python3: generalise make commands
PATCH 1/5 python3: generalise make commands
We're repeating the same make invocation over and over, twice without setting OPT=${CFLAGS} which doesn't seem right.
Centralise the make invocation to clean up the tasks.
--- meta/recipes-devtools/python/python3_3.5.5.bb | 48 +++++++++------------------ 1 file changed, 16 insertions(+), 32 deletions(-)
meta/recipes-devtools/python/python3_3.5.5.bb +16 -32
@@ -95,6 +95,18 @@ do_configure_append() { autoreconf -Wcross --verbose --install --force --exclude=autopoint ../Python-${PV}/Modules/_ctypes/libffi } +run_make() { + oe_runmake HOSTPGEN=${STAGING_BINDIR_NATIVE}/python3-native/pgen \ + HOSTPYTHON=${STAGING_BINDIR_NATIVE}/python3-native/python3 \ + STAGING_LIBDIR=${STAGING_LIBDIR} \ + STAGING_INCDIR=${STAGING_INCDIR} \ + STAGING_BASELIBDIR=${STAGING_BASELIBDIR} \ + LIB=${baselib} \ + ARCH=${TARGET_ARCH} \ + OPT="${CFLAGS}" \ + "$@" +} + do_compile() { # regenerate platform specific files, because they depend on system headers cd ${S}/Lib/plat-linux*
@@ -125,14 +137,7 @@ do_compile() { # then call do_install twice we get Makefile.orig == Makefile.sysroot install -m 0644 Makefile Makefile.sysroot - oe_runmake HOSTPGEN=${STAGING_BINDIR_NATIVE}/python3-native/pgen \ - HOSTPYTHON=${STAGING_BINDIR_NATIVE}/python3-native/python3 \ - STAGING_LIBDIR=${STAGING_LIBDIR} \ - STAGING_INCDIR=${STAGING_INCDIR} \ - STAGING_BASELIBDIR=${STAGING_BASELIBDIR} \ - LIB=${baselib} \ - ARCH=${TARGET_ARCH} \ - OPT="${CFLAGS}" profile-opt + run_make profile-opt if ${@bb.utils.contains('MACHINE_FEATURES', 'qemu-usermode', 'true', 'false', d)}; then qemu_binary="${@qemu_wrapper_cmdline(d, '${STAGING_DIR_TARGET}', ['${B}', '${STAGING_DIR_TARGET}/${base_libdir}'])}"
@@ -145,14 +150,7 @@ EOF ./pgo-image-qemuwrapper ${B}/python ${PYTHON3_PROFILE_TASK} || true fi - oe_runmake HOSTPGEN=${STAGING_BINDIR_NATIVE}/python3-native/pgen \ - HOSTPYTHON=${STAGING_BINDIR_NATIVE}/python3-native/python3 \ - STAGING_LIBDIR=${STAGING_LIBDIR} \ - STAGING_INCDIR=${STAGING_INCDIR} \ - STAGING_BASELIBDIR=${STAGING_BASELIBDIR} \ - LIB=${baselib} \ - ARCH=${TARGET_ARCH} \ - OPT="${CFLAGS}" clean_and_use_profile + run_make clean_and_use_profile } do_install() {
@@ -165,23 +163,9 @@ do_install() { # rerun the build once again with original makefile this time # run install in a separate step to avoid compile/install race - oe_runmake HOSTPGEN=${STAGING_BINDIR_NATIVE}/python3-native/pgen \ - HOSTPYTHON=${STAGING_BINDIR_NATIVE}/python3-native/python3 \ - STAGING_LIBDIR=${STAGING_LIBDIR} \ - STAGING_INCDIR=${STAGING_INCDIR} \ - STAGING_BASELIBDIR=${STAGING_BASELIBDIR} \ - LIB=${baselib} \ - ARCH=${TARGET_ARCH} \ - DESTDIR=${D} LIBDIR=${libdir} build_all_use_profile + run_make DESTDIR=${D} LIBDIR=${libdir} build_all_use_profile - oe_runmake HOSTPGEN=${STAGING_BINDIR_NATIVE}/python3-native/pgen \ - HOSTPYTHON=${STAGING_BINDIR_NATIVE}/python3-native/python3 \ - STAGING_LIBDIR=${STAGING_LIBDIR} \ - STAGING_INCDIR=${STAGING_INCDIR} \ - STAGING_BASELIBDIR=${STAGING_BASELIBDIR} \ - LIB=${baselib} \ - ARCH=${TARGET_ARCH} \ - DESTDIR=${D} LIBDIR=${libdir} install + run_make DESTDIR=${D} LIBDIR=${libdir} install # avoid conflict with 2to3 from Python 2 rm -f ${D}/${bindir}/2to3
Signature · 1 line
2.11.0
PATCH 2/5 python3: add toggle for profile-guided-optimisation
Add a PACKAGECONFIG for profile-guided-optimisation, and default to whether qemu-usermode is available.
