| Message ID | 1474642729-40702-1-git-send-email-abrodkin@synopsys.com |
|---|---|
| State | Accepted |
| Headers | show |
Hi Alexey, Excellent commit log! On 23-09-16 16:58, Alexey Brodkin wrote: > Some Qt5 sub-projects as well as 3rd-party apps written on Qt > are failing to compile with gcc 6.x like that: > ---------------------------->8------------------------- > In file included from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/bits/stl_algo.h:59:0, > from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/algorithm:62, > from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qglobal.h:88, > from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qpair.h:37, > from qmediametadata.h:37, > from qmediametadata.cpp:28: > XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/cstdlib:75:25: fatal error: stdlib.h: No such file or directory > #include_next <stdlib.h> > ^ > compilation terminated. > ---------------------------->8------------------------- > > That happens because qmake trying to play smart passes some > include paths in Makefile with "-isystem" prefix. > Which in some cases lead to build failure well described in [1]. > > A little bit more details below on what really happens: > > 1. In "configure" script Qt determines default include paths of the > toolchain and stores them in DEFAULT_INCDIRS variable, see [2]. > > 2. On qmake execution when it creates Makefile out of .pro-file > it parses headers in INCLUDEPATH variable and if a path matches > one in DEFAULT_INCDIRS then in CXXFLAGS that path is written > with $QMAKE_CFLAGS_ISYSTEM prefix, otherwise non-matching include > path ends up in CXXFLAGS with normal "-I" prefix. > > 3. By default for gcc "QMAKE_CFLAGS_ISYSTEM = -isystem", see [3]. > > 4. gcc fails to find stdlib.h, again refer to Jörg's explanation in [1]. There is one difference here, however: In Jörg's explanation, it's about headers which really aren't system headers (protobuf). But in this case, the headers really are system headers. > > What we do here we force set QMAKE_CFLAGS_ISYSTEM to "" and so qmake > won't use "-isystem" any longer instead expected "-I" will be used for > all headers, see [4]. Another option would be to set it to -idirafter, which puts the directory at the end of the default search path and I think it also still treats it as a system directory. But that would possibly lead to even worse side effects, so maybe not such a good idea. I still don't understand how the -I$(STAGING_DIR)/usr/include gets added, by the way. Isn't pkg-config supposed to suppress those, because they already are in the default search path? For instance, on my system I have a xf86dgaproto.pc which contains Cflags: -I${includedir}, but pkg-config --cflags xf86dgaproto gives empty. But in the end it will probably come down to this patch anyway, so Reviewed-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be> However, I propose to wait another week or so to see if someone else comes up with a better idea. Regards, Arnout > > That fixes building of Qt5Webkit on ARM with gcc 6.x and a number of > autobuilder failures for ARC (the an arch that uses gcc 6 by default) like: > http://autobuild.buildroot.net/results/56a/56a6700774af692e7f5a99b452b15e4e8592310f > http://autobuild.buildroot.net/results/697/697412b29bf031bf8f246cc3af97ebcbf6bf6d1b > > [1] https://git.buildroot.net/buildroot/commit/?id=e79272fa7ff3d66c18de3514b912cd9d68d121a4 > [2] http://code.qt.io/cgit/qt/qtbase.git/tree/configure?h=5.6.1#n3660 > [3] http://code.qt.io/cgit/qt/qtbase.git/tree/mkspecs/common/gcc-base.conf?h=5.6.1#n47 > [4] http://code.qt.io/cgit/qt/qtbase.git/tree/qmake/generators/unix/unixmake2.cpp?&h=5.6.1#n193 > > Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com> > Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com> > Cc: Peter Seiderer <ps.report@gmx.net> > Cc: Julien Corjon <corjon.j@ecagroup.com> > Cc: "Yann E. MORIN" <yann.morin.1998@free.fr> > --- > package/qt5/qt5base/qmake.conf | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/package/qt5/qt5base/qmake.conf b/package/qt5/qt5base/qmake.conf > index 2152d3c..49cf898 100644 > --- a/package/qt5/qt5base/qmake.conf > +++ b/package/qt5/qt5base/qmake.conf > @@ -19,6 +19,7 @@ QMAKE_CXXFLAGS_RELEASE += -O3 > CONFIG += nostrip > > QMAKE_LIBS += -lrt -lpthread -ldl > +QMAKE_CFLAGS_ISYSTEM = > > include(../common/linux_device_post.conf) > load(qt_config) >
