Bug 10599

The best fix is probably -fno-strict-aliasing, I don't see a need for silly tricks with union{}.

Hmm, I just tried to compile dbus-launch-x11.c with -Wstrict-aliasing, and I got no warnings:

dbus-1.2.12
$ gcc --version
gcc (Gentoo 4.3.2 p1.0) 4.3.2

I should have followed up to this bug; we use -fno-strict-aliasing by default as of the following commit.  It's still worth investigating at some point whether there are noticeable performance gains in certain parts of the code from making it strict-aliasing compatible, but for now we don't do it.

commit 98495b896dc77493db21b1847f508f64dc6ee7d3
Author: Colin Walters <walters@verbum.org>
Date:   Fri Dec 19 20:02:14 2008 -0500

    Enable -Werror by default with --enable-maintainer-mode, and change warnings
    
    Important compiler warnings were being lost in the noise from warnings
    we know about but aren't problems, and moreover made using -Werror
    difficult.  Now we expect *all* developers and testers to be using
    -Werror.

n@laptop /usr/portage/distfiles/dbus-1.2.12/tools $ i686-pc-linux-gnu-gcc -DHAVE_CONFIG_H -I. -I.. -I..   -DDBUS_LOCALEDIR=\"/usr/share/locale\" -DDBUS_COMPILATION -DDBUS_DAEMONDIR=\"/usr/bin\" -DDBUS_MACHINE_UUID_FILE=\""/var/lib/dbus/machine-id"\"    -ffunction-sections -fdata-sections -g -O2 -Wall -MT dbus-launch-x11.o -MD -MP -MF .deps/dbus-launch-x11.Tpo -c -o dbus-launch-x11.o dbus-launch-x11.c -Wextra -Wstrict-aliasing -fstrict-aliasing
dbus-launch-x11.c: In function 'x_io_error_handler':
dbus-launch-x11.c:44: warning: unused parameter 'xdisplay'
stian@laptop /usr/portage/distfiles/dbus-1.2.12/tools $ 

But I guess there might be aliasing issues in other source-files

Under clang, we also have to compile with -Wno-error=cast-align to suppress (fatal) warnings about inappropriate casts (Bug #89243).

This is a very long-standing issue: the DBusString code has always used "aliasing" that is not, strictly speaking, Standard C, and compensates for it by second-guessing the compiler and doing the alignment "by hand". I've wanted to replace this with the correct C way to do it (via a pointer to a union) for a while, but have never got round to it.

Alternatively, I might try just replacing the aliased accesses with a memcpy() into a correctly-aligned local variable, and seeing whether it makes any practical difference on benchmarks. My guess is that it won't actually produce a measurable difference.

Under gcc these warnings masked by the fact that the x86 family have essentially no mandatory alignment requirements (unaligned accesses are just slow); building for ARM, where unaligned access silently does the wrong thing, has raised these warnings for years. clang treats the multi-byte integers' preferred alignment as if it was mandatory even on x86, which is a stricter interpretation.

The union trick is not standard C either. The only correct way is memcpy.

-- GitLab Migration Automatic Message --

This bug has been migrated to freedesktop.org's GitLab instance and has been closed from further activity.

You can subscribe and participate further through the new bug through this link to our GitLab instance: https://gitlab.freedesktop.org/dbus/dbus/issues/4.