Move --enable-optimizations to the pgo optimisation as all this currently does is use the PGO rules, causing a performance hit if PGO isn't actually used.
--- meta/recipes-devtools/python/python3_3.5.5.bb | 54 +++++++++++++++------------ 1 file changed, 31 insertions(+), 23 deletions(-)
meta/recipes-devtools/python/python3_3.5.5.bb +31 -23
@@ -73,7 +73,8 @@ CACHED_CONFIGUREVARS = "ac_cv_have_chflags=no \ TARGET_CC_ARCH += "-DNDEBUG -fno-inline" SDK_CC_ARCH += "-DNDEBUG -fno-inline" EXTRA_OEMAKE += "CROSS_COMPILE=yes" -EXTRA_OECONF += "CROSSPYTHONPATH=${STAGING_LIBDIR_NATIVE}/python${PYTHON_MAJMIN}/lib-dynload/ --without-ensurepip --enable-optimizations" +EXTRA_OECONF += "CROSSPYTHONPATH=${STAGING_LIBDIR_NATIVE}/python${PYTHON_MAJMIN}/lib-dynload/ --without-ensurepip" + PYTHON3_PROFILE_TASK ?= "${S}/Tools/pybench/pybench.py -n 1" export CROSS_COMPILE = "${TARGET_PREFIX}"
@@ -87,8 +88,10 @@ export CROSSPYTHONPATH = "${STAGING_LIBDIR_NATIVE}/python${PYTHON_MAJMIN}/lib-dy # No ctypes option for python 3 PYTHONLSBOPTS = "" -PACKAGECONFIG ??= "readline" +PACKAGECONFIG ??= "readline ${@bb.utils.contains('MACHINE_FEATURES', 'qemu-usermode', 'pgo', '', d)}" PACKAGECONFIG[readline] = ",,readline" +# Use profile guided optimisation by running PyBench inside qemu-user +PACKAGECONFIG[pgo] = "--enable-optimizations" do_configure_append() { rm -f ${S}/Makefile.orig
@@ -108,16 +111,16 @@ run_make() { } do_compile() { - # regenerate platform specific files, because they depend on system headers - cd ${S}/Lib/plat-linux* - include=${STAGING_INCDIR} ${STAGING_BINDIR_NATIVE}/python3-native/python3 \ - ${S}/Tools/scripts/h2py.py -i '(u_long)' \ - ${STAGING_INCDIR}/dlfcn.h \ - ${STAGING_INCDIR}/linux/cdrom.h \ - ${STAGING_INCDIR}/netinet/in.h \ - ${STAGING_INCDIR}/sys/types.h - sed -e 's,${STAGING_DIR_HOST},,g' -i *.py - cd - + # regenerate platform specific files, because they depend on system headers + cd ${S}/Lib/plat-linux* + include=${STAGING_INCDIR} ${STAGING_BINDIR_NATIVE}/python3-native/python3 \ + ${S}/Tools/scripts/h2py.py -i '(u_long)' \ + ${STAGING_INCDIR}/dlfcn.h \ + ${STAGING_INCDIR}/linux/cdrom.h \ + ${STAGING_INCDIR}/netinet/in.h \ + ${STAGING_INCDIR}/sys/types.h + sed -e 's,${STAGING_DIR_HOST},,g' -i *.py + cd - # remove any bogus LD_LIBRARY_PATH sed -i -e s,RUNSHARED=.*,RUNSHARED=, Makefile
@@ -137,20 +140,21 @@ do_compile() { # then call do_install twice we get Makefile.orig == Makefile.sysroot install -m 0644 Makefile Makefile.sysroot - run_make profile-opt - - if ${@bb.utils.contains('MACHINE_FEATURES', 'qemu-usermode', 'true', 'false', d)}; then - qemu_binary="${@qemu_wrapper_cmdline(d, '${STAGING_DIR_TARGET}', ['${B}', '${STAGING_DIR_TARGET}/${base_libdir}'])}" - cat > pgo-image-qemuwrapper << EOF + if ${@bb.utils.contains('PACKAGECONFIG', 'pgo', 'true', 'false', d)}; then + run_make profile-opt + qemu_binary="${@qemu_wrapper_cmdline(d, '${STAGING_DIR_TARGET}', ['${B}', '${STAGING_DIR_TARGET}/${base_libdir}'])}" + cat > pgo-image-qemuwrapper << EOF #!