Hi Arnout, On Sat, 2016-09-24 at 00:11 +0200, Arnout Vandecappelle wrote: > Hi Alexey, > > Excellent commit log! That a pleasure to read this kind of comments really :) > On 23-09-16 16:58, Alexey Brodkin wrote: > > > > Some Qt5 sub-projects as well as 3rd-party apps written on Qt > > are failing to compile with gcc 6.x like that: > > ---------------------------->8------------------------- > > In file included from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/bits/stl_algo.h:59:0, > > from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/algorithm:62, > > from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qglobal.h:88, > > from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qpair.h:37, > > from qmediametadata.h:37, > > from qmediametadata.cpp:28: > > XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/cstdlib:75:25: fatal error: stdlib.h: No such file > > or directory > > #include_next <stdlib.h> > > ^ > > compilation terminated. > > ---------------------------->8------------------------- > > > > That happens because qmake trying to play smart passes some > > include paths in Makefile with "-isystem" prefix. > > Which in some cases lead to build failure well described in [1]. > > > > A little bit more details below on what really happens: > > > > 1. In "configure" script Qt determines default include paths of the > > toolchain and stores them in DEFAULT_INCDIRS variable, see [2]. > > > > 2. On qmake execution when it creates Makefile out of .pro-file > > it parses headers in INCLUDEPATH variable and if a path matches > > one in DEFAULT_INCDIRS then in CXXFLAGS that path is written > > with $QMAKE_CFLAGS_ISYSTEM prefix, otherwise non-matching include > > path ends up in CXXFLAGS with normal "-I" prefix. > > > > 3. By default for gcc "QMAKE_CFLAGS_ISYSTEM = -isystem", see [3]. > > > > 4. gcc fails to find stdlib.h, again refer to Jörg's explanation in [1]. > > There is one difference here, however: In Jörg's explanation, it's about > headers which really aren't system headers (protobuf). But in this case, the > headers really are system headers. Indeed. > > > > > > What we do here we force set QMAKE_CFLAGS_ISYSTEM to "" and so qmake > > won't use "-isystem" any longer instead expected "-I" will be used for > > all headers, see [4]. > > Another option would be to set it to -idirafter, which puts the directory at > the end of the default search path and I think it also still treats it as a > system directory. But that would possibly lead to even worse side effects, so > maybe not such a good idea. > > I still don't understand how the -I$(STAGING_DIR)/usr/include gets added, by > the way. Isn't pkg-config supposed to suppress those, because they already are > in the default search path? For instance, on my system I have a xf86dgaproto.pc > which contains Cflags: -I${includedir}, but pkg-config --cflags xf86dgaproto > gives empty. That's how: --------------------->8-------------------- export PATH=/XXX/output/host/usr/bin/:$PATH /XXX/output/build/qt5webkit-b889f460280ad98c89ede179bd3b9ce9cb02002b/Tools/qmake/config.tests/icu/pkg-config_wrapper.sh --libs icu-i18n -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata --------------------->8-------------------- Looks like .pc files from "XXX/output/host/usr/lib/pkgconfig" are used. And just FYI contents of "pkg-config_wrapper.sh": --------------------->8-------------------- ccat /XXX/output/build/qt5webkit-b889f460280ad98c89ede179bd3b9ce9cb02002b/Tools/qmake/config.tests/icu/pkg- config_wrapper.sh #!/bin/sh PKG_CONFIG_SYSROOT_DIR=/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot export PKG_CONFIG_SYSROOT_DIR PKG_CONFIG_LIBDIR=/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/pkgconfig export PKG_CONFIG_LIBDIR exec pkg-config "$@" --------------------->8-------------------- -Alexey
