Like in my previous plan, when cross-compiling we run the same
configure script separately for a native build configuration on the
build platform, in a temporary subdirectory.
Now use a fixed name "CONF-FOR-BUILD" for that subdirectory, so that
it is easy to edit out that path component from those build
environment variables that contain it.
Pass more of the native build environment variables up to the main
configure and propagate those to the build environment suffixed with
_FOR_BUILD: INPATH, OUTPATH OUTDIR, PATH, SOLARINC, SOLARLIB,
WORKDIR. Whether these all will actually be needed remains to be seen,
the set can be reduced later.
The environment setting file (*Env.Set.sh) for the native build is
copied here to the top directory under the name Env.Build.sh, and the
environment variables set in it that contain pathnames are modified to
point directly to this top directory, not the temporary CONF-FOR-BUILD
subdirectory.
When doing a cross-compiling build, we first do a build of the
necessary build-time tools for the build platform. This is done in the
same source tree. As the directories where build results are stored
include the platform specification (OUTPATH or in some cases INPATH),
there should be no clashes.
Don't run the download script from ./bootstrap(.1). We are running it
from Makefile already anyway often enough. This could also do with
some clean-up; the ./g -f clone phase is a bit slow, I am not sure if
it really is necessary every time? Also, we should not overwrite
ooo.lst if its contents isn't changing.
Use INPATH_FOR_BUILD in SOLARBINDIR so that the self-built tools like
idlc that we run are for the build platform, not the host
platform.
Attempt to get rid of the makefile.rc and makefile.mk files. Surely it
should be enough with just Makefile(.in) (and then GNUmakefile.mk for
its own so far special gbuild purposes). Instead of invoking dmake to
do "clean" or "distclean" from Makefile(.in), we already just do the
same directly in Makefile(.in). This way we don't need to first build
dmake in order to be able to do a make clean, which will then finally
clean out dmake again;)
Ideally I would like to get rid of bootstrap(.1), too. It should be
possible to merge its tasks into configure(.in) or Makefile(.in) as
appropriate.
And actually, maybe also what set_soenv(.in) does could well be merged
into configure(.in)?
Actually NSIS_PATH is crack, it is actually the *directory* where the
makensis program is located, so why not simply just have a variable
with the full pathname to that instead (or just the command name, in
case it is in the default PATH anyway, like on openSUSE with
/usr/bin/makensis)? That would simplify the stuff in download.pm, too.
Re-introduce the old --with-mingw option but now called
--with-mingw-cross-compiler. Its purpose is now specifically to give
the cross-compiler used when building the ODK, if Java is enabled, and
if building the unowinreg.dll. It has now nothing to do with
cross-compiling LibreOffice itself.
Correspondingly, the WITH_MINGW variable now has meaning only when
building LibreOffice for Windows: If using MinGW, whether natively on
Windows itself (which we as such don't intend to support, I hope), or
cross-compiling, it is set to "yes".
Automate and simplify the search for the MinGW cross-compiler when
intending to build unowinreg.dll on Unix.
Look for the usual tool-chain tools ar, nm, objdump, pkg-config,
ranlib, strip, and for Windows alto dlltool and windres using
AC_CHECK_TOOL so that the proper cross tools are found when
needed. Propagate to environment. As such these are not used except in
the MinGW mk files so far.
Other minor cleanups.
When looking for the db,h header, use Autoconf mechanisms instead of
manual checks in hardcoded directories. So yeah, this means that you
need to make sure the correct -I flag is passed if you have db
installed in a weird place where the compiler doesn't find it.
Use checks that require only compiling, not running code. Nice.
Don't AC_SUBST variables that are not used.
AC_SUBST also EXEEXT_FOR_BUILD and use that in Makefile.in.
As winemv.set.sh is now called WindowsMSVCEnv.Set.sh, with capital E
and S like all the others, we can simplify the glob pattern for the
Set.sh file.
Don't attempt to download and/or run unpackers for dependencies
relevant only when using MSVC if using MinGW.
Misc other Windows host vs. build fixes.
Still a long way from working, of course.
The configure script now runs to finish on Linux with --host=mingw32.
It is no longer an error if Windows SDK or DirectX SDK are not found
by the logic in the configure script. It might well be that the user
has included relevant -I and -L flags in CC or CXX that makes the
compilations work anyway, or something. We should not try to be too
clever and try to predict how the compiler or linker work in the
configure script.
We now define the FOO_FOR_BUILD environment variables in set_soenv.in
even when not cross-compiling (identically as the plain FOO ones in
that case, obviously). This should make some makefiles and stuff that
build tools to run on the build host a bit simpler.
In a from-scratch build, when running the configure script, which then
runs the set_soenv script, the default_images symlink (that is passed
to this function in the form of a Windows path, for some reason) does
not exist yet, and realpath fails anyway. So don't bother calling
realpath on Windows paths.
Although I don't know whether it then will cause a problem that the
cygpath -m call won't be able to expand the symlink as it doesn't
exist anyway. This is a mess. And if cygpath -m expands symlinks
anyway, why is the realpath needed at all?