/bin/sh set -x $qemu_binary "\$@" EOF - chmod +x pgo-image-qemuwrapper - ./pgo-image-qemuwrapper ${B}/python ${PYTHON3_PROFILE_TASK} || true + chmod +x pgo-image-qemuwrapper + ./pgo-image-qemuwrapper ${B}/python ${PYTHON3_PROFILE_TASK} || true + run_make clean_and_use_profile + else + run_make libpython3.so + run_make fi - - run_make clean_and_use_profile } do_install() {
@@ -163,8 +167,12 @@ do_install() { # rerun the build once again with original makefile this time # run install in a separate step to avoid compile/install race - run_make DESTDIR=${D} LIBDIR=${libdir} build_all_use_profile - + if ${@bb.utils.contains('PACKAGECONFIG', 'pgo', 'true', 'false', d)}; then + run_make DESTDIR=${D} LIBDIR=${libdir} build_all_use_profile + else + run_make DESTDIR=${D} LIBDIR=${libdir} + fi + run_make DESTDIR=${D} LIBDIR=${libdir} install # avoid conflict with 2to3 from Python 2
Signature · 1 line
2.11.0
PATCH 3/5 python3: trim PGO patch
In reply to Ross BurtonThere's no need to delete the line that removes the profile data, as we're not using it after the build. This reduces the size of the patch, making it easier to maintain.
--- .../0002-Makefile-add-target-to-split-profile-generation.patch | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
meta/recipes-devtools/python/python3/0002-Makefile-add-target-to-split-profile-generation.patch +2 -2
@@ -19,7 +19,7 @@ diff --git a/Makefile.pre.in b/Makefile.pre.in index 84bc3ff..017a2c4 100644 --- a/Makefile.pre.in +++ b/Makefile.pre.in -@@ -469,13 +469,11 @@ profile-opt: +@@ -469,13 +469,12 @@ profile-opt: $(MAKE) profile-removal $(MAKE) build_all_generate_profile $(MAKE) profile-removal
@@ -31,7 +31,7 @@ index 84bc3ff..017a2c4 100644 @echo "Rebuilding with profile guided optimizations:" $(MAKE) clean $(MAKE) build_all_use_profile -- $(MAKE) profile-removal + $(MAKE) profile-removal build_all_generate_profile: $(MAKE) @DEF_MAKE_RULE@ CFLAGS_NODIST="$(CFLAGS) $(EXTRA_CFLAGS) $(PGO_PROF_GEN_FLAG) @LTOFLAGS@" LDFLAGS="$(LDFLAGS) $(PGO_PROF_GEN_FLAG) @LTOFLAGS@" LIBS="$(LIBS)"
Signature · 1 line
2.11.0
PATCH 4/5 python3: don't use runtime checks to identify float endianism
In reply to Ross BurtonPython uses AC_RUN_IFELSE to determine the byte order for floats and doubles, and falls back onto "I don't know" if it can't run code. This results in crippled floating point numbers in Python, and the regression tests fail.
Instead of running code, take a macro from autoconf-archive which compiles C with a special double in which has an ASCII representation, and then greps the binary to identify the format.
