Name: pypy Version: 1.4.1 Release: 3%{?dist} Summary: Python implementation with a Just-In-Time compiler Group: Development/Languages # LGPL and another free license we'd need to ask spot about are present in some # java jars that we're not building with atm (in fact, we're deleting them # before building). If we restore those we'll have to work out the new # licensing terms License: MIT and Python and UCD URL: http://pypy.org/ BuildRoot: %{_tmppath}/%{name}-%{version}-%{release}-root-%(%{__id_u} -n) # High-level configuration of the build: # PyPy consists of an implementation of an interpreter (with JIT compilation) # for the full Python language written in a high-level language, leaving many # of the implementation details as "pluggable" policies. # # The implementation language is then compiled down to .c code, from which we # obtain a binary. # # This allows us to build a near-arbitrary collection of different # implementations of Python with differing tradeoffs # # (As it happens, the implementation language is itself Python, albeit a # restricted subset "RPython", chosen to making it amenable to being compiled. # The result implements the full Python language though) # We could build many different implementations of Python. # For now, let's focus on the implementation that appears to be receiving the # most attention upstream: the JIT-enabled build, with all standard # optimizations # Building a configuration can take significant time: # A build of pypy (with jit) on i686 took 77 mins: # [Timer] Timings: # [Timer] annotate --- 583.3 s # [Timer] rtype_lltype --- 760.9 s # [Timer] pyjitpl_lltype --- 567.3 s # [Timer] backendopt_lltype --- 375.6 s # [Timer] stackcheckinsertion_lltype --- 54.1 s # [Timer] database_c --- 852.2 s # [Timer] source_c --- 1007.3 s # [Timer] compile_c --- 419.9 s # [Timer] =========================================== # [Timer] Total: --- 4620.5 s # # A build of pypy (nojit) on x86_64 took about an hour: # [Timer] Timings: # [Timer] annotate --- 537.5 s # [Timer] rtype_lltype --- 667.3 s # [Timer] backendopt_lltype --- 385.4 s # [Timer] stackcheckinsertion_lltype --- 42.5 s # [Timer] database_c --- 625.3 s # [Timer] source_c --- 1040.2 s # [Timer] compile_c --- 273.9 s # [Timer] =========================================== # [Timer] Total: --- 3572.0 s # # # A build of pypy-stackless on i686 took about 87 mins: # [Timer] Timings: # [Timer] annotate --- 584.2 s # [Timer] rtype_lltype --- 777.3 s # [Timer] backendopt_lltype --- 365.9 s # [Timer] stackcheckinsertion_lltype --- 39.3 s # [Timer] database_c --- 1089.6 s # [Timer] source_c --- 1868.6 s # [Timer] compile_c --- 490.4 s # [Timer] =========================================== # [Timer] Total: --- 5215.3 s # Should we build a "pypy" binary? (with jit) # pypy-1.4/pypy/jit/backend/detect_cpu.py:getcpuclassname currently supports the # following options: # 'i386', 'x86' # 'x86-without-sse2': # 'x86_64' # 'cli' # 'llvm' %ifarch %{ix86} x86_64 # FIXME: is there a better way of expressing "intel" here? %global with_jit 1 %else %global with_jit 0 %endif # Should we build a "pypy-stackless" binary? %global with_stackless 0 # Easy way to enable/disable verbose logging: %global verbose_logs 0 %global pypyprefix %{_libdir}/pypy-%{version} %global pylibver 2.5.2 # We refer to this subdir of the source tree in a few places during the build: %global goal_dir pypy/translator/goal # Turn off the brp-python-bytecompile postprocessing script # We manually invoke it later on, using the freshly built pypy binary %global __os_install_post \ %(echo '%{__os_install_post}' | sed -e 's!/usr/lib[^[:space:]]*/brp-python-bytecompile[[:space:]].