On 26-09-16 13:30, Alexey Brodkin wrote: > Hi Arnout, > > On Sat, 2016-09-24 at 00:11 +0200, Arnout Vandecappelle wrote: [snip] >> I still don't understand how the -I$(STAGING_DIR)/usr/include gets added, by >> the way. Isn't pkg-config supposed to suppress those, because they already are >> in the default search path? For instance, on my system I have a xf86dgaproto.pc >> which contains Cflags: -I${includedir}, but pkg-config --cflags xf86dgaproto >> gives empty. I mean that it's _our_ pkg-config that does the wrong thing. ------ output/staging/usr/lib/pkgconfig/libpcre.pc -------- # Package Information for pkg-config prefix=/usr exec_prefix=/usr libdir=${exec_prefix}/lib includedir=${prefix}/include Name: libpcre Description: PCRE - Perl compatible regular expressions C library with 8 bit character support Version: 8.39 Libs: -L${libdir} -lpcre Libs.private: Cflags: -I${includedir} ------------------------------------------------------------ output/host/usr/bin/pkg-config --cflags libpcre -IXXX/output/host/usr/arm-buildroot-linux-musleabihf/sysroot/usr/include According to me, this should have returned empty. So I've checked, and it looks like this is a feature of the full-fledged pkg-config that pkgconf doesn't have: pkg-config --cflags libpcre <empty> output/host/usr/bin/pkgconf --cflags libpcre -I/usr/include So ideally, we should patch pkgconf (and send upstream) to remove the default search paths. > > That's how: > --------------------->8-------------------- > export PATH=/XXX/output/host/usr/bin/:$PATH > /XXX/output/build/qt5webkit-b889f460280ad98c89ede179bd3b9ce9cb02002b/Tools/qmake/config.tests/icu/pkg-config_wrapper.sh > --libs icu-i18n > -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata > --------------------->8-------------------- > > Looks like .pc files from "XXX/output/host/usr/lib/pkgconfig" are used. No, it's from the staging dir XXX/output/host/usr/<tuple>/sysroot/usr/lib/pkgconfig > > And just FYI contents of "pkg-config_wrapper.sh": > --------------------->8-------------------- > ccat /XXX/output/build/qt5webkit-b889f460280ad98c89ede179bd3b9ce9cb02002b/Tools/qmake/config.tests/icu/pkg- > config_wrapper.sh > #!/bin/sh > PKG_CONFIG_SYSROOT_DIR=/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot > export PKG_CONFIG_SYSROOT_DIR > PKG_CONFIG_LIBDIR=/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/pkgconfig See? Regards, Arnout > export PKG_CONFIG_LIBDIR > exec pkg-config "$@" > --------------------->8-------------------- > > -Alexey >
Hi Arnout, On Wed, 2016-09-28 at 00:33 +0200, Arnout Vandecappelle wrote: > > On 26-09-16 13:30, Alexey Brodkin wrote: > > > > Hi Arnout, > > > > On Sat, 2016-09-24 at 00:11 +0200, Arnout Vandecappelle wrote: > [snip] > > > > > > > > I still don't understand how the -I$(STAGING_DIR)/usr/include gets added, by > > > the way. Isn't pkg-config supposed to suppress those, because they already are > > > in the default search path? For instance, on my system I have a xf86dgaproto.pc > > > which contains Cflags: -I${includedir}, but pkg-config --cflags xf86dgaproto > > > gives empty. > > I mean that it's _our_ pkg-config that does the wrong thing. > > ------ output/staging/usr/lib/pkgconfig/libpcre.pc -------- > # Package Information for pkg-config > > prefix=/usr > exec_prefix=/usr > libdir=${exec_prefix}/lib > includedir=${prefix}/include > > Name: libpcre > Description: PCRE - Perl compatible regular expressions C library with 8 bit > character support > Version: 8.39 > Libs: -L${libdir} -lpcre > Libs.private: > Cflags: -I${includedir} > ------------------------------------------------------------ > > output/host/usr/bin/pkg-config --cflags libpcre > -IXXX/output/host/usr/arm-buildroot-linux-musleabihf/sysroot/usr/include > > According to me, this should have returned empty. > > So I've checked, and it looks like this is a feature of the full-fledged > pkg-config that pkgconf doesn't have: > > pkg-config --cflags libpcre > <empty> > > output/host/usr/bin/pkgconf --cflags libpcre > -I/usr/include Well note in our case pkg-config is just a wrapper script that is created out of BR's "package/pkgconf/pkg-config.in": --------------------->8-------------------- cat ./output/host/usr/bin/pkg-config #!