[3.13] gh-89640: harden float word ordering (#125571 and #126387) (#126429)

Properly detect float word ordering on Linux (gh-125571)

autoconf-archive patch by Dan Amelang.

(cherry picked from commit 26d627779f)

Hardcode WASM float word ordering to little endian (gh-126387)

(cherry picked from commit 532fc08102)

.github/workflows/build.yml

@@ -89,7 +89,7 @@ jobs:
     # reproducible: to get the same tools versions (autoconf, aclocal, ...)
     runs-on: ubuntu-24.04
     container:
-      image: ghcr.io/python/autoconf:2024.10.11.11293396815
+      image: ghcr.io/python/autoconf:2024.10.16.11360930377
     timeout-minutes: 60
     needs: check_source
     if: needs.check_source.outputs.run_tests == 'true'

Misc/NEWS.d/next/Build/2024-10-16-09-37-51.gh-issue-89640.UDsW-j.rst

@@ -0,0 +1,2 @@
+Improve detection of float word ordering on Linux when link-time optimizations
+are enabled.

Misc/NEWS.d/next/Build/2024-11-04-09-42-04.gh-issue-89640.QBv05o.rst

@@ -0,0 +1 @@
+Hard-code float word ordering as little endian on WASM.

Tools/build/regen-configure.sh

@@ -5,7 +5,7 @@ set -e -x
 # The check_autoconf_regen job of .github/workflows/build.yml must kept in
 # sync with this script. Use the same container image than the job so the job
 # doesn't need to run autoreconf in a container.
-IMAGE="ghcr.io/python/autoconf:2024.10.06.11200919239"
+IMAGE="ghcr.io/python/autoconf:2024.10.16.11360930377"
 AUTORECONF="autoreconf -ivf -Werror"
 
 WORK_DIR="/src"

aclocal.m4

@@ -41,32 +41,81 @@ m4_ifndef([AC_CONFIG_MACRO_DIRS], [m4_defun([_AM_CONFIG_MACRO_DIRS], [])m4_defun
 #   If neither value is found, the user is instructed to specify the
 #   ordering.
 #
+#   Early versions of this macro (i.e., before serial 12) would not work
+#   when interprocedural optimization (via link-time optimization) was
+#   enabled. This would happen when, say, the GCC/clang "-flto" flag, or the
+#   ICC "-ipo" flag was used, for example. The problem was that under
+#   these conditions, the compiler did not allocate for and write the special
+#   float value in the data segment of the object file, since doing so might
+#   not prove optimal once more context was available. Thus, the special value
+#   (in platform-dependent binary form) could not be found in the object file,
+#   and the macro would fail.
+#
+#   The solution to the above problem was to:
+#
+#     1) Compile and link a whole test program rather than just compile an
+#        object file. This ensures that we reach the point where even an
+#        interprocedural optimizing compiler writes values to the data segment.
+#
+#     2) Add code that requires the compiler to write the special value to
+#        the data segment, as opposed to "optimizing away" the variable's
+#        allocation. This could be done via compiler keywords or options, but
+#        it's tricky to make this work for all versions of all compilers with
+#        all optimization settings. The chosen solution was to make the exit
+#        code of the test program depend on the storing of the special value
+#        in memory (in the data segment). Because the exit code can be
+#        verified, any compiler that aspires to be correct will produce a
+#        program binary that contains the value, which the macro can then find.
+#
+#   How does the exit code depend on the special value residing in memory?
+#   Memory, unlike variables and registers, can be addressed indirectly at run
+#   time. The exit code of this test program is a result of indirectly reading
+#   and writing to the memory region where the special value is supposed to
+#   reside. The actual memory addresses used and the values to be written are
+#   derived from the the program input ("argv") and are therefore not known at
+#   compile or link time. The compiler has no choice but to defer the
+#   computation to run time, and to prepare by allocating and populating the
+#   data segment with the special value. For further details, refer to the
+#   source code of the test program.
+#
+#   Note that the test program is never meant to be run. It only exists to host
+#   a double float value in a given platform's binary format. Thus, error
+#   handling is not included.
+#
 # LICENSE
 #
-#   Copyright (c) 2008 Daniel Amelang <dan@amelang.net>
+#   Copyright (c) 2008, 2023 Daniel Amelang <dan@amelang.net>
 #
 #   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
+#serial 12
 
 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([[
+AC_LINK_IFELSE([AC_LANG_SOURCE([[
 
-double d = 90904234967036810337470478905505011476211692735615632014797120844053488865816695273723469097858056257517020191247487429516932130503560650002327564517570778480236724525140520121371739201496540132640109977779420565776568942592.0;
+#include <stdlib.h>
+
+static double m[] = {9.090423496703681e+223, 0.0};
+
+int main (int argc, char *argv[])
+{
+    m[atoi (argv[1])] += atof (argv[2]);
+    return m[atoi (argv[3])] > 0.0;
+}
 
 ]])], [
 
-if grep noonsees conftest.$ac_objext >/dev/null ; then
+if grep noonsees conftest$EXEEXT >/dev/null ; then
   ax_cv_c_float_words_bigendian=yes
 fi
-if grep seesnoon conftest.$ac_objext >/dev/null ; then
+if grep seesnoon conftest$EXEEXT >/dev/null ; then
   if test "$ax_cv_c_float_words_bigendian" = unknown; then
     ax_cv_c_float_words_bigendian=no
   else

configure

@@ -23703,18 +23703,26 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 
-double d = 90904234967036810337470478905505011476211692735615632014797120844053488865816695273723469097858056257517020191247487429516932130503560650002327564517570778480236724525140520121371739201496540132640109977779420565776568942592.0;
+#include <stdlib.h>
+
+static double m[] = {9.090423496703681e+223, 0.0};
+
+int main (int argc, char *argv[])
+{
+    m[atoi (argv[1])] += atof (argv[2]);
+    return m[atoi (argv[3])] > 0.0;
+}
 