--- .../python/python3/float-endian.patch | 212 +++++++++++++++++++++ meta/recipes-devtools/python/python3_3.5.5.bb | 1 + 2 files changed, 213 insertions(+) create mode 100644 meta/recipes-devtools/python/python3/float-endian.patch
meta/recipes-devtools/python/python3/float-endian.patch new file +212 -0
@@ -0,0 +1,212 @@ +Python uses AC_RUN_IFELSE to determine the byte order for floats and doubles, +and falls back onto "I don't know" if it can't run code. This results in +crippled floating point numbers in Python, and the regression tests fail. + +Instead of running code, take a macro from autoconf-archive which compiles C +with a special double in which has an ASCII representation, and then greps the +binary to identify the format. + +Upstream-Status: Submitted [https://bugs.python.org/issue34585] +Signed-off-by: Ross Burton <[email protected]> + +From 50df2a4c3a65ed06322be7c26d42b06ce81730c1 Mon Sep 17 00:00:00 2001 +From: Ross Burton <[email protected]> +Date: Wed, 5 Sep 2018 11:45:52 +0100 +Subject: [PATCH] Don't do runtime test to get float byte order + +--- + configure.ac | 74 +++++------------------------------ + m4/ax_c_float_words_bigendian.m4 | 83 ++++++++++++++++++++++++++++++++++++++++ + 2 files changed, 92 insertions(+), 65 deletions(-) + create mode 100644 m4/ax_c_float_words_bigendian.m4 + +diff --git a/configure.ac b/configure.ac +index c9b755f0f4..1215969871 100644 +--- a/configure.ac ++++ b/configure.ac +@@ -9,6 +9,8 @@ AC_PREREQ(2.65) + + AC_INIT(python, PYTHON_VERSION, https://bugs.python.org/) + ++AC_CONFIG_MACRO_DIR(m4) ++ + AC_SUBST(BASECPPFLAGS) + if test "$srcdir" != . -a "$srcdir" != "$(pwd)"; then + # If we're building out-of-tree, we need to make sure the following +@@ -4128,77 +4130,19 @@ fi + # * Check for various properties of floating point * + # ************************************************** + +-AC_MSG_CHECKING(whether C doubles are little-endian IEEE 754 binary64) +-AC_CACHE_VAL(ac_cv_little_endian_double, [ +-AC_RUN_IFELSE([AC_LANG_SOURCE([[ +-#include <string.h> +-int main() { +- double x = 9006104071832581.0; +- if (memcmp(&x, "\x05\x04\x03\x02\x01\xff\x3f\x43", 8) == 0) +- return 0; +- else +- return 1; +-} +-]])], +-[ac_cv_little_endian_double=yes], +-[ac_cv_little_endian_double=no], +-[ac_cv_little_endian_double=no])]) +-AC_MSG_RESULT($ac_cv_little_endian_double) +-if test "$ac_cv_little_endian_double" = yes +-then +- AC_DEFINE(DOUBLE_IS_LITTLE_ENDIAN_IEEE754, 1, +- [Define if C doubles are 64-bit IEEE 754 binary format, stored +- with the least significant byte first]) +-fi +- +-AC_MSG_CHECKING(whether C doubles are big-endian IEEE 754 binary64) +-AC_CACHE_VAL(ac_cv_big_endian_double, [ +-AC_RUN_IFELSE([AC_LANG_SOURCE([[ +-#include <string.h> +-int main() { +- double x = 9006104071832581.0; +- if (memcmp(&x, "\x43\x3f\xff\x01\x02\x03\x04\x05", 8) == 0) +- return 0; +- else +- return 1; +-} +-]])], +-[ac_cv_big_endian_double=yes], +-[ac_cv_big_endian_double=no], +-[ac_cv_big_endian_double=no])]) +-AC_MSG_RESULT($ac_cv_big_endian_double) +-if test "$ac_cv_big_endian_double" = yes ++AX_C_FLOAT_WORDS_BIGENDIAN ++if test "$ax_cv_c_float_words_bigendian" = "yes" + then + AC_DEFINE(DOUBLE_IS_BIG_ENDIAN_IEEE754, 1, + [Define if C doubles are 64-bit IEEE 754 binary format, stored + with the most significant byte first]) +-fi +- +-# Some ARM platforms use a mixed-endian representation for doubles. +-# While Python doesn't currently