/bin/sh PKG_CONFIG_LIBDIR=${PKG_CONFIG_LIBDIR:-/XXX/output/host/usr/arc-buildroot-linux- uclibc/sysroot/usr/lib/pkgconfig:/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/share/pkgconfig} PKG_CONFIG_SYSROOT_DIR=${PKG_CONFIG_SYSROOT_DIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot} $(dirname $0)/pkgconf $@ --------------------->8-------------------- So it all boils down to: 1) "pkg-config" case: --------------------->8-------------------- PKG_CONFIG_LIBDIR=${PKG_CONFIG_LIBDIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/pkgconfig} PKG_CONFIG_SYSROOT_DIR=${PKG_CONFIG_SYSROOT_DIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot} ./output/host/usr/bin/pkgconf --libs icu-i18n -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata --------------------->8-------------------- 2. Pure "pkgconf" case: --------------------->8-------------------- ./output/host/usr/bin/pkgconf --libs icu-i18n -licui18n -licuuc -licudata --------------------->8-------------------- > So ideally, we should patch pkgconf (and send upstream) to remove the default > search paths. Given my comments above I'm not really sure it's pkgconf who's guilty. Any thoughts? -Alexey
Hello, On Wed, 28 Sep 2016 00:33:16 +0200, Arnout Vandecappelle wrote: > I mean that it's _our_ pkg-config that does the wrong thing. > > ------ output/staging/usr/lib/pkgconfig/libpcre.pc -------- > # Package Information for pkg-config > > prefix=/usr > exec_prefix=/usr > libdir=${exec_prefix}/lib > includedir=${prefix}/include > > Name: libpcre > Description: PCRE - Perl compatible regular expressions C library with 8 bit > character support > Version: 8.39 > Libs: -L${libdir} -lpcre > Libs.private: > Cflags: -I${includedir} > ------------------------------------------------------------ > > output/host/usr/bin/pkg-config --cflags libpcre > -IXXX/output/host/usr/arm-buildroot-linux-musleabihf/sysroot/usr/include > > According to me, this should have returned empty. So in other words, you're withdrawing your Reviewed-by on Alexey's patch, and say that we should fix host-pkgconf instead, right ? Thomas
On 28-09-16 10:54, Alexey Brodkin wrote: > So it all boils down to: > 1) "pkg-config" case: > --------------------->8-------------------- > PKG_CONFIG_LIBDIR=${PKG_CONFIG_LIBDIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/pkgconfig} > PKG_CONFIG_SYSROOT_DIR=${PKG_CONFIG_SYSROOT_DIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot} > ./output/host/usr/bin/pkgconf --libs icu-i18n > -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata > --------------------->8-------------------- > > 2. Pure "pkgconf" case: > --------------------->8-------------------- > ./output/host/usr/bin/pkgconf --libs icu-i18n > -licui18n -licuuc -licudata > --------------------->8-------------------- Your pure pkgconf case is using the host icu-i18n, that's why the -L flag is missing. The host icu-i18n doesn't have a -L bit in its Libs: stanza. You should compare: 1) pkgconf case PKG_CONFIG_LIBDIR=$STAGING_DIR/usr/lib/pkgconfig PKG_CONFIG_SYSROOT_DIR=$STAGING_DIR/sysroot output/host/usr/bin/pkgconf --libs icu-i18n -licui18n -L$STAGING_DIR/usr/lib -licuuc -licudata 2) Real pkg-config case: PKG_CONFIG_LIBDIR=$STAGING_DIR/usr/lib/pkgconfig PKG_CONFIG_SYSROOT_DIR=$STAGING_DIR/sysroot /usr/bin/pkg-config --libs icu-i18n -licui18n -licuuc -licudata We switched from pkg-config to pkgconf because it builds a whole lot faster (pkg-config has a whole lot of dependencies). Obviously we give up features with that, but this is the first one where I think it is an important feature that we're missing. Regards, Arnout > >> > So ideally, we should patch pkgconf (and send upstream) to remove the default >> > search paths. > Given my comments above I'm not really sure it's pkgconf who's guilty.
