Illegal instruction on start

[xiota]

Segfaults on start. Building 1.20.2 and 81f1c2283 from source on Arch Linux. Prebuilt binaries work, but would still be good to figure out why this is happening.

cmake "../${pkgname}_src" \
  -DCMAKE_INSTALL_PREFIX=/usr \
  -DCACHE_NAME_SUFFIX="-git" \
  -DCMAKE_BUILD_TYPE=Release \
  -DPROC_TARGET_NUMBER=2 \
  -DWITH_LTO="ON" \
  -DENABLE_LIBRAW="ON" \
  -DENABLE_OCIO="ON" \
  -DBUILD_SHARED="ON"

make

gdb backtrace:

Program received signal SIGILL, Illegal instruction.
0x0000555555847fb5 in _sub_I_65535_0.0 ()
(gdb) bt
#0  0x0000555555847fb5 in _sub_I_65535_0.0 ()
#1  0x00007ffff5445dfe in call_init (env=<optimized out>, argv=0x7fffffffd888, argc=1) at ../csu/libc-start.c:145
#2  __libc_start_main_impl
    (main=0x555555840a30 <main(int, char**)>, argc=1, argv=0x7fffffffd888, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>, stack_end=0x7fffffffd878) at ../csu/libc-start.c:347
#3  0x00005555558495e5 in _start ()
(gdb) q

Potentially related to #288.

[agriggio]

Hi, did you try with a different compiler?

[xiota]

I don’t know how to change the compiler away from the default. Will search internet and try. Would help if you could provide instructions.

-- The C compiler identification is GNU 13.2.1
-- The CXX compiler identification is GNU 13.2.1

I do see this message in the log:

-- WARNING: gcc 13.2.1 is known to miscompile ART when using --ffp-contract=fast, forcing the option to be off

Will also try to figure out what options are being used.

[xiota]

Switched to clang by adding options to cmake:

-DCMAKE_C_COMPILER=clang
-DCMAKE_CXX_COMPILER=clang++

Linking fails:

[100%] Linking CXX executable ART
clang-16: error: unable to execute command: Segmentation fault
clang-16: error: linker command failed due to signal (use -v to see invocation)
make[2]: *** [rtgui/CMakeFiles/art.dir/build.make:2596: rtgui/ART] Error 1
make[1]: *** [CMakeFiles/Makefile2:191: rtgui/CMakeFiles/art.dir/all] Error 2
make: *** [Makefile:136: all] Error 2

‌

[xiota]

Link failure when building with clang.

[xiota]

Default compiler flags on system, before cmake is run.

CFLAGS
-march=x86-64 -mtune=generic -O2 -pipe -fno-plt -fexceptions -Wp,-D_FORTIFY_SOURCE=2 -Wformat -Werror=format-security -fstack-clash-protection -fcf-protection

CXXFLAGS
-march=x86-64 -mtune=generic -O2 -pipe -fno-plt -fexceptions -Wp,-D_FORTIFY_SOURCE=2 -Wformat -Werror=format-security -fstack-clash-protection -fcf-protection -Wp,-D_GLIBCXX_ASSERTIONS

LDFLAGS
-Wl,-O1,--sort-common,--as-needed,-z,relro,-z,now

‌

[agriggio]

Did you try disabling lto? That has always given me troubles… I would try using ` -DWITH_LTO="OFF”`

[xiota]

I didn’t realize the cmake options I was pasting had LTO enabled. With it disabled, I still get the illegal instruction, but gdb backtrace looks more useful:

Program received signal SIGILL, Illegal instruction.
0x00005555558cd26f in __static_initialization_and_destruction_0 ()
    at /usr/src/debug/art-rawconverter-git/art-rawconverter-git_src/rtgui/blackwhite.cc:49
49      };
(gdb) bt
#0  0x00005555558cd26f in __static_initialization_and_destruction_0() ()
    at /usr/src/debug/art-rawconverter-git/art-rawconverter-git_src/rtgui/blackwhite.cc:49
#1  0x00007ffff5445dfe in call_init (env=<optimized out>, argv=0x7fffffffd878, argc=1) at ../csu/libc-start.c:145
#2  __libc_start_main_impl
    (main=0x5555558ce380 <main(int, char**)>, argc=1, argv=0x7fffffffd878, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>, stack_end=0x7fffffffd868) at ../csu/libc-start.c:347
#3  0x00005555558d7865 in _start ()

‌

I just tried changing -DPROC_TARGET_NUMBER=2 to different values, and it doesn’t crash with some. The computer I’m building on has an AMD processor, but the one I’m using is Intel. So I think maybe this is the problem/solution? What number do you recommend for binaries that will be used on reasonably recent AMD and Intel processors?

[xiota]

The PROC_TARGET_NUMBER option appears to override existing -march settings in CXXFLAGS. According to x86 Options, -march produces instructions that may not run on other processors. Building without setting PROC_TARGET_NUMBER allows options already set in CXXFLAGS to be used.

Thank you for your help figuring this out.

[xiota]

Problem caused by setting PROC_TARGET_NUMBER=2, which overrides -march settings in CXXFLAGS. This generate instructions that may not work on different processors. Solution is to not set PROC_TARGET_NUMBER when building binaries that will be used with multiple processors.