V2 [PATCH] PKG_CHECK_MODULES: Check if $pkg_cv_[]$1[]_LIBS works

[Simon Marchi]

On 2020-07-28 6:45 a.m., H.J. Lu wrote:
> On Mon, Jul 27, 2020 at 12:32 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>>
>> On Mon, Jul 27, 2020 at 12:14 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>>>
>>> On Mon, Jul 27, 2020 at 9:11 AM Aaron Merey <amerey@redhat.com> wrote:
>>>>
>>>> On Mon, Jul 27, 2020 at 11:32 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>>
>>>>> On Sat, Jul 25, 2020 at 9:01 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>>> This caused:
>>>>>>
>>>>>> https://sourceware.org/bugzilla/show_bug.cgi?id=26301
>>>>>>
>>>>>
>>>>> It is quite normal to have debuginfod headers without libdebuginfod on
>>>>> multilib OSes.  Restore AC_CHECK_LIB to check if libdebuginfod exists.
>>>>> And always define HAVE_LIBDEBUGINFOD to 0 or 1 for
>>>>>
>>>>> binutils/dwarf.c:#if HAVE_LIBDEBUGINFOD
>>>>> binutils/dwarf.c:#if HAVE_LIBDEBUGINFOD
>>>>> binutils/dwarf.c:#if HAVE_LIBDEBUGINFOD
>>>>> binutils/dwarf.h:#if HAVE_LIBDEBUGINFOD
>>>>> binutils/objdump.c:#if HAVE_LIBDEBUGINFOD
>>>>> binutils/objdump.c:#endif /* HAVE_LIBDEBUGINFOD */
>>>>> binutils/readelf.c:#if HAVE_LIBDEBUGINFOD
>>>>> binutils/readelf.c:#endif /* HAVE_LIBDEBUGINFOD */
>>>>> gdb/top.c:#if HAVE_LIBDEBUGINFOD
>>>>>
>>>>> OK for master?
>>>>
>>>> Thanks for spotting this. Normally PKG_CHECH_MODULES would correctly
>>>> detect whether the .so and header are installed and build accordingly,
>>>> but when cross compiling the AC_CHECK_LIB may be needed.
>>>
>>> I am not cross compiling.  I am simply using "gcc -m32".   The problem
>>> is PKG_CHECK_MODULES which doesn't check if $pkg_cv_[]$1[]_LIBS
>>> actually works.   Here is the updated patch to fix PKG_CHECK_MODULES.
>>> Any comments or objections?
>>>
>>>
>>
>> HAVE_LIBDEBUGINFOD is a separate issue.  Here is the updated patch
>> which only adds AC_TRY_LINK to PKG_CHECK_MODULES to check if
>> $pkg_cv_[]$1[]_LIBS works.
>>
> 
> I am checking it in.
> 
> -- 
> H.J.
> 

You said that you are not cross-compiling, but technically I'd say you are cross compiling, since
you are building for a different architecture than what the compiler is running.  You are probably
configuring with --host=i686-something-something?

Anyway regardless of vocabulary, I don't think there was a problem to begin with (not that I blame
you, it's not made in an intuitive way).  The problem is that you were using pkg-config as
configured to look up x86_64 packages.  It looks up .pc files in (amongst others)
/usr/lib64/pkgconfig, which provides information about x86_64 packages, which are in turn obviously
not suitable not suitable to build a i686 program.  Just like you cross-compile "for real" (say,
for an ARM host), you need to set PKG_CONFIG or the PKG_CONFIG_* variables to returns packages for
the --host architecture.  That means searching in /usr/lib/pkgconfig instead of /usr/lib64/pkgconfig.

You could for example set the PKG_CONFIG_LIBDIR variable to /usr/lib/pkgconfig:/usr/share/pkgconfig

This way, if you don't install the elfutils-debuginfod-client-devel.i686 package, your binutils won't
try to link with libdebuginfod (because pkg-config won't find it).  If you install it, then your
binutils will be built against the i686 libdebuginfod.

Ideally, distros would ship a i686-something-something-pkg-config that automatically searchs in paths
that make sense for that architecture (just like you have arm-linux-gnueabihf-pkg-config when cross
compiling for ARM), but that doesn't seem to exist.  But this is just like you have to explicitly set
CC="gcc -m32" instead of using some i686-something-something-gcc.

You can always make it yourself, create, say, a `i686-pc-linux-gnu-pkg-config` file somewhere in $PATH,
with:

  export PKG_CONFIG_LIBDIR=/usr/lib/pkgconfig:/usr/share/pkgconfig
  exec pkg-config $*

Then, when you configure with --host=i686-pc-linux-gnu, AC_PATH_TOOL will automatically pick up that as
the pkg-config to use, and everything will work seamlessly.

So, I concede that it's not intuitive, but I think your patch is not right because it just hides the
mis-configuration.  If `pkg-config` says a lib exists but we are not able to link with it, there is a
bigger problem than "lib not found".  I think it should be a hard error (abort configure) and tell the
user about it: "pkg-config says that libfoo is available but we can't link with it, are you maybe using
the wrong pkg-config, or a wrong pkg-config path?".

Finally, the file you modified is maintained upstream here:

  https://cgit.freedesktop.org/pkg-config/tree/pkg.m4.in

Do you intend to submit your changes there?  Otherwise, they will be overwritten next time we sync with
upstream.

Simon