*$!!g') # Source and patches: Source0: http://pypy.org/download/pypy-%{version}-src.tar.bz2 # Edit a translator file for linux in order to configure our cflags and dynamic libffi Patch0: pypy-1.4-config.patch # By default, if built at a tty, the translation process renders a Mandelbrot # set to indicate progress. # This obscures useful messages, and may waste CPU cycles, so suppress it, and # merely render dots: Patch1: pypy-1.2-suppress-mandelbrot-set-during-tty-build.patch # test_commmands fails on SELinux systems due to a change in the output # of "ls" (http://bugs.python.org/issue7108) Patch2: fix-test_commands-expected-ls-output-issue7108.patch # When locating the pypy standard libraries, look first within # LIBRARY_INSTALLATION_PATH. # We convert this from being a non-existant variable into a string literal # with the value of "pypyprefix" in the "prep" phase below. # # We still use the scanning relative to the binary location when invoking a # pypy binary during the build (e.g. during "check") # # Sent upstream (with caveats) as: # https://codespeak.net/issue/pypy-dev/issue614 Patch3: pypy-1.4.1-add-LIBRARY_INSTALLATION_PATH.patch # Build-time requirements: BuildRequires: python-devel # FIXME: I'm seeing errors like this in the logs: # [translation:WARNING] The module '_rawffi' is disabled # [translation:WARNING] because importing pypy.rlib.libffi raised ImportError # [translation:WARNING] 'libffi.a' not found in ['/usr/lib/libffi', '/usr/lib'] # Presumably we need to fix things to support dynamically-linked libffi BuildRequires: libffi-devel BuildRequires: zlib-devel BuildRequires: bzip2-devel BuildRequires: ncurses-devel BuildRequires: expat-devel BuildRequires: openssl-devel BuildRequires: valgrind-devel # Used by the selftests, though not by the build: BuildRequires: gc-devel BuildRequires: /usr/bin/execstack # pypy is bundling these so we delete them in %%prep. I don't think they are # needed unless we build pypy targetted at running on the jvm. #BuildRequires: jna #BuildRequires: jasmin # Not yet in Fedora # Metadata for the core package (the JIT build): Requires: pypy-libs = %{version}-%{release} %description PyPy's implementation of Python, featuring a Just-In-Time compiler, and various optimized implementations of the standard types (strings, dictionaries, etc) %package libs Group: Development/Languages Summary: Run-time libraries used by PyPy implementations of Python %description libs Libraries required by the various PyPy implementations of Python. %if 0%{with_stackless} %package stackless Group: Development/Languages Summary: Stackless Python interpreter built using PyPy Requires: pypy-libs = %{version}-%{release} %description stackless Build of PyPy with support for micro-threads for massive concurrency %endif %if 0%{with_stackless} %package stackless Group: Development/Languages Summary: Stackless Python interpreter built using PyPy Requires: pypy-libs = %{version}-%{release} %description stackless Build of PyPy with support for micro-threads for massive concurrency %endif %prep %setup -q -n pypy-%{version}-src %patch0 -p1 -b .configure-fedora %patch1 -p1 -b .suppress-mandelbrot-set-during-tty-build pushd lib-python/%{pylibver} %patch2 -p0 popd # Look for the pypy libraries within LIBRARY_INSTALLATION_PATH first: %patch3 -p1 # Fixup LIBRARY_INSTALLATION_PATH to be a string literal containing our value # for "pypyprefix": sed -i \ -e 's|LIBRARY_INSTALLATION_PATH|"%{pypyprefix}"|' \ pypy/translator/goal/app_main.py # Replace /usr/local/bin/python shebangs with /usr/bin/python: find -name "*.py" -exec \ sed \ -i -e "s|/usr/local/bin/python|/usr/bin/python|" \ "{}" \ \; find . -name '*.jar' -exec rm \{\} \; # Remove stray ".svn" directories present within the 1.4.1 tarball # (reported as https://codespeak.net/issue/pypy-dev/issue612 ) find . -path '*/.svn*' -delete # Remove DOS batch files: find -name "*.bat"|xargs rm -f # The "demo" directory gets auto-installed by virture of being listed in %doc # Remove shebang lines from demo .py files, and remove executability from them: for f in demo/bpnn.py ; do # Detect shebang lines && remove them: sed -e '/^#!