On 28-09-16 10:57, Thomas Petazzoni wrote: > Hello, > > On Wed, 28 Sep 2016 00:33:16 +0200, Arnout Vandecappelle wrote: > >> I mean that it's _our_ pkg-config that does the wrong thing. >> >> ------ output/staging/usr/lib/pkgconfig/libpcre.pc -------- >> # Package Information for pkg-config >> >> prefix=/usr >> exec_prefix=/usr >> libdir=${exec_prefix}/lib >> includedir=${prefix}/include >> >> Name: libpcre >> Description: PCRE - Perl compatible regular expressions C library with 8 bit >> character support >> Version: 8.39 >> Libs: -L${libdir} -lpcre >> Libs.private: >> Cflags: -I${includedir} >> ------------------------------------------------------------ >> >> output/host/usr/bin/pkg-config --cflags libpcre >> -IXXX/output/host/usr/arm-buildroot-linux-musleabihf/sysroot/usr/include >> >> According to me, this should have returned empty. > > So in other words, you're withdrawing your Reviewed-by on Alexey's > patch, and say that we should fix host-pkgconf instead, right ? > > Thomas Au contraire. Ideally host-pkgconf should be improved, but that's not a simple task. Until then, Alexey's patch will work. Regards, Arnout
On 29-09-16 23:05, Arnout Vandecappelle wrote: > > > On 28-09-16 10:54, Alexey Brodkin wrote: >> So it all boils down to: >> 1) "pkg-config" case: >> --------------------->8-------------------- >> PKG_CONFIG_LIBDIR=${PKG_CONFIG_LIBDIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/pkgconfig} >> PKG_CONFIG_SYSROOT_DIR=${PKG_CONFIG_SYSROOT_DIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot} >> ./output/host/usr/bin/pkgconf --libs icu-i18n >> -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata >> --------------------->8-------------------- >> >> 2. Pure "pkgconf" case: >> --------------------->8-------------------- >> ./output/host/usr/bin/pkgconf --libs icu-i18n >> -licui18n -licuuc -licudata >> --------------------->8-------------------- > > Your pure pkgconf case is using the host icu-i18n, that's why the -L flag is > missing. The host icu-i18n doesn't have a -L bit in its Libs: stanza. Hm, actually, icu-uc.pc *does* have a -L bit in its Libs: stanza. But for the host, I guess it recognizes that ${libdir} == /usr/lib so it can be eliminated. So perhaps that can be extended to cover a sysroot as well. BTW, I did a quick test with pkgconf 1.0.1 and it still behaves the same. Regards, Arnout > > You should compare: > > 1) pkgconf case > PKG_CONFIG_LIBDIR=$STAGING_DIR/usr/lib/pkgconfig > PKG_CONFIG_SYSROOT_DIR=$STAGING_DIR/sysroot output/host/usr/bin/pkgconf --libs > icu-i18n > -licui18n -L$STAGING_DIR/usr/lib -licuuc -licudata > > 2) Real pkg-config case: > PKG_CONFIG_LIBDIR=$STAGING_DIR/usr/lib/pkgconfig > PKG_CONFIG_SYSROOT_DIR=$STAGING_DIR/sysroot /usr/bin/pkg-config --libs icu-i18n > -licui18n -licuuc -licudata > > We switched from pkg-config to pkgconf because it builds a whole lot faster > (pkg-config has a whole lot of dependencies). Obviously we give up features with > that, but this is the first one where I think it is an important feature that > we're missing. > > Regards, > Arnout > >> >>>> So ideally, we should patch pkgconf (and send upstream) to remove the default >>>> search paths. >> Given my comments above I'm not really sure it's pkgconf who's guilty. >
Hi Arnout, On Thu, 2016-09-29 at 23:05 +0200, Arnout Vandecappelle wrote: > > On 28-09-16 10:54, Alexey Brodkin wrote: > > > > So it all boils down to: > > 1) "pkg-config" case: > > --------------------->8-------------------- > > PKG_CONFIG_LIBDIR=${PKG_CONFIG_LIBDIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/pkgconfig} > > PKG_CONFIG_SYSROOT_DIR=${PKG_CONFIG_SYSROOT_DIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot} > > ./output/host/usr/bin/pkgconf --libs icu-i18n > > -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata > > --------------------->8-------------------- > > > > 2. Pure "pkgconf" case: > > --------------------->8-------------------- > > ./output/host/usr/bin/pkgconf --libs icu-i18n > > -licui18n -licuuc -licudata > > --------------------->8-------------------- > > Your pure pkgconf case is using the host icu-i18n, that's why the -L flag is > missing. The host