode', '_operator', '_pickle', '_posixsubprocess', '_queue', '_random', '_sha1', '_sha256', '_sha3', '_sha512', '_signal', '_socket', '_sre', '_ssl', '_stat', '_statistics', '_string', '_struct', '_symtable', '_testbuffer', '_testimportmultiple', '_testinternalcapi', '_testmultiphase', '_thread', '_tokenize', '_tracemalloc', '_typing', '_warnings', '_weakref', '_xxsubinterpreters', '_xxtestfuzz', '_zoneinfo', 'array', 'atexit', 'audioop', 'binascii', 'builtins', 'cmath', 'errno', 'faulthandler', 'fcntl', 'gc', 'grp', 'itertools', 'marshal', 'math', 'mmap', 'nis', 'ossaudiodev', 'posix', 'pwd', 'pyexpat', 'readline', 'resource', 'select', 'spwd', 'sys', 'syslog', 'termios', 'time', 'unicodedata', 'xx', 'xxlimited', 'xxlimited_35', 'xxsubtype', 'zlib')
$ ldd python
        linux-vdso.so.1 (0x00007fffa1f40000)
        libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f6032fe2000)
        libdl.so.2 => /lib64/libdl.so.2 (0x00007f6032fdb000)
        libutil.so.1 => /lib64/libutil.so.1 (0x00007f6032fd6000)
        libm.so.6 => /lib64/libm.so.6 (0x00007f6032e92000)
        libcrypt.so.2 => /lib64/libcrypt.so.2 (0x00007f6032e58000)
        libnsl.so.2 => /lib64/libnsl.so.2 (0x00007f6032e3c000)
        libtirpc.so.3 => /lib64/libtirpc.so.3 (0x00007f6032e0a000)
        libbz2.so.1 => /lib64/libbz2.so.1 (0x00007f6032df7000)
        liblzma.so.5 => /lib64/liblzma.so.5 (0x00007f6032dcb000)
        libz.so.1 => /lib64/libz.so.1 (0x00007f6032db1000)
        libreadline.so.8 => /lib64/libreadline.so.8 (0x00007f6032d59000)
        libssl.so.1.1 => /lib64/libssl.so.1.1 (0x00007f6032cbc000)
        libcrypto.so.1.1 => /lib64/libcrypto.so.1.1 (0x00007f60329cc000)
        libncursesw.so.6 => /lib64/libncursesw.so.6 (0x00007f603298d000)
        libtinfo.so.6 => /lib64/libtinfo.so.6 (0x00007f603295e000)
        libpanel.so.6 => /lib64/libpanel.so.6 (0x00007f6032958000)
        libc.so.6 => /lib64/libc.so.6 (0x00007f6032789000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f6033029000)
        libgssapi_krb5.so.2 => /lib64/libgssapi_krb5.so.2 (0x00007f6032732000)
        libkrb5.so.3 => /lib64/libkrb5.so.3 (0x00007f6032652000)
        libk5crypto.so.3 => /lib64/libk5crypto.so.3 (0x00007f603263a000)
        libcom_err.so.2 => /lib64/libcom_err.so.2 (0x00007f6032633000)
        libncurses.so.6 => /lib64/libncurses.so.6 (0x00007f6032606000)
        libkrb5support.so.0 => /lib64/libkrb5support.so.0 (0x00007f60325f5000)
        libkeyutils.so.1 => /lib64/libkeyutils.so.1 (0x00007f60325ec000)
        libresolv.so.2 => /lib64/libresolv.so.2 (0x00007f60325d2000)
        libselinux.so.1 => /lib64/libselinux.so.1 (0x00007f60325a6000)
        libpcre2-8.so.0 => /lib64/libpcre2-8.so.0 (0x00007f603250f000)

[tiran]

$ find build/lib.linux-x86_64-3.11/ -name '*.so' | sort
build/lib.linux-x86_64-3.11/_ctypes.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_ctypes_test.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_dbm.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_decimal.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_gdbm.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_multiprocessing.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_posixshmem.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_sqlite3.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_testcapi.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_tkinter.cpython-311-x86_64-linux-gnu.so
build/lib.linux-x86_64-3.11/_uuid.cpython-311-x86_64-linux-gnu.so

[tiran]

We can detect majority of our dependencies with pkg-config. The use of pkg-config has some benefits:

• Distro's provide the .pc files in their -dev / -devel packages. The presence of a .pc file indicates that all development dependencies are available.

