Clarify %_build*, %_host* and %_target* macros

[Vogtinator]

I stumbled across an issue in relation to the %_host* and %_target* macros which subtly broke some builds because of armv7l/armv7hl differences. This led me to dig into this a bit further and at this point I'm now unsure how those macros are actually meant to be used.

macros.in says:

#==============================================================================
# ---- config.guess platform macros.
#       Macro(s) similar to the tokens used by configure.
#
%_build                 %{_host}
%_build_alias           %{_host_alias}
%_build_cpu             %{_host_cpu}
%_build_vendor          %{_host_vendor}
%_build_os              %{_host_os}
%_host                  @host@
%_host_alias            @host_alias@%{nil}
%_host_cpu              @host_cpu@
%_host_vendor           @host_vendor@
%_host_os               @host_os@
%_target                %{_host}
%_target_alias          %{_host_alias}
%_target_cpu            %{_host_cpu}
%_target_vendor         %{_host_vendor}
%_target_os             %{_host_os}

The @host*@ placeholders are set to whatever CMake runs on (e.g. CMAKE_HOST_SYSTEM_PROCESSOR), i.e. on what platform RPM was built on. I think this is already wrong. In the (currently unlikely) case that RPM was cross compiled, e.g. on an x86_64 machine to run on riscv64, @host_cpu@ and thus %_host_cpu would be x86_64. What machine RPM was built on should not have any effect on RPM though.

And then I'm confused about the role of build/host/target here. From my understanding, having all three of them implies the toolchain semantics:

build -> where the build happens
host -> where the built binary runs on
target -> what the built binary (e.g. compiler) targets

Question remains from which PoV this is. From RPM's PoV or the .spec file's PoV?

The rpmbuild --target option changes which architecture the built RPMs can be installed on and sets %_target* accordingly. This points to RPM's PoV. The %configure macro however uses %{_configure} --host=%{_host} --build=%{_build}, which means it's from the .spec file's PoV. This doesn't match. Example:

> cat test.spec 
Name: test
Version: 0
Release: 0
License: WTFPL
Summary: Test
%description

%build
echo %_target
set +e
%configure
touch l

%files
%license l
> rpmbuild -bb --target=i686-unknown-linux test.spec 
Building target platforms: i686-unknown-linux
Building for target i686-unknown-linux
Executing(%build): /usr/bin/bash -e /var/tmp/rpm-tmp.AGeJEj
[...]
+ echo i686-linux
i686-linux
+ set +e
+ CONFIG_SHELL=/usr/bin/bash
+ export CONFIG_SHELL
+ CFLAGS='-O2 -g -m32 -march=i686 -mtune=i686 -fmessage-length=0 -D_FORTIFY_SOURCE=2 -fstack-protector -funwind-tables -fasynchronous-unwind-tables'
+ export CFLAGS
[...]
+ ./configure --host=x86_64-suse-linux --build=x86_64-suse-linux --program-prefix= --disable-dependency-tracking --prefix=/usr --exec-prefix=/usr --bindir=/usr/bin --sbindir=/usr/sbin --sysconfdir=/etc --datadir=/usr/share --includedir=/usr/include --libdir=/usr/lib --libexecdir=/usr/libexec --localstatedir=/var --sharedstatedir=/var/lib --mandir=/usr/share/man --infodir=/usr/share/info
/var/tmp/rpm-tmp.AGeJEj: line 46: ./configure: No such file or directory
+ touch l
+ RPM_EC=0
++ jobs -p
+ exit 0
Processing files: test-0-0.i686
[...]
Provides: test = 0-0 test(x86-32) = 0-0
Requires(rpmlib): rpmlib(CompressedFileNames) <= 3.0.4-1 rpmlib(FileDigests) <= 4.6.0-1 rpmlib(PayloadFilesHavePrefix) <= 4.0-1
Checking for unpackaged file(s): /usr/lib/rpm/check-files /home/fvogt/rpmbuild/BUILDROOT/test-0-0.i386
Wrote: /home/fvogt/rpmbuild/RPMS/i686/test-0-0.i686.rpm
[...]

As you can see, configure was invoked to create a binary to run on x86_64-suse-linux but it spat out an RPM for i686: test-0-0.i686.rpm.

[pmatilai]

Nobody knows. See #1650

[Vogtinator]

Thanks, so I'm not alone.

I do wonder how it was initially intended to work, either the --target option is wrong or the %configure macro is...

[pmatilai]

AFAICS --target was always wrong from the autoconf cross-compilation terminology point of view. For what it does in rpm, --host would be closer to the mark. But outside a autoconf terminology explanation, would anybody ever guess that --host somehow relates to output architecture and stuff?