icu-i18n doesn't have a -L bit in its Libs: stanza. > > You should compare: > > 1) pkgconf case > PKG_CONFIG_LIBDIR=$STAGING_DIR/usr/lib/pkgconfig > PKG_CONFIG_SYSROOT_DIR=$STAGING_DIR/sysroot output/host/usr/bin/pkgconf --libs > icu-i18n > -licui18n -L$STAGING_DIR/usr/lib -licuuc -licudata > > 2) Real pkg-config case: > PKG_CONFIG_LIBDIR=$STAGING_DIR/usr/lib/pkgconfig > PKG_CONFIG_SYSROOT_DIR=$STAGING_DIR/sysroot /usr/bin/pkg-config --libs icu-i18n > -licui18n -licuuc -licudata > > We switched from pkg-config to pkgconf because it builds a whole lot faster > (pkg-config has a whole lot of dependencies). Obviously we give up features with > that, but this is the first one where I think it is an important feature that > we're missing. Thanks for explanation that definitely makes sense. -Alexey
Hi Arnout, On Thu, 2016-09-29 at 23:34 +0200, Arnout Vandecappelle wrote: > > On 29-09-16 23:05, Arnout Vandecappelle wrote: > > > > > > > > On 28-09-16 10:54, Alexey Brodkin wrote: > > > > > > So it all boils down to: > > > 1) "pkg-config" case: > > > --------------------->8-------------------- > > > PKG_CONFIG_LIBDIR=${PKG_CONFIG_LIBDIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/pkgconfig} > > > PKG_CONFIG_SYSROOT_DIR=${PKG_CONFIG_SYSROOT_DIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot} > > > ./output/host/usr/bin/pkgconf --libs icu-i18n > > > -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata > > > --------------------->8-------------------- > > > > > > 2. Pure "pkgconf" case: > > > --------------------->8-------------------- > > > ./output/host/usr/bin/pkgconf --libs icu-i18n > > > -licui18n -licuuc -licudata > > > --------------------->8-------------------- > > > > Your pure pkgconf case is using the host icu-i18n, that's why the -L flag is > > missing. The host icu-i18n doesn't have a -L bit in its Libs: stanza. > > Hm, actually, icu-uc.pc *does* have a -L bit in its Libs: stanza. But for the > host, I guess it recognizes that ${libdir} == /usr/lib so it can be eliminated. > So perhaps that can be extended to cover a sysroot as well. > > BTW, I did a quick test with pkgconf 1.0.1 and it still behaves the same. Shall we then file an issue here https://github.com/pkgconf/pkgconf/issues ? -Alexey
Hi Thomas, Arnout, On Fri, 2016-09-30 at 11:55 +0300, Alexey Brodkin wrote: > Hi Arnout, > > On Thu, 2016-09-29 at 23:34 +0200, Arnout Vandecappelle wrote: > > > > > > On 29-09-16 23:05, Arnout Vandecappelle wrote: > > > > > > > > > > > > > > > On 28-09-16 10:54, Alexey Brodkin wrote: > > > > > > > > > > > > So it all boils down to: > > > > 1) "pkg-config" case: > > > > --------------------->8-------------------- > > > > PKG_CONFIG_LIBDIR=${PKG_CONFIG_LIBDIR:-/XXX/output/host/usr/arc-buildroot-linux- > > > > uclibc/sysroot/usr/lib/pkgconfig} > > > > PKG_CONFIG_SYSROOT_DIR=${PKG_CONFIG_SYSROOT_DIR:-/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot} > > > > ./output/host/usr/bin/pkgconf --libs icu-i18n > > > > -licui18n -L/XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib -licuuc -licudata > > > > --------------------->8-------------------- > > > > > > > > 2. Pure "pkgconf" case: > > > > --------------------->8-------------------- > > > > ./output/host/usr/bin/pkgconf --libs icu-i18n > > > > -licui18n -licuuc -licudata > > > > --------------------->8-------------------- > > > > > > Your pure pkgconf case is using the host icu-i18n, that's why the -L flag is > > > missing. The host icu-i18n doesn't have a -L bit in its Libs: stanza. > > > > Hm, actually, icu-uc.pc *does* have a -L bit in its Libs: stanza. But for the > > host, I guess it recognizes that ${libdir} == /usr/lib so it can be eliminated. > > So perhaps that can be extended to cover a sysroot as well. > > > > BTW, I did a quick test with pkgconf 1.0.1 and it still behaves the same. > > Shall we then file an issue here https://github.com/pkgconf/pkgconf/issues ? I'm wondering if there're any other thoughts on this one? We're seeing more and more instances of that problem, for example with qt5serialport here http://autobuild.buildroot.net/results/b14/b143bc806d45a91d52c43fa49227e88dc8a105d5/ So if there're no objections we may apply this one and see if it breaks anything while we're far enough from the next BR release and in the worst case we'll just revert it further down the line. -Alexey