 
 _ACEOF
-if ac_fn_c_try_compile "$LINENO"
+if ac_fn_c_try_link "$LINENO"
 then :
 
 
-if grep noonsees conftest.$ac_objext >/dev/null ; then
+if grep noonsees conftest$EXEEXT >/dev/null ; then
   ax_cv_c_float_words_bigendian=yes
 fi
-if grep seesnoon conftest.$ac_objext >/dev/null ; then
+if grep seesnoon conftest$EXEEXT >/dev/null ; then
   if test "$ax_cv_c_float_words_bigendian" = unknown; then
     ax_cv_c_float_words_bigendian=no
   else
@@ -23724,7 +23732,8 @@ fi
 
 
 fi
-rm -f core conftest.err conftest.$ac_objext conftest.beam conftest.$ac_ext
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
+    conftest$ac_exeext conftest.$ac_ext
 fi
 { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_c_float_words_bigendian" >&5
 printf "%s\n" "$ax_cv_c_float_words_bigendian" >&6; }
@@ -23732,41 +23741,34 @@ printf "%s\n" "$ax_cv_c_float_words_bigendian" >&6; }
 case $ax_cv_c_float_words_bigendian in
   yes)
 
-printf "%s\n" "#define FLOAT_WORDS_BIGENDIAN 1" >>confdefs.h
+printf "%s\n" "#define DOUBLE_IS_BIG_ENDIAN_IEEE754 1" >>confdefs.h
  ;;
   no)
+
+printf "%s\n" "#define DOUBLE_IS_LITTLE_ENDIAN_IEEE754 1" >>confdefs.h
  ;;
   *)
-    as_fn_error $? "
+    case $host_cpu in #(
+  *arm*) :
+    # 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.
+                     # FLOAT_WORDS_BIGENDIAN doesn't actually detect this case,
+                     # but if it's not big or little, then it must be this?
 
-Unknown float word ordering. You need to manually preset
-ax_cv_c_float_words_bigendian=no (or yes) according to your system.
+printf "%s\n" "#define DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754 1" >>confdefs.h
+ ;; #(
+  wasm*) :
 
-    " "$LINENO" 5 ;;
+printf "%s\n" "#define DOUBLE_IS_LITTLE_ENDIAN_IEEE754 1" >>confdefs.h
+ ;; #(
+  *) :
+     ;;
+esac ;;
 esac
 
 
-if test "$ax_cv_c_float_words_bigendian" = "yes"
-then
-
-printf "%s\n" "#define DOUBLE_IS_BIG_ENDIAN_IEEE754 1" >>confdefs.h
-
-elif test "$ax_cv_c_float_words_bigendian" = "no"
-then
-
-printf "%s\n" "#define DOUBLE_IS_LITTLE_ENDIAN_IEEE754 1" >>confdefs.h
-
-else
-  # 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.
-  # FLOAT_WORDS_BIGENDIAN doesn't actually detect this case, but if it's not big
-  # or little, then it must be this?
-
-printf "%s\n" "#define DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754 1" >>confdefs.h
-
-fi
 
 # The short float repr introduced in Python 3.1 requires the
 # correctly-rounded string <-> double conversion functions from

configure.ac

@@ -5856,28 +5856,26 @@ AS_VAR_IF([ac_cv_gcc_asm_for_x64], [yes], [
 # * Check for various properties of floating point *
 # **************************************************
 
-AX_C_FLOAT_WORDS_BIGENDIAN
-if test "$ax_cv_c_float_words_bigendian" = "yes"
-then
-  AC_DEFINE([DOUBLE_IS_BIG_ENDIAN_IEEE754], [1],
-  [Define if C doubles are 64-bit IEEE 754 binary format, stored
-   with the most significant byte first])
-elif test "$ax_cv_c_float_words_bigendian" = "no"
-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])
-else
-  # 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.
-  # FLOAT_WORDS_BIGENDIAN doesn't actually detect this case, but if it's not big
-  # or little, then it must be this?
+AX_C_FLOAT_WORDS_BIGENDIAN(
+  [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])],
+  [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])],
+  [AS_CASE([$host_cpu],
+           [*arm*], [# 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.
+                     # FLOAT_WORDS_BIGENDIAN doesn't actually detect this case,
+                     # but if it's not big or little, then it must be this?
                      AC_DEFINE([DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754], [1],
-  [Define if C doubles are 64-bit IEEE 754 binary format, stored
-   in ARM mixed-endian order (byte order 45670123)])
-fi
+                               [Define if C doubles are 64-bit IEEE 754 binary format,
+                                stored in ARM mixed-endian order (byte order 45670123)])],
+           [wasm*], [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])])])
 
 # The short float repr introduced in Python 3.1 requires the
 # correctly-rounded string <-> double conversion functions from

pyconfig.h.in

@@ -47,10 +47,6 @@
 /* Define if --enable-ipv6 is specified */
 #undef ENABLE_IPV6
 
-/* Define to 1 if your system stores words within floats with the most
-   significant word first */
-#undef FLOAT_WORDS_BIGENDIAN
-
 /* Define if getpgrp() must be called as getpgrp(0). */
 #undef GETPGRP_HAVE_ARG