have full support for these platforms +-# (see e.g., issue 1762561), we can at least make sure that float <-> string +-# conversions work. +-AC_MSG_CHECKING(whether C doubles are ARM mixed-endian IEEE 754 binary64) +-AC_CACHE_VAL(ac_cv_mixed_endian_double, [ +-AC_RUN_IFELSE([AC_LANG_SOURCE([[ +-#include <string.h> +-int main() { +- double x = 9006104071832581.0; +- if (memcmp(&x, "\x01\xff\x3f\x43\x05\x04\x03\x02", 8) == 0) +- return 0; +- else +- return 1; +-} +-]])], +-[ac_cv_mixed_endian_double=yes], +-[ac_cv_mixed_endian_double=no], +-[ac_cv_mixed_endian_double=no])]) +-AC_MSG_RESULT($ac_cv_mixed_endian_double) +-if test "$ac_cv_mixed_endian_double" = yes ++elif test "$ax_cv_c_float_words_bigendian" = "no" + then +- AC_DEFINE(DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754, 1, ++ AC_DEFINE(DOUBLE_IS_LITTLE_ENDIAN_IEEE754, 1, + [Define if C doubles are 64-bit IEEE 754 binary format, stored +- in ARM mixed-endian order (byte order 45670123)]) ++ with the least significant byte first]) ++else ++ AC_MSG_ERROR([Cannot identify floating point byte order]) + fi + + # The short float repr introduced in Python 3.1 requires the +diff --git a/m4/ax_c_float_words_bigendian.m4 b/m4/ax_c_float_words_bigendian.m4 +new file mode 100644 +index 0000000000..216b90d803 +--- /dev/null ++++ b/m4/ax_c_float_words_bigendian.m4 +@@ -0,0 +1,83 @@ ++# =============================================================================== ++# https://www.gnu.org/software/autoconf-archive/ax_c_float_words_bigendian.html ++# =============================================================================== ++# ++# SYNOPSIS ++# ++# AX_C_FLOAT_WORDS_BIGENDIAN([ACTION-IF-TRUE], [ACTION-IF-FALSE], [ACTION-IF-UNKNOWN]) ++# ++# DESCRIPTION ++# ++# Checks the ordering of words within a multi-word float. This check is ++# necessary because on some systems (e.g. certain ARM systems), the float ++# word ordering can be different from the byte ordering. In a multi-word ++# float context, "big-endian" implies that the word containing the sign ++# bit is found in the memory location with the lowest address. This ++# implementation was inspired by the AC_C_BIGENDIAN macro in autoconf. ++# ++# The endianness is detected by first compiling C code that contains a ++# special double float value, then grepping the resulting object file for ++# certain strings of ASCII values. The double is specially crafted to have ++# a binary representation that corresponds with a simple string. In this ++# implementation, the string "noonsees" was selected because the ++# individual word values ("noon" and "sees") are palindromes, thus making ++# this test byte-order agnostic. If grep finds the string "noonsees" in ++# the object file, the target platform stores float words in big-endian ++# order. If grep finds "seesnoon", float words are in little-endian order. ++# If neither value is found, the user is instructed to specify the ++# ordering. ++# ++# LICENSE ++# ++# Copyright (c) 2008 Daniel Amelang <[email protected]> ++# ++# Copying and distribution of this file, with or without modification, are ++# permitted in any medium without royalty provided the copyright notice ++# and this notice are preserved. This file is offered as-is, without any ++# warranty. ++ ++#serial 11 ++ ++AC_DEFUN([AX_C_FLOAT_WORDS_BIGENDIAN], ++ [AC_CACHE_CHECK(whether float word ordering is bigendian, ++ ax_cv_c_float_words_bigendian, [ ++ ++ax_cv_c_float_words_bigendian=unknown ++AC_COMPILE_IFELSE([AC_LANG_SOURCE([[ ++ ++double d = 90904234967036810337470478905505011476211692735615632014797120844053488865816695273723469097858056257517020191247487429516932130503560650002327564517570778480236724525140520121371739201496540132640109977779420565776568942592.0; ++ ++]])], [ ++ ++if grep noonsees conftest.