/Q 0' -e 'Q 1' $f \ && sed -i '1d' $f chmod a-x $f done %build BuildPyPy() { ExeName=$1 Options=$2 echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "STARTING BUILD OF: $ExeName" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" pushd %{goal_dir} # The build involves invoking a python script, passing in particular # arguments, environment variables, etc. # Some notes on those follow: # The generated binary embeds copies of the values of all environment # variables. We need to unset "RPM_BUILD_ROOT" to avoid a fatal error from # /usr/lib/rpm/check-buildroot # during the postprocessing of the rpmbuild, complaining about this # reference to the buildroot # By default, pypy's autogenerated C code is placed in # /tmp/usession-N # # and it appears that this stops rpm from extracting the source code to the # debuginfo package # # The logic in pypy-1.4/pypy/tool/udir.py indicates that it is generated in: # $PYPY_USESSION_DIR/usession-$PYPY_USESSION_BASENAME-N # and so we set PYPY_USESSION_DIR so that this tempdir is within the build # location, and set $PYPY_USESSION_BASENAME so that the tempdir is unique # for each invocation of BuildPyPy # Compilation flags for C code: # pypy-1.4/pypy/translator/c/genc.py:gen_makefile # assembles a Makefile within # THE_UDIR/testing_1/Makefile # calling out to platform.gen_makefile # For us, that's # pypy-1.4/pypy/translator/platform/linux.py: class BaseLinux(BasePosix): # which by default has: # CFLAGS = ['-O3', '-pthread', '-fomit-frame-pointer', # '-Wall', '-Wno-unused'] # plus all substrings from CFLAGS in the environment. # This is used to generate a value for CFLAGS that's written into the Makefile # https://bugzilla.redhat.com/show_bug.cgi?id=588941#c18 # The generated Makefile compiles the .c files into assembler (.s), rather # than direct to .o It then post-processes this assembler to locate # garbage-collection roots (building .lbl.s and .gcmap files, and a # "gcmaptable.s"). (The modified .lbl.s files have extra code injected # within them). # Unfortunately, the code to do this: # pypy-1.4/pypy/translator/c/gcc/trackgcroot.py # doesn't interract well with the results of using our standard build flags. # For now, filter our CFLAGS of everything that could be conflicting with # pypy. Need to check these and reenable ones that are okay later. export CFLAGS=$(echo "$RPM_OPT_FLAGS" | sed -e 's/-Wp,-D_FORTIFY_SOURCE=2//' -e 's/-fexceptions//' -e 's/-fstack-protector//' -e 's/--param=ssp-buffer-size=4//' -e 's/-O2//' -e 's/-fasynchronous-unwind-tables//' -e 's/-march=i686//' -e 's/-mtune=atom//') # If we're already built the JIT-enabled "pypy", then use it for subsequent # builds (of other configurations): if test -x './pypy' ; then INTERP='./pypy' else INTERP='python' fi # Here's where we actually invoke the build: time \ RPM_BUILD_ROOT= \ PYPY_USESSION_DIR=$(pwd) \ PYPY_USESSION_BASENAME=$ExeName \ $INTERP translate.py \ %if 0%{verbose_logs} --translation-verbose \ %endif --cflags="$CFLAGS" \ --batch \ --output=$ExeName \ $Options echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "FINISHED BUILDING: $ExeName" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" popd } %if 0%{with_jit} BuildPyPy \ pypy \ "-Ojit" %endif %if 0%{with_stackless} BuildPyPy \ pypy-stackless \ "--stackless" %endif %install rm -rf $RPM_BUILD_ROOT # Install the various executables: InstallPyPy() { ExeName=$1 install -m 755 %{goal_dir}/$ExeName %{buildroot}/%{_bindir} # The generated machine code doesn't need an executable stack, but # one of the assembler files (gcmaptable.s) doesn't have the necessary # metadata to inform gcc of that, and thus gcc pessimistically assumes # that the built binary does need an executable stack. # # Reported upstream as: https://codespeak.net/issue/pypy-dev/issue610 # # I tried various approaches involving fixing the build, but the simplest # approach is to postprocess the ELF file: execstack --clear-execstack %{buildroot}/%{_bindir}/$ExeName } mkdir -p %{buildroot}/%{_bindir} %if 0%{with_jit} InstallPyPy pypy %endif %if 0%{with_stackless} InstallPyPy pypy-stackless %endif # Install the various support libraries as described at: # http://codespeak.net/pypy/dist/pypy/doc/getting-started-python.html#installation # which refers to a "PREFIX" found relative to the location of the binary. # Given that the pypy binaries will be in /usr/bin, PREFIX can be # "../share/pypy-1.2" relative to that directory, i.e. /usr/share/pypy-1.2 # # Running "strace" on a built binary indicates that it searches within # PREFIX/lib-python/modified-2.5.2 # not # PREFIX/lib-python/modified.2.5.2 # as given on the above page, i.e. it uses '-' not '.' mkdir -p %{buildroot}/%{pypyprefix} cp -a lib-python %{buildroot}/%{pypyprefix} cp -a lib_pypy %{buildroot}/%{pypyprefix} # Remove a text file that documents which selftests fail on Win32: rm %{buildroot}/%{pypyprefix}/lib-python/win32-failures.txt # Remove shebang lines from .py files that aren't executable, and # remove executability from .py files that don't have a shebang line: find \ %{buildroot} \ -name "*.py" \ \( \ \( \! -perm /u+x,g+x,o+x -exec sed -e '/^#!/Q 0' -e 'Q 1' {} \; \ -print -exec sed -i '1d' {} \; \ \) \ -o \ \( \ -perm /u+x,g+x,o+x ! -exec grep -m 1 -q '^#!' {} \; \ -exec chmod a-x {} \; \ \) \ \) mkdir -p %{buildroot}/%{pypyprefix}/site-packages # pypy uses .pyc files by default (--objspace-usepycfiles), but has a slightly # different bytecode format to CPython. It doesn't use .pyo files: the -O flag # is treated as a "dummy optimization flag for compatibility with C Python" # # pypy-1.4/pypy/module/imp/importing.py has this comment: # XXX picking a magic number is a mess. So far it works because we # have only two extra opcodes, which bump the magic number by +1 and # +2 respectively, and CPython leaves a gap of 10 when it increases # its own magic number. To avoid assigning exactly the same numbers # as CPython we always add a +2. We'll have to think again when we # get at the fourth new opcode :-( # # * CALL_LIKELY_BUILTIN +1 # * CALL_METHOD +2 # # In other words: # # default_magic -- used by CPython without the -U option # default_magic + 1 -- used by CPython with the -U option # default_magic + 2 -- used by PyPy without any extra opcode # ... # default_magic + 5 -- used by PyPy with both extra opcodes # # pypy-1.4/pypy/interpreter/pycode.py has: # # default_magic = (62141+2) | 0x0a0d0000 # this PyPy's magic # # (62131=CPython 2.5.1) # giving a value for "default_magic" for PyPy of 0xa0df2bf. # Note that this corresponds to the "default_magic + 2" from the comment above # In my builds: # $ ./pypy --info | grep objspace.opcodes # objspace.opcodes.CALL_LIKELY_BUILTIN: False # objspace.opcodes.CALL_METHOD: True # so I'd expect the magic number to be: # 0x0a0df2bf + 2 (the flag for CALL_METHOD) # giving # 0x0a0df2c1 # # I'm seeing # c1 f2 0d 0a # as the first four bytes of the .pyc files, which is consistent with this. # Bytecompile all of the .py files we ship, using our pypy binary, giving us # .pyc files for pypy. The script actually does the work twice (passing in -O # the second time) but it's simplest to reuse that script. # # The script has special-casing for .py files below # /usr/lib{64}/python[0-9].