office-gobmx/README.cross

432 lines
17 KiB
Text
Raw Normal View History

2011-05-27 10:29:43 -05:00
Cross-compiling LibreOffice
===========================
Cross-compilation work has been done, to various degrees, for four
platforms: Windows, iOS, Android and PowerPC Mac OS X.
2011-05-27 10:29:43 -05:00
General
-------
In GNU Autoconf terminology, "build" is the platform on which you are
running a build of some software and "host" is the platform on which
2011-05-27 10:29:43 -05:00
the software you are building will run. Only in the specific case of
building compilers and other programming tools is the term "target"
used to indicate the platform for which the tools your are building
will produce code. As LibreOffice is not a compiler, the "target" term
should not be used in the context of cross-compilation.
(For a case where all three of "build", "host" and "target" are
different: consider a gcc cross-compiler running on Windows, producing
code for Android, where the cross-compiler itself was built on
Linux. (This is a real case.) An interesting tidbit is that such
configurations are called "Canadian Cross".)
Even though the LibreOffice build mechanism is highly unorthodox, the
configure script takes the normal --build and --host options like any
GNU Autoconf -based configure script. To cross-compile, you basically
need just to specify a suitable --host option and things should work
out nicely. In practise, many details needed to be handled. See
2011-05-27 10:29:43 -05:00
examples below.
What is so hard, then?
----------------------
Despite the fact that the configure script takes normal --build and
--host options, that is just the beginning. In practise a lot of work
was necessary to separate tests for "host" and "build" platforms in
the configure script. See the git log for details. And the reasonably
"standard" configure.in is just the top level; when we get down to the
actual makefilery used to build the bits of LibreOffice, it gets much
worse.
Windows
-------
There is some support in LibreOffice already (from OpenOffice.org) for
2012-01-22 02:18:47 -06:00
building it locally on Windows with the GNU tool-chain
(MinGW). Apparently, that work has never attempted cross-compilation.
2011-05-27 10:29:43 -05:00
This OOo-originated MinGW support attempts to support both running
2011-08-28 03:14:47 -05:00
Cygwin gcc in its -mno-cygwin mode, and a native MinGW compiler. The
2011-05-27 10:29:43 -05:00
-mno-cygwin mechanism in the Cygwin gcc is rapidly being obsoleted, if
2012-01-22 02:18:47 -06:00
it isn't already, and we have not attempted to try to keep it working;
in fact we have activly cleaned out mechanisms related to this. Ditto
2011-08-27 03:11:14 -05:00
for native MinGW. If one compiles natively on Windows, just use
Microsoft's compiler. OOo/LO has been built for Windows all the time
using that.
2011-06-18 03:31:50 -05:00
2012-01-22 02:18:47 -06:00
The only case where it makes sense to use MinGW is for
2011-06-18 03:31:50 -05:00
cross-compilation. There is just too much crack on Windows anyway, and
it is a semi-miracle (well, make that the result of years of work)
that the MSVC build under Cygwin works as nicely as it does.
2011-05-27 10:29:43 -05:00
MinGW is available as cross-build toolchains pre-packaged in more or
2012-01-22 02:18:47 -06:00
less official packages for many Linux distros including Debian,
Fedora, openSUSE and SLE. For instance the mingw32 packages in the
Open Build Service, running on openSUSE:
http://download.opensuse.org/repositories/windows:/mingw:/win32/
For example, you can install it like this:
zypper ar http://download.opensuse.org/repositories/windows:/mingw:/win32/<your_os>/windows:mingw:win32.repo
where <your_os> is one of SLE_11, SLE_11_SP1, openSUSE_11.3, openSUSE_11.4 or
openSUSE_Factory.
