Re: [PULL 29/86] meson: move libudev test
On Wed, 30 Sep 2020 at 21:01, Paolo Bonzini wrote:
>
> Signed-off-by: Paolo Bonzini
> ---
> configure | 14 --
> meson.build | 7 ---
> 2 files changed, 4 insertions(+), 17 deletions(-)
Hi; this commit seems to have broken my static build.
Previously configure did not include libudev in the link
for a static build (there is no libudev.a, at least on my
system). Now it does, and then the link fails with
/usr/bin/ld: cannot find -ludev
> ##
> -# Do we have libudev
> -if test "$libudev" != "no" ; then
> - if $pkg_config libudev && test "$static" != "yes"; then
> -libudev="yes"
> -libudev_libs=$($pkg_config --libs libudev)
> - else
> -libudev="no"
> - fi
> -fi
This is the old code, which doesn't enable libudev for
static builds.
> --- a/meson.build
> +++ b/meson.build
> @@ -257,8 +257,8 @@ if 'CONFIG_CURL' in config_host
> link_args: config_host['CURL_LIBS'].split())
> endif
> libudev = not_found
> -if 'CONFIG_LIBUDEV' in config_host
> - libudev = declare_dependency(link_args:
> config_host['LIBUDEV_LIBS'].split())
> +if targetos == 'linux' and (have_system or have_tools)
> + libudev = dependency('libudev', static: enable_static)
> endif
I'm not very confident about reading meson.build logic, but it
looks like this trusts meson/pkg-config to tell it about
whether it can do a static link against this library,
which doesn't work on my system, at least. (Ubuntu 18.04.4).
> brlapi = not_found
> if 'CONFIG_BRLAPI' in config_host
thanks
-- PMM
Re: [PULL 29/86] meson: move libudev test
On 01/10/20 18:19, Peter Maydell wrote: > Hi; this commit seems to have broken my static build. > Previously configure did not include libudev in the link > for a static build (there is no libudev.a, at least on my > system). Now it does, and then the link fails with > /usr/bin/ld: cannot find -ludev > >> ## >> -# Do we have libudev >> -if test "$libudev" != "no" ; then >> - if $pkg_config libudev && test "$static" != "yes"; then >> -libudev="yes" >> -libudev_libs=$($pkg_config --libs libudev) >> - else >> -libudev="no" >> - fi >> -fi > > This is the old code, which doesn't enable libudev for > static builds. [...] > I'm not very confident about reading meson.build logic, but it > looks like this trusts meson/pkg-config to tell it about > whether it can do a static link against this library, > which doesn't work on my system, at least. (Ubuntu 18.04.4). Yes, and the same was of course true of pkg-config without meson. You probably got the same warning that you reported on v7. In fact, my guess is that the "test $static != yes" was added in reply to a similar complaint; the commit that introduced it has a note that the test was added by the maintainer: commit 3efac6ebb88e4d099f07fef65178ebaa595ae770 Author: Tomáš Golembiovský Date: Tue Oct 23 13:23:10 2018 +0200 configure: add test for libudev Signed-off-by: Tomáš Golembiovský Reviewed-by: Marc-André Lureau *make libudev optional to avoid breaking existing build/test environments *disable libudev for --static builds Signed-off-by: Michael Roth But I don't think --static should affect the build this way, especially since libudev was the only library that had this test in the configure script (checked in 5.1). Debian doesn't package "libgtk-3.a" either and yet GTK+ it is not special cased in the configure script, so I'm not sure what it is that makes libudev special. Do you actually build just with "--static"? Or do you have a list of "--disable" options so you can add one more? Paolo
Re: [PULL 29/86] meson: move libudev test
On Thu, 1 Oct 2020 at 17:55, Paolo Bonzini wrote: > On 01/10/20 18:19, Peter Maydell wrote: > > I'm not very confident about reading meson.build logic, but it > > looks like this trusts meson/pkg-config to tell it about > > whether it can do a static link against this library, > > which doesn't work on my system, at least. (Ubuntu 18.04.4). > > Yes, and the same was of course true of pkg-config without meson. You > probably got the same warning that you reported on v7. > > In fact, my guess is that the "test $static != yes" was added in reply > to a similar complaint; the commit that introduced it has a note that > the test was added by the maintainer: > > commit 3efac6ebb88e4d099f07fef65178ebaa595ae770 > Author: Tomáš Golembiovský > Date: Tue Oct 23 13:23:10 2018 +0200 > > configure: add test for libudev > > Signed-off-by: Tomáš Golembiovský > Reviewed-by: Marc-André Lureau > *make libudev optional to avoid breaking existing build/test > environments > *disable libudev for --static builds > Signed-off-by: Michael Roth > > But I don't think --static should affect the build this way, especially > since libudev was the only library that had this test in the configure > script (checked in 5.1). Debian doesn't package "libgtk-3.a" either and > yet GTK+ it is not special cased in the configure script, so I'm not > sure what it is that makes libudev special. > > Do you actually build just with "--static"? Or do you have a list of > "--disable" options so you can add one more? Yes, I have a lot of --disable-foo options. Ideally I wouldn't need any, because our configure/build system would identify "there isn't actually a static version of this dependency present" rather than blindly trusting pkg-config when it lies to us. (Is it possible to get Meson to just always do a "test that a trivial program with these cflags and libs will build" as part of whatever magic it does as part of dependency() ? Or to have a qemu_dependency() wrapper that does ?) The useful thing about libgtk is that it's pretty obvious that the --disable option to use to stop QEMU linking to it is --disable-gtk. It's much less obvious what would be the --disable option to use to stop us linking against libgio, or libudev, because those dependencies aren't closely and obviously tied to QEMU features. So for that kind of library dependency it's much more useful to have configure do enough checks that it doesn't try to link against something that doesn't exist. thanks -- PMM
Re: [PULL 29/86] meson: move libudev test
On 01/10/20 20:51, Peter Maydell wrote: > Yes, I have a lot of --disable-foo options. Ideally I wouldn't > need any, because our configure/build system would identify "there > isn't actually a static version of this dependency present" > rather than blindly trusting pkg-config when it lies to us. > (Is it possible to get Meson to just always do a "test that > a trivial program with these cflags and libs will build" as > part of whatever magic it does as part of dependency() ? Unfortunately there are many special cases, including libraries that require symbols in the executable, so its not possible to do that. For this reason the "static library not found for dependency" is just a warning. We can add such a compile test ourselves. Of course it might add to the compile time if we do it for every dependency, but perhaps we can do one with all the dependencies and if it fails loop on single dependencies to give a better error. Paolo
