From 2ff58e62fd7b05e9092452498e4b84fb65ef0e49 Mon Sep 17 00:00:00 2001 From: DongHun Kwak Date: Thu, 11 Apr 2024 10:36:38 +0900 Subject: [PATCH] =?utf8?q?AX=5FC=5FFLOAT=5FWORDS=5FBIGENDIAN:=20Fix=20bug?= =?utf8?q?=20"AX=5FC=5FFLOAT=5FWORDS=5FBIGENDIAN=20fails=E2=80=A6=20?= =?utf8?q?=E2=80=A6=20whenever=20interprocedural=20optimization=20is=20ena?= =?utf8?q?bled."?= MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Previous versions of this macro 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 test object file. Thus, the special value (in platform-dependent binary form) could not be found, 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. 2) Add code that requires the compiler to write the special value to the data segment of the program executable. For further details, refer to the documentation in the macro source file. Note that this change already increments the serial number of the file. This change resolves the following issues: CPython issue #89640 python/cpython#89640 Gentoo bug #700012 https://bugs.gentoo.org/700012 Cairo issue #459 https://gitlab.freedesktop.org/cairo/cairo/-/issues/459 Poky (Yocto Project) #12699 https://lists.yoctoproject.org/g/poky/topic/89131836 Change-Id: I7fd6401e3dcb0ffeeb2e51c06fe8630f146d12ad --- m4/ax_c_float_words_bigendian.m4 | 61 ++++++++++++++++++++++++++++++++++++---- 1 file changed, 55 insertions(+), 6 deletions(-) diff --git a/m4/ax_c_float_words_bigendian.m4 b/m4/ax_c_float_words_bigendian.m4 index 216b90d..c66f9c8 100644 --- a/m4/ax_c_float_words_bigendian.m4 +++ b/m4/ax_c_float_words_bigendian.m4 @@ -27,32 +27,81 @@ # 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 +# Copyright (c) 2008, 2023 Daniel Amelang # # 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 13 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([[ + +#include + +static double m[] = {9.090423496703681e+223, 0.0}; -double d = 90904234967036810337470478905505011476211692735615632014797120844053488865816695273723469097858056257517020191247487429516932130503560650002327564517570778480236724525140520121371739201496540132640109977779420565776568942592.0; +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 -- 2.7.4