2011-08-24 03:26:11 -05:00
zypper in mingw32-cross-gcc mingw32-cross-gcc-c++ mingw32-python-devel \
mingw32-python \
2011-08-24 03:26:11 -05:00
mingw32-libexpat-devel mingw32-libexpat mingw32-boost-devel \
mingw32-libhyphen-devel mingw32-libhyphen mingw32-hyphen-en \
mingw32-liblpsolve mingw32-liblpsolve-devel \
mingw32-libxml2-devel mingw32-libxslt-devel mingw32-libicu \
mingw32-libicu-devel mingw32-libgraphite2 mingw32-libgraphite2-devel \
mingw32-libcairo2 mingw32-cairo-devel mingw32-librsvg mingw32-librsvg-devel \
mingw32-hunspell mingw32-hunspell-devel mingw32-libcurl \
mingw32-libcurl-devel mingw32-libneon mingw32-libneon-devel \
mingw32-libopenssl mingw32-libopenssl-devel mingw32-libexttextcat \
mingw32-libexttextcat-devel mingw32-libdb mingw32-libdb-devel \
2011-08-24 03:26:11 -05:00
mingw32-cross-pkg-config mingw32-pkg-config mingw32-libcppunit \
mingw32-libcppunit-devel mingw32-libredland mingw32-libredland-devel \
mingw32-libmythes mingw32-libmythes-devel
You also need wine, ideally:
zypper ar http://download.opensuse.org/repositories/Emulators:/Wine/<your_os>/Emulators:Wine.repo
zypper in wine wine-devel wine-devel-32bit
And in order to be able to use 'winegcc -m32', also
zypper in glibc-devel-32bit gcc-32bit
There might be more that are missing, please read carefully what autogen.sh
tells you, and either remove one of the --with-system-*, or install the
missing dependency.
It also looks like graphite2.pc needs tweaking in order to work right; but
2011-08-28 03:14:47 -05:00
that's likely to be fixed in the openSUSE project.
2011-05-27 10:29:43 -05:00
It is somewhat unclear how well thought-out the conditionals and code
2011-06-18 03:31:50 -05:00
for MinGW inside the OOo-originated code in LibreOffice actually
2012-01-22 02:18:47 -06:00
are. It often seems a bit randomish, with copy-pasting having been
preferred to factoring out differences.
2011-05-27 10:29:43 -05:00
Most of the configuration settings are maintained in the
distro-configs/LibreOfficeMinGW.conf file, so in your autogen.lastrun,
you can use:
CC=ccache i686-w64-mingw32-gcc
CXX=ccache i686-w64-mingw32-g++
CC_FOR_BUILD=ccache gcc
CXX_FOR_BUILD=ccache g++
--with-distro=LibreOfficeMinGW
Alternatively, you can use something like the following; but the
preferred way is to keep the LibreOfficeMinGW.conf file up-to-date.
2011-05-27 10:29:43 -05:00
CC=ccache i686-w64-mingw32-gcc
CXX=ccache i686-w64-mingw32-g++
CC_FOR_BUILD=ccache gcc
CXX_FOR_BUILD=ccache g++
--build=x86_64-unknown-linux-gnu
--host=i686-w64-mingw32
--with-distro=LibreOfficeWin32
--disable-activex
--disable-binfilter
2011-05-27 10:29:43 -05:00
--disable-build-mozilla
--disable-directx
2011-05-27 10:29:43 -05:00
--disable-ext-nlpsolver
--disable-ext-pdfimport
--disable-ext-presenter-console
--disable-ext-presenter-minimizer
--disable-ext-report-builder
--disable-ext-scripting-beanshell
--disable-ext-scripting-javascript
--disable-ext-wiki-publisher
--disable-ext-wiki-publisher
--disable-mozilla
--disable-nss-module
2011-05-27 10:29:43 -05:00
--disable-zenity
2011-06-13 07:46:26 -05:00
--enable-python=system
2011-05-27 10:29:43 -05:00
--with-system-altlinuxhyph
--with-system-boost
2011-06-18 03:31:50 -05:00
--with-system-cairo
--with-system-cppunit
2011-05-27 10:29:43 -05:00
--with-system-curl
--with-system-db
--with-system-expat
2011-06-13 07:46:26 -05:00
--with-system-gettext
2011-05-27 10:29:43 -05:00
--with-system-hunspell
--with-system-icu
2011-06-18 03:31:50 -05:00
--with-system-libpng
2011-05-27 10:29:43 -05:00
--with-system-libwpd
--with-system-libwpg
2011-06-13 07:46:26 -05:00
--with-system-libwps
2011-05-27 10:29:43 -05:00
--with-system-libxml
2011-06-13 07:46:26 -05:00
--with-system-lpsolve
2011-05-27 10:29:43 -05:00
--with-system-mythes
2011-06-13 07:46:26 -05:00
--with-system-neon
--with-system-openssl
--with-system-redland
2011-05-27 10:29:43 -05:00
--with-vendor=no
--without-help
--without-helppack-integration
--without-myspell-dicts
2011-05-27 10:29:43 -05:00
Once you have compiled it, you may want to try to run it, for instance
using Wine:
2011-09-27 05:02:01 -05:00
$ cd /tmp
$ tar xf <your-build-dir>/instsetoo_native/wntgcci.pro/LibreOffice_Dev/archive/install/en-US/LibO-Dev_3.6.0alpha0_Win_x86_install-arc_en-US.tar.gz
$ cd LibO-Dev_3.6.0alpha0_Win_x86_install-arc_en-US/LOdev\ 3.6/program
2011-09-27 05:02:01 -05:00
$ wine soffice.exe
NB. it is important to unpack somewhere low in the hierarchy structure (like
in /tmp as advised above), otherwise you'll get BerkeleyDB errors on startup.
