Travis shows that clang (but not gcc) fails to build after our recent gnulib updates, due to: CC domtop/domtop.o In file included from dommigrate/dommigrate.c:26: In file included from ../gnulib/lib/stdlib.h:100: In file included from ../gnulib/lib/unistd.h:40: In file included from /usr/include/unistd.h:638: In file included from ../gnulib/lib/sys/select.h:110: In file included from ../gnulib/lib/signal.h:67: ../gnulib/lib/pthread.h:74:3: error: "Please include config.h first." #error "Please include config.h first." This probably stems from gcc and clang having a subtle difference in -isystem vs. -I include behaviors on #include_next, where that difference then results in clang trying to pick up same-named gnulib headers merely because their directory was in the search path, while gcc did not hit that problem. But for our examples that are intended to be standalone, we want to work with the bare-bones libc, without any interference from our <config.h> or from gnulib. The easiest way to do that is by using per-binary CPPFLAGS, and including gnulib headers only for example binaries that still use <config.h>. Signed-off-by: Eric Blake <eblake(a)redhat.com&gt; --- I don't have a local clang setup, so I'm relying on CI testing to prove whether this fixes the build failures that Travis reported... --- examples/Makefile.am | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/examples/Makefile.am b/examples/Makefile.am index 61cf2af4a5..d85e1d5ace 100644 --- a/examples/Makefile.am +++ b/examples/Makefile.am @@ -34,7 +34,8 @@ EXTRA_DIST = \ AM_CPPFLAGS = \ - -I$(top_builddir)/include -I$(top_srcdir)/include -I$(top_srcdir) \ + -I$(top_builddir)/include -I$(top_srcdir)/include -I$(top_srcdir) +GNULIB_CPPFLAGS = \ -I$(top_builddir)/gnulib/lib -I$(top_srcdir)/gnulib/lib LDADD = $(STATIC_BINARIES) $(WARN_CFLAGS) $(COVERAGE_LDFLAGS) \ $(top_builddir)/src/libvirt.la $(top_builddir)/gnulib/lib/libgnu.la \ @@ -49,13 +50,16 @@ noinst_PROGRAMS=dominfo/info1 dommigrate/dommigrate domsuspend/suspend \ dominfo_info1_SOURCES = dominfo/info1.c dommigrate_dommigrate_SOURCES = dommigrate/dommigrate.c domsuspend_suspend_SOURCES = domsuspend/suspend.c +domsuspend_suspend_CPPFLAGS = $(AM_CPPFLAGS) $(GNULIB_CPPFLAGS) domtop_domtop_SOURCES = domtop/domtop.c +domtop_domtop_CPPFLAGS = $(AM_CPPFLAGS) $(GNULIB_CPPFLAGS) hellolibvirt_hellolibvirt_SOURCES = hellolibvirt/hellolibvirt.c object_events_event_test_CFLAGS = \ $(WARN_CFLAGS) \ $(NULL) object_events_event_test_SOURCES = object-events/event-test.c +object_events_event_test_CPPFLAGS = $(AM_CPPFLAGS) $(GNULIB_CPPFLAGS) openauth_openauth_SOURCES = openauth/openauth.c rename_rename_SOURCES = rename/rename.c @@ -67,6 +71,7 @@ admin_client_limits_SOURCES = admin/client_limits.c admin_client_info_SOURCES = admin/client_info.c admin_client_close_SOURCES = admin/client_close.c admin_logging_SOURCES = admin/logging.c +admin_logging_CPPFLAGS = $(AM_CPPFLAGS) $(GNULIB_CPPFLAGS) INSTALL_DATA_LOCAL = UNINSTALL_LOCAL = -- 2.20.1

On 1/7/19 5:35 AM, Ján Tomko wrote: Aha, so it wasn't a compiler difference, but a libc difference. I can update the commit message accordingly. Indeed, now that I see your thread, I still like this solution better (the examples really are meant to be standalone, so let's get gnulib out of the way rather than polluting the files with an otherwise useless config.h include). I'm fine with reverting your reversion in order to with this approach, once we solve the remaining issues... Rather, the point is that if you -Ignulib, you must also -lgnulib. If your examples are intended to be standalone (no -lgnulib), then you do NOT want to compile with gnulib headers OR with config.h. Yes, I need to fix the wording in that paragraph to point out that the real culprit is that a recent gnulib change caused BSD libc headers to cycle through a different inclusion order than previously (and not a compiler difference), while still emphasizing that the real solution is not to include config.h first, but to get rid of the gnulib headers. _If_ config.h is needed, then it is ALWAYS needed first (before ANY system headers), whether or not we also link against gnulib. But either logging.c links against gnulib (in which case your suggestion of hoisting config.h first is right), or it doesn't (in which case we should get rid of the config.h include altogether, to prove that the example is standalone; as well as tweak my patch to remove the GNULIB_CPPFLAGS addition to admin_logging_CPPFLAGS). I'll post a v2 along those lines. Yes, I'll check ALL of the examples in my v2. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org