$ac_objext >/dev/null ; then ++ ax_cv_c_float_words_bigendian=yes ++fi ++if grep seesnoon conftest.$ac_objext >/dev/null ; then ++ if test "$ax_cv_c_float_words_bigendian" = unknown; then ++ ax_cv_c_float_words_bigendian=no ++ else ++ ax_cv_c_float_words_bigendian=unknown ++ fi ++fi ++ ++])]) ++ ++case $ax_cv_c_float_words_bigendian in ++ yes) ++ m4_default([$1], ++ [AC_DEFINE([FLOAT_WORDS_BIGENDIAN], 1, ++ [Define to 1 if your system stores words within floats ++ with the most significant word first])]) ;; ++ no) ++ $2 ;; ++ *) ++ m4_default([$3], ++ [AC_MSG_ERROR([ ++ ++Unknown float word ordering. You need to manually preset ++ax_cv_c_float_words_bigendian=no (or yes) according to your system. ++ ++ ])]) ;; ++esac ++ ++])# AX_C_FLOAT_WORDS_BIGENDIAN +-- +2.11.0 +
meta/recipes-devtools/python/python3_3.5.5.bb +1 -0
@@ -41,6 +41,7 @@ SRC_URI += "\ file://pass-missing-libraries-to-Extension-for-mul.patch \ file://Use-correct-CFLAGS-for-extensions-when-cross-compili.patch \ file://0002-Makefile-add-target-to-split-profile-generation.patch \ + file://float-endian.patch \ " SRC_URI[md5sum] = "f3763edf9824d5d3a15f5f646083b6e0" SRC_URI[sha256sum] = "063d2c3b0402d6191b90731e0f735c64830e7522348aeb7ed382a83165d45009"
Signature · 1 line
2.11.0
PATCH 5/5 python3: use regrtest instead of PyBench for profile-guided-optimisation
In reply to Ross BurtonPyBench takes a long time to run, also upstream have removed it from Python and instead use test.regrtest —pgo to profile the interpreter.
The results are good: not only does Python compile faster (~300s vs ~600s on my machine) but Phoronix’s PyBench test runs in 2130ms compared to 2229ms when using PyBench to train (and 2345ms with PGO disabled).
--- meta/recipes-devtools/python/python3_3.5.5.bb | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-)
meta/recipes-devtools/python/python3_3.5.5.bb +7 -5
@@ -76,7 +76,7 @@ SDK_CC_ARCH += "-DNDEBUG -fno-inline" EXTRA_OEMAKE += "CROSS_COMPILE=yes" EXTRA_OECONF += "CROSSPYTHONPATH=${STAGING_LIBDIR_NATIVE}/python${PYTHON_MAJMIN}/lib-dynload/ --without-ensurepip" -PYTHON3_PROFILE_TASK ?= "${S}/Tools/pybench/pybench.py -n 1" +PYTHON3_PROFILE_TASK ?= "./python -m test.regrtest --pgo test_grammar test_opcodes test_dict test_builtin test_exceptions test_types test_support || true" export CROSS_COMPILE = "${TARGET_PREFIX}" export _PYTHON_PROJECT_BASE = "${B}"
@@ -144,13 +144,15 @@ do_compile() { if ${@bb.utils.contains('PACKAGECONFIG', 'pgo', 'true', 'false', d)}; then run_make profile-opt qemu_binary="${@qemu_wrapper_cmdline(d, '${STAGING_DIR_TARGET}', ['${B}', '${STAGING_DIR_TARGET}/${base_libdir}'])}" - cat > pgo-image-qemuwrapper << EOF + cat >pgo-wrapper <<EOF #!/bin/sh -set -x +cd ${B} $qemu_binary "\$@" EOF - chmod +x pgo-image-qemuwrapper - ./pgo-image-qemuwrapper ${B}/python ${PYTHON3_PROFILE_TASK} || true + chmod +x pgo-wrapper + bbnote Gathering profiling data + ./pgo-wrapper ${PYTHON3_PROFILE_TASK} + bbnote Profiling data gathered, rebuilding run_make clean_and_use_profile else run_make libpython3.so
Signature · 1 line
2.11.0