2011-09-27 05:02:01 -05:00
And if you are brave enough, you can even debug it. First you have to add the
URE dll's to the wine's PATH using 'wine regedit' - see
http://www.winehq.org/docs/wineusr-guide/environment-variables, and add
Z:\tmp\LibO-Dev_3.6.0alpha0_Win_x86_install-arc_en-US\LOdev 3.6\URE\bin
to "Path" in My Computer->HKEY_CURRENT_USER->Environment.
2011-09-27 05:02:01 -05:00
Then run linkoo, so that when you rebuild something, you can directly see the
changes the next time you run it:
solenv/bin/linkoo '/tmp/LibO-Dev_3.6.0alpha0_Win_x86_install-arc_en-US/LOdev 3.6' <your_clone_dir>
2011-09-27 05:02:01 -05:00
And start debugging:
$ winedbg soffice.bin
Would be great to be able to use winedbg --gdb, but it was crashing here :-( -
but maybe you'll be more lucky.
2012-03-19 09:27:25 -05:00
Tricks of some use with winedbg:
--------------------------------
To examine OUStrings, you might want to use the following trick (prints 50
unicode characters of rLibName OUString - the +10 is where the buffer starts):
Wine-dbg>x /50u rLibName->pData+10
0x0909b6c8: vnd.sun.star.expand:$LO_LIB_DIR/abplo.dll
2011-08-25 06:11:26 -05:00
TODO:
- make the debugging more convenient on (native) Windows
- check possibilities like WinGDB - http://www.wingdb.com/
- or find / write a MSVS / WinDBG extension that can read MinGW debugging
symbols natively; more info
http://windbg.info/forum/12-symbol-and-source-files-/21-debugging-mingwgcc-built-dll-in-visual-studio.html
2011-08-25 06:11:26 -05:00
- installation
- so far the make_installer.pl calls makecab.exe, uuidgen.exe, and
others; would be best to avoid that if at all possible (using a free
cab implementation, part of Wine or something)
- MSI generation
- if at all possible, the make dev-install installation (with links
back to the build) should be done so that it would be directly
executable via wine after doing make dev-install :-)
- runtime
- no idea if the entire thing works after the installation at all; I
suppose there will be runtime problems to look at too
- cleanup
- enable & fix pieces that are currently disabled
- --without-myspell-dicts
- --disable-directx
- --disable-activex
- --disable-mozilla
- much of the stuff currently relies on --with-system-*, and
consequently on the mingw32-* openSUSE packages; might be good to be
able to build with as few dependencies as possible - but that is low
2012-01-22 02:18:47 -06:00
prio
2011-08-25 06:11:26 -05:00
- profiling
- when all the above is sorted out, we should look at the speed of
this vs. the speed of the MSVC version
2011-05-27 10:29:43 -05:00
iOS
---
iOS is the operating system of Apple's mobile devices. Clearly for a
device like the iPad it would be totally unacceptable to run a normal
LibreOffice application with a overlapping windows and mouse-oriented
GUI widgets. No work has been done (at least publicly) by others to
design a touch GUI for LibreOffice, so that is something that needs to
be done.
Obviously it will make sense to use only a part of LibreOffice's code
for iOS. Most likely lots of the GUI-oriented code should be left out,
and some iOS app(s) that eventually wants to use the remaining bits
will handle all its GUI in a platform-dependent manner. How well it
will be possible to do such a split remains to be seen.