• pkg-config does the right thing for non-standard include and libraries directories as well as multiarch builds. In case a library or header is not on the default search path, pkg-config returns necessary -I and -L flags.

• At least the pkgconf implementation of pkg-config standard search /usr/local and ~/.local/ directories for .pc files. Cases like bpo-45774: Autoconfiscate SQLite detection (GH-29507) #29507 are handled correctly. On FreeBSD "pkgconf sqlite3 --cflags --libs" returns "-I/usr/local/include -L/usr/local/lib -lsqlite3".

• pkg-config understands dependencies. For example "pkg-config --libs tk" returns linker flags for TK *and* TCL.

• pkg-config can check for module version, e.g. "pkg-config sqlite3 --atleast-version=3.7.15"

pkg-config modules:

readline, libedit
ncursesw, ncurses, panel, tinfo
sqlite3
zlib
bzip2
liblzma
expat
uuid (Linux's util-linux uuid)
libffi
libnsl, libtirpc
libcrypt
tcl, tk
openssl, libssl, libcrypto

modules / libraries without pkg-config modules:

decimal: libmpdec
gdbm: gdbm
dbm: gdbm_compat, ndbm, libdb (bdb)

To simplify use of flags in Modules/Setup, I propose to add two make variables for each module that needs cflags and ldflags:

f"MODULE_{ext.name.upper()}CFLAGS"
f"MODULE{ext.name.upper()}_LDFLAGS"

e.g. for the _ssl module:

  MODULE__SSL_CFLAGS=
  MODULE__SSL_LDFLAGS=-lssl -lcrypto

Then use the flags from Makefile in setup.py:

    def update_extension_flags(self, ext):
        name = ext.name.upper()
        cflags = sysconfig.get_config_var(f"MODULE_{name}_CFLAGS")
        if cflags:
            ext.extra_compile_args.extend(shlex.split(cflags))
        ldflags = sysconfig.get_config_var(f"MODULE_{name}_LDFLAGS")
        if ldflags:
            ext.extra_link_args.extend(shlex.split(ldflags))
        return ext

Finally update Modules/makesetup to use the new variables, too.

[brettcannon]

SGTM!

[ned-deily]

SGTM,2

This all sounds great. I think a goal here should be to remove all header and lib file searching from setup.py as that has always been a bug magnet. Perhaps one workaround for those libs that don't (yet) provide .pc files would be for us to supply reasonable default .pc for them, if possible, so that most builders wouldn't have to set the MODULE_xxx_*FLAGS variables while still removing the header/lib file searching from setup.py? Also it would be great to document how using pkg-config works in generic cross-compiling cases; we should strive to make at least the most common cases just work with minimal tweaking of configure arguments.

[tiran]

gdbmmodule and dbmmodule need special treatment anyway. macOS has dbm-API build into libc. Linux has either libgdbm_compat, libndbm, or libdb. The --with-dbmliborder makes it even more interesting. Users can override in which order they want to probe for gdbm, ndbm, and libdb. We have to keep the "manual" library and header checks.

Cross-compilation with pkg-config uses a trivial wrapper script, https://autotools.io/pkgconfig/cross-compiling.html . The build system has to create and provide a script with correct sysroot setting.

Is traditional cross-compiling work still useful these days anyway? Emulation has made big leaps in the last decade. Emulated cross compiling with qemu has become widespread. It also has the big advantage that you can run the test suite on the emulated hardware and verify that your binaries work. AFAIK Fedora's build system uses qemu for a bunch of hardware targets.