Hello, On Fri, 23 Sep 2016 17:58:49 +0300, Alexey Brodkin wrote: > Some Qt5 sub-projects as well as 3rd-party apps written on Qt > are failing to compile with gcc 6.x like that: > ---------------------------->8------------------------- > In file included from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/bits/stl_algo.h:59:0, > from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/algorithm:62, > from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qglobal.h:88, > from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qpair.h:37, > from qmediametadata.h:37, > from qmediametadata.cpp:28: > XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/cstdlib:75:25: fatal error: stdlib.h: No such file or directory > #include_next <stdlib.h> > ^ > compilation terminated. > ---------------------------->8------------------------- Applied to master, thanks a lot! Thomas
diff --git a/package/qt5/qt5base/qmake.conf b/package/qt5/qt5base/qmake.conf index 2152d3c..49cf898 100644 --- a/package/qt5/qt5base/qmake.conf +++ b/package/qt5/qt5base/qmake.conf @@ -19,6 +19,7 @@ QMAKE_CXXFLAGS_RELEASE += -O3 CONFIG += nostrip QMAKE_LIBS += -lrt -lpthread -ldl +QMAKE_CFLAGS_ISYSTEM = include(../common/linux_device_post.conf) load(qt_config)
Some Qt5 sub-projects as well as 3rd-party apps written on Qt are failing to compile with gcc 6.x like that: ---------------------------->8------------------------- In file included from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/bits/stl_algo.h:59:0, from XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/algorithm:62, from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qglobal.h:88, from XXX/output/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/include/qt5/QtCore/qpair.h:37, from qmediametadata.h:37, from qmediametadata.cpp:28: XXX/output/host/usr/arc-buildroot-linux-uclibc/include/c++/6.2.1/cstdlib:75:25: fatal error: stdlib.h: No such file or directory #include_next <stdlib.h> ^ compilation terminated. ---------------------------->8------------------------- That happens because qmake trying to play smart passes some include paths in Makefile with "-isystem" prefix. Which in some cases lead to build failure well described in [1]. A little bit more details below on what really happens: 1. In "configure" script Qt determines default include paths of the toolchain and stores them in DEFAULT_INCDIRS variable, see [2]. 2. On qmake execution when it creates Makefile out of .pro-file it parses headers in INCLUDEPATH variable and if a path matches one in DEFAULT_INCDIRS then in CXXFLAGS that path is written with $QMAKE_CFLAGS_ISYSTEM prefix, otherwise non-matching include path ends up in CXXFLAGS with normal "-I" prefix. 3. By default for gcc "QMAKE_CFLAGS_ISYSTEM = -isystem", see [3]. 4. gcc fails to find stdlib.h, again refer to Jörg's explanation in [1]. What we do here we force set QMAKE_CFLAGS_ISYSTEM to "" and so qmake won't use "-isystem" any longer instead expected "-I" will be used for all headers, see [4]. That fixes building of Qt5Webkit on ARM with gcc 6.x and a number of autobuilder failures for ARC (the an arch that uses gcc 6 by default) like: http://autobuild.buildroot.net/results/56a/56a6700774af692e7f5a99b452b15e4e8592310f http://autobuild.buildroot.net/results/697/697412b29bf031bf8f246cc3af97ebcbf6bf6d1b [1] https://git.buildroot.net/buildroot/commit/?id=e79272fa7ff3d66c18de3514b912cd9d68d121a4 [2] http://code.qt.io/cgit/qt/qtbase.git/tree/configure?h=5.6.1#n3660 [3] http://code.qt.io/cgit/qt/qtbase.git/tree/mkspecs/common/gcc-base.conf?h=5.6.1#n47 [4] http://code.qt.io/cgit/qt/qtbase.git/tree/qmake/generators/unix/unixmake2.cpp?&h=5.6.1#n193 Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com> Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com> Cc: Peter Seiderer <ps.report@gmx.net> Cc: Julien Corjon <corjon.j@ecagroup.com> Cc: "Yann E. MORIN" <yann.morin.1998@free.fr> --- package/qt5/qt5base/qmake.conf | 1 + 1 file changed, 1 insertion(+)