2011-05-27 10:29:43 -05:00
Technically, one important special aspect of iOS is that apps are not
allowed to load own dynamic libraries. (System libraries are used in
2011-11-08 07:38:18 -06:00
the form of dynamic libraries, just like on Mac OS X, of which iOS is a
2011-05-27 10:29:43 -05:00
variant.) So all the libraries in LibreOffice that normally are shared
libraries (DLLs on Windows, shared objects (.so) on Linux, dynamic
2011-11-08 07:38:18 -06:00
libraries on Mac OS X (.dylib)) need to be built as static archives
2011-05-27 10:29:43 -05:00
instead. Obviously this will have some interesting consequences for
how UNO is implemented and used. None of that has been spared much
thought yet.
The Apple tool-chain for iOS cross-building is available only for
2012-01-22 02:18:47 -06:00
Mac OS X.
2011-05-27 10:29:43 -05:00
Here is an autogen.lastrun for iOS (device) using Xcode 4.3 and clang,
and Xcode 3 for the build platform parts:
2012-02-29 16:40:49 -06:00
CXX=ccache /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang++ -arch armv7 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS5.0.sdk
CC=ccache /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang -arch armv7 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS5.0.sdk
2011-11-10 23:52:01 -06:00
CC_FOR_BUILD=ccache /Xcode3/usr/bin/gcc-4.0 -mmacosx-version-min=10.4
CXX_FOR_BUILD=ccache /Xcode3/usr/bin/g++-4.0 -mmacosx-version-min=10.4
2011-05-27 10:29:43 -05:00
--with-distro=LibreOfficeiOS
2012-02-22 12:41:58 -06:00
--with-system-nss
2011-11-10 23:52:01 -06:00
--without-help
--without-helppack-integration
--without-myspell-dicts
2011-05-27 10:29:43 -05:00
2012-02-22 12:41:58 -06:00
The --with-system-nss doesn't mean that any Mozilla NSS would be used,
it's just a trick to bypass some tests in the configure script.
For the iOS simulator, using Xcode 4.3 and clang both for the host and
build platform:
CXX=ccache /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang++ -arch i386 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator5.1.sdk -mmacosx-version-min=10.6
CC=ccache /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang -arch i386 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator5.1.sdk -mmacosx-version-min=10.6
CC_FOR_BUILD=ccache clang -arch i386 -mmacosx-version-min=10.6
CXX_FOR_BUILD=ccache clang++ -arch i386 -mmacosx-version-min=10.6
2012-02-22 12:41:58 -06:00
--enable-werror
--with-build-platform-configure-options=--with-macosx-version-min-required=10.6 --with-macosx-sdk=10.6
2012-02-22 12:41:58 -06:00
--with-distro=LibreOfficeiOS
--with-system-nss
--without-help
--without-helppack-integration
--without-myspell-dicts
Note that you will have to install autoconf and automake yourself
before running autogen.sh. They are no longer included in Xcode 4.3
(not even in the add-on "command line tools").
The -mmacosx-version-min=10.6 is necessary when building for the iOS
simulator to avoid clang replacing simple calls to fprintf with calls
to fwrite$UNIX2003 which Xcode then warns that doesn't exist on iOS.
2012-02-22 12:41:58 -06:00
2011-05-27 10:29:43 -05:00
Android
-------
From a technical point of view the core Android OS is Linux, but
everything else is different. Unlike iOS, an Android app can use
shared objects just fine, so that aspect of UNO doesn't need special
handling. Except that there is a silly low limit in the Android
dynamic linker on the number of libraries you can dlopen. Note that
this is a limitation of user-level code, not the kernel. Here are
various ways around this problem that are being examined.
2011-05-27 10:29:43 -05:00
As for the GUI, the same holds as said above for iOS.
2012-01-22 02:18:47 -06:00
Android cross-compilation work has been done on Linux (openSUSE in
2011-11-08 07:38:18 -06:00
particular) and Mac OS X. The Android cross-buld tool-chain (the
"Native Development Kit", or NDK) is available for Linux and Mac OS
X. Sure, for Windows, too, but trying to cross-compile LO from Windows
will probably drive you insane.