[0-9] # but given that we're installing into a different path, the supplied "default" # implementation gets used instead. # # Note that some of the test files deliberately contain syntax errors, so # we pass 0 for the second argument ("errors_terminate"): /usr/lib/rpm/brp-python-bytecompile \ %{buildroot}/%{_bindir}/pypy \ 0 %check topdir=$(pwd) SkipTest() { # Append the given test name to TESTS_TO_SKIP TEST_NAME=$1 TESTS_TO_SKIP="$TESTS_TO_SKIP $TEST_NAME" } CheckPyPy() { # We'll be exercising one of the freshly-built binaries using the # test suite from the standard library (overridden in places by pypy's # modified version) ExeName=$1 echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "STARTING TEST OF: $ExeName" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" pushd %{goal_dir} # Gather a list of tests to skip, due to known failures # TODO: report these failures to pypy upstream TESTS_TO_SKIP="" # Test failures relating to missing codecs # Hopefully when pypy merges to 2.7 support we'll gain these via a # refreshed stdlib: # test_codecencodings_cn: # all tests fail, with one of # unknown encoding: gb18030 # unknown encoding: gb2312 # unknown encoding: gbk SkipTest test_codecencodings_cn # test_codecencodings_hk: # all tests fail, with: # unknown encoding: big5hkscs SkipTest test_codecencodings_hk # test_codecencodings_jp: # all tests fail, with one of: # unknown encoding: cp932 # unknown encoding: euc_jisx0213 # unknown encoding: euc_jp # unknown encoding: shift_jis # unknown encoding: shift_jisx0213 SkipTest test_codecencodings_jp # test_codecencodings_kr: # all tests fail, with one of: # unknown encoding: cp949 # unknown encoding: euc_kr # unknown encoding: johab SkipTest test_codecencodings_kr # test_codecencodings_tw: # all tests fail, with: # unknown encoding: big5 SkipTest test_codecencodings_tw # test_multibytecodec: # 14 failures out of 15, due to: # unknown encoding: euc-kr # unknown encoding: gb2312 # unknown encoding: jisx0213 # unknown encoding: cp949 # unknown encoding: iso2022-jp # unknown encoding: gb18030 SkipTest test_multibytecodec # # Other failures: # # test_asynchat: # seems to hang on this test, within test_line_terminator SkipTest test_asynchat # test_compiler: # 4 errors out of 13: # testSourceCodeEncodingsError # testWith # testWithAss # testYieldExpr SkipTest test_compiler # test_ctypes: # failures=17, errors=20, out of 132 tests SkipTest test_ctypes # test_frozen: # TestFailed: import __hello__ failed:No module named __hello__ SkipTest test_frozen # test_iterlen: # 24 failures out of 25, apparently all due to TypeError SkipTest test_iterlen # test_parser: # 12 failures out of 15 SkipTest test_parser # test_platform: # Koji builds show: # test test_platform failed -- errors occurred in test.test_platform.PlatformTest SkipTest test_platform # test_socket: # testSockName can fail in Koji with: # my_ip_addr = socket.gethostbyname(socket.gethostname()) # gaierror: (-3, 'Temporary failure in name resolution') SkipTest test_socket # test_sort: # some failures SkipTest test_sort # test_sqlite: # 3 of the sqlite3.test.dbapi.ExtensionTests raise: # ProgrammingError: Incomplete statement '' SkipTest test_sqlite # test_traceback: # works when run standalone; failures seen when run as part of a suite SkipTest test_traceback # test_zlib: # failure seen in Koji, not sure of reason why: # test test_zlib failed -- Traceback (most recent call last): # File "/builddir/build/BUILD/pypy-1.4.1-src/lib-python/2.5.2/test/test_zlib.py", line 72, in test_baddecompressobj # self.assertRaises(ValueError, zlib.decompressobj, 0) # AssertionError: ValueError not raised SkipTest test_zlib # Run the built binary through the selftests: time ./$ExeName -m test.regrtest -x $TESTS_TO_SKIP popd # Doublecheck pypy's own test suite, using the built pypy binary: # Disabled for now: # x86_64 shows various failures inside: # jit/backend/x86/test # followed by a segfault inside # jit/backend/x86/test/test_runner.py # # i686 shows various failures inside: # jit/backend/x86/test # with the x86_64 failure leading to cancellation of the i686 build # Here's the disabled code: # pushd pypy # time translator/goal/$ExeName test_all.py # popd echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "FINISHED TESTING: $ExeName" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" echo "--------------------------------------------------------------" } %if 0%{with_jit} CheckPyPy pypy %endif %if 0%{with_stackless} CheckPyPy pypy-stackless %endif %clean rm -rf $RPM_BUILD_ROOT %files libs %defattr(-,root,root,-) %doc LICENSE README demo %dir %{pypyprefix} %dir %{pypyprefix}/lib-python %{pypyprefix}/lib-python/%{pylibver}/ %{pypyprefix}/lib-python/modified-%{pylibver}/ %{pypyprefix}/lib-python/conftest.py* %{pypyprefix}/lib_pypy/ %{pypyprefix}/site-packages/ %if 0%{with_jit} %files %defattr(-,root,root,-) %doc LICENSE README %{_bindir}/pypy %endif %if 0%{with_stackless} %files stackless %defattr(-,root,root,-) %doc LICENSE README %{_bindir}/pypy-stackless %endif %changelog * Wed Dec 22 2010 David Malcolm - 1.4.1-3 - try to respect the FHS by installing libraries below libdir, rather than datadir; patch app_main.py to look in this installation location first when scanning for the pypy library directories. - clarifications and corrections to the comments in the specfile * Wed Dec 22 2010 David Malcolm - 1.4.1-2 - remove .svn directories - disable verbose logging - add a %%check section - introduce %%goal_dir variable, to avoid repetition - remove shebang line from demo/bpnn.py, as we're treating this as a documentation file - regenerate patch 2 to apply without generating a .orig file * Tue Dec 21 2010 David Malcolm - 1.4.1-1 - 1.4.1; fixup %%setup to reflect change in toplevel directory in upstream source tarball - apply SELinux fix to the bundled test_commands.py (patch 2) * Wed Dec 15 2010 David Malcolm - 1.4-4 - rename the jit build and subpackge to just "pypy", and remove the nojit and sandbox builds, as upstream now seems to be focussing on the JIT build (with only stackless called out in the getting-started-python docs); disable stackless for now - add a verbose_logs specfile boolean; leave it enabled for now (whilst fixing build issues) - add more comments, and update others to reflect 1.2 -> 1.4 changes - re-enable debuginfo within CFLAGS ("-g") - add the LICENSE and README to all subpackages - ensure the built binaries don't have the "I need an executable stack" flag - remove DOS batch files during %%prep (idlelib.bat) - remove shebang lines from .py files that aren't executable, and remove executability from .py files that don't have a shebang line (taken from our python3.spec) - bytecompile the .py files into .pyc files in pypy's bytecode format * Sun Nov 28 2010 Toshio Kuratomi - 1.4-3 - BuildRequire valgrind-devel - Install pypy library from the new directory - Disable building with our CFLAGS for now because they are causing a build failure. - Include site-packages directory * Sat Nov 27 2010 Toshio Kuratomi - 1.4-2 - Add patch to configure the build to use our CFLAGS and link libffi dynamically * Sat Nov 27 2010 Toshio Kuratomi - 1.4-1 - Update to 1.4 - Drop patch for py2.6 that's in this build - Switch to building pypy with itself once pypy is built once as recommended by upstream - Remove bundled, prebuilt java libraries - Fix license tag - Fix source url - Version pypy-libs Req * Tue May 4 2010 David Malcolm - 1.2-2 - cherrypick r72073 from upstream SVN in order to fix the build against python 2.6.5 (patch 2) * Wed Apr 28 2010 David Malcolm - 1.2-1 - initial packaging