2011-05-27 10:29:43 -05:00
2012-01-22 02:18:47 -06:00
Here is an autogen.lastrun for Android, when cross-compiling from Mac OS X:
CC_FOR_BUILD=ccache /Xcode3/usr/bin/gcc-4.0 -mmacosx-version-min=10.4
CXX_FOR_BUILD=ccache /Xcode3/usr/bin/g++-4.0 -mmacosx-version-min=10.4
--with-android-ndk=/Users/tml/android-ndk-r7b
--with-android-sdk=/Users/tml/android-sdk-macosx
--build=x86_64-apple-darwin11
--disable-zenity
--enable-werror
--with-distro=LibreOfficeAndroid
--disable-python
--without-helppack-integration
--without-myspell-dicts
--with-java
2012-01-22 02:18:47 -06:00
And here is an autogen.lastrun for Android when cross-compiling from Linux:
2011-05-27 10:29:43 -05:00
CC_FOR_BUILD=ccache gcc
CXX_FOR_BUILD=ccache g++
--with-android-ndk=/home/tml/android-ndk-r7b
--with-android-sdk=/home/tml/android-sdk-linux_x86
--with-ant-home=/opt/apache-ant-1.8.2
2011-05-27 10:29:43 -05:00
--build=x86_64-unknown-linux-gnu
--disable-zenity
--with-distro=LibreOfficeAndroid
2011-06-22 01:05:05 -05:00
--disable-python
--with-num-cpus=1
--with-max-jobs=1
--without-helppack-integration
--without-myspell-dicts
2011-05-27 10:29:43 -05:00
2012-01-03 05:48:29 -06:00
There is no interactive end-user "app" you could run yet that would
use LibreOffice code, but you can build some non-interactive unit
2012-01-22 02:18:47 -06:00
tests and run them on the emulator (or, presumably, on a device,
although that hasn't been tested as nobody with an actual Android
device has worked on it yet...) The simple unit tests will succeed,
the complex one still fails.
2012-01-03 05:48:29 -06:00
These unit tests *are* proper "apps" from Android'd point of view, but
they don't have any GUI and thus don't take part in the normal Android
message passing and Android thinks they are stuck and offers to kill
them...
The activity used for these apps is in android/Bootstrap. See
README.Android.
To build the complex unit test for Calc functionality which invokes the
"ucalc" unit test code from sc, cd to android/qa/sc and run a "make". The
Makefile here is completely manually written, this stuff is so experimental it
doesn't make sense yet to try to integrate with the normal gbuild mechanism.
Note that lately (as of May 2012) the android/qa/sc test has not been
maintained, and might not build cleanly.
2012-01-03 05:48:29 -06:00
Then to run the unit test, do "make install" followed by "make
run". You most likely want to have an "adb logcat" running in another
window.
To debug, do manually what "make run" would do, adding args "-e lo-main-delay
20" to the command line, and when the app has started, run
ndk-gdb. Unfortunately the gdb in NDK r7 and r8 is a bit broken, you can use
the one in a NDK build with newer versions of gcc and gdb from
http://code.google.com/p/mingw-and-ndk/ instead.
2012-01-03 05:48:29 -06:00
Running strace on the unit test in progress is often useful to find
out what is going wrong. Pass something like -e lo-strace '-tt -f -e
trace=file,process,network -o /system/sc/strace.out' to the am start
command line.
The Android test app which has been exercised lately is in
android/experimental/DocumentLoader.
2011-05-27 10:29:43 -05:00
PowerPC Mac OS X
----------------
Cross-compiling for PowerPC Mac OS X from Intel Mac OS X will probably
be easy. The APIs available should after all be closely identical to
those on Intel Mac OS X, and LibreOffice builds fine natively on
2012-01-22 02:18:47 -06:00
PowerPC Mac already. Only a little experimenting has been done with
it. An autogen.lastrun looks like this:
CC=ccache /Xcode3/usr/bin/gcc-4.0 -arch ppc
CXX=ccache /Xcode3/usr/bin/g++-4.0 -arch ppc
CC_FOR_BUILD=ccache /Xcode3/usr/bin/gcc-4.0
CXX_FOR_BUILD=ccache /Xcode3/usr/bin/g++-4.0
--build=i386-apple-darwin10.7.0
--host=powerpc-apple-darwin10
--disable-mozilla
--disable-build-mozilla