strepsirrhini/office-gobmx
9
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
office-gobmx/static/README.wasm.md
Stephan Bergmann 887ebfd6e5 Embind: Centrally initialize via Module.initUno() in a new uno.js 
...so that the unoObject function can be moved there too (so that other code
outside embindtest.js can reuse it)

Change-Id: Id3edb7cede56321db29ba435f221cb7702a3513c
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/168700
Reviewed-by: Stephan Bergmann <stephan.bergmann@allotropia.de>
Tested-by: Jenkins
2024-06-12 15:44:12 +02:00

16 KiB
Raw Blame History

Support for Emscripten Cross Build

This subdirectory provides support for building LibreOffice as WASM, with the Emscripten toolchain.

You can build LibreOffice for WASM for two separate purposes: 1) Either to produce a WASM binary of LibreOffice as such, using Qt5 for its GUI, or 2) just compiling LibreOffice core ("LibreOffice Technology") to WASM without any UI for use in other software that provides the UI, like Collabora Online built as WASM.

The first purpose was the original reason for the WASM port and this document was originally written with that in mind. For the second purpose, look towards the end of the document for the section "Building headless LibreOffice as WASM for use in another product".

Status of LibreOffice as WASM with Qt

The build generates a Writer-only LO build. You should be able to run either

$ emrun --serve_after_close instdir/program/qt_soffice.html
$ emrun --serve_after_close workdir/LinkTarget/Executable/qt_vcldemo.html
$ emrun --serve_after_close workdir/LinkTarget/Executable/qt_wasm-qt5-mandelbrot.html

REMINDER: Always start new tabs in the browser, reload might fail / cache! INFO: latest browser won't work anymore with 0.0.0.0 and need 127.0.0.1.

Setup for the LO WASM build (with Qt)

We're using Qt 5.15.2 with Emscripten 3.1.46. There are a bunch of Qt patches to fix the most grave bugs. Also there's rapid development in Emscripten, so using another version often causes arbitrary problems.

  • See below under Docker build for another build option

Setup emscripten

https://emscripten.org/docs/getting_started/index.html

git clone https://github.com/emscripten-core/emsdk.git
./emsdk install 3.1.46
./emsdk activate 3.1.46

Example bashrc scriptlet:

EMSDK_ENV=$HOME/Development/libreoffice/git_emsdk/emsdk_env.sh
[ -f "$EMSDK_ENV" ] && \. "$EMSDK_ENV" 1>/dev/null 2>&1

Setup Qt

https://doc.qt.io/qt-5/wasm.html

Most of the information from https://doc.qt.io/qt-6/wasm.html is still valid for Qt5; generally the Qt6 WASM documentation is much better, because it incorporated many information from the Qt Wiki.

FWIW: Qt 5.15 LTS is not maintained publicly and Qt WASM has quite a few bugs. Most WASM fixes from Qt 6 are needed for Qt 5.15 too. Allotropia offers a Qt repository with the necessary patches cherry-picked.

With "-opensource -confirm-license" you agree to the open source license.

git clone https://github.com/allotropia/qt5.git
cd qt5
git checkout v5.15.2+wasm
./init-repository --module-subset=qtbase
./configure -opensource -confirm-license -xplatform wasm-emscripten -feature-thread -prefix <whatever> QMAKE_CFLAGS+=-sSUPPORT_LONGJMP=wasm QMAKE_CXXFLAGS+=-sSUPPORT_LONGJMP=wasm
make -j<CORES> module-qtbase

Optionally you can add the configure flag "-compile-examples". But then you also have to patch at least mkspecs/wasm-emscripten/qmake.conf with EXIT_RUNTIME=0, otherwise they will fail to run. In addition, building with examples will break with some of them, but at that point Qt already works and also most examples. Or just skip them. Other interesting flags might be "-nomake tests -no-pch -ccache".

Linking takes quite a long time, because emscripten-finalize rewrites the whole WASM files with some options. This way the LO WASM possibly needs 64GB RAM. For faster link times add "-s WASM_BIGINT=1", change to ASSERTIONS=1 and use -g3 to prevent rewriting the WASM file and generating source maps (see emscripten.py, finalize_wasm, and avoid modify_wasm = True). This is just needed for Qt examples, as LO already uses the correct flags!

It's needed to install Qt5 to the chosen prefix. Else LO won't find all needed files in the right place. For installation you can do

make -j<CORES> install

or make -j8 -C qtbase/src install_subtargets

Current Qt fails to start the demo webserver: https://bugreports.qt.io/browse/QTCREATORBUG-24072

Use emrun --serve_after_close to run Qt WASM demos.

Setup LO

autogen.sh is patched to use emconfigure. That basically sets various environment vars, especially EMMAKEN_JUST_CONFIGURE, which will create the correct output file names, checked by configure (a.out).

There's a distro config for WASM, but it just provides --host=wasm32-local-emscripten, which should be enough setup. The build itself is a cross build and the cross-toolset just depends on a minimal toolset (gcc, libc-dev, flex, bison); all else is build from source, because the final result is not depending on the build system at all.

Recommended configure setup is thusly:

  • grab defaults --with-distro=LibreOfficeWASM32

  • local config QT5DIR=/dir/of/qt5/install/prefix

  • if you want to use ccache on both sides of the build --with-build-platform-configure-options=--enable-ccache --enable-ccache

FWIW: it's also possible to build an almost static Linux LibreOffice by just using --disable-dynloading --enable-customtarget-components. System externals are still linked dynamically, but everything else is static.

"Deploying" soffice.wasm

tar -chf wasm.tar --xform 's/.*program/lo-wasm/' instdir/program/soffice.* \
    instdir/program/qt*

Your HTTP server needs to provide additional headers:

  • add_header Cross-Origin-Opener-Policy same-origin
  • add_header Cross-Origin-Embedder-Policy require-corp

The default html to use should be qt_soffice.html

Debugging setup

Since a few months you can use DWARF information embedded by LLVM into the WASM to debug WASM in Chrome. You need to enable an experimental feature and install an additional extension. The whole setup is described in:

https://developer.chrome.com/blog/wasm-debugging-2020/

This way you don't need source maps (much faster linking!) and can resolve local WASM variables to C++ names!

Per default, the WASM debug build splits the DWARF information into an additional WASM file, postfixed '.debug.wasm'.

Using Docker to cross-build with emscripten

If you prefer a controlled environment (sadly emsdk install/activate is not stable over time, as e.g. nodejs versions evolve), that is easy to replicate across different machines - consider the docker images we're providing.

Config/setup file see https://git.libreoffice.org/lode/+/ccb36979563635b51215477455953252c99ec013

Run

docker-compose build

in the lode/docker dir to get the container prepared. Run

PARALLELISM=4 BUILD_OPTIONS= BUILD_TARGET=build docker-compose run --rm \
    -e PARALLELISM -e BUILD_TARGET -e BUILD_OPTIONS builder

to perform an actual srcdir != builddir build; the container mounts checked-out git repo and output dir via docker-compose.yml (so make sure the path names there match your setup):

The lode setup expects, inside the lode/docker subdir, the following directories:

  • core (git checkout)
  • workdir (the output dir - gets written into)
  • cache (ccache tree)
  • tarballs (external project tarballs gets written and cached there)

Ideas for an UNO bridge implementation

My post to Discord #emscripten:

"I'm looking for a way to do an abstract call from one WASM C++ object to another WASM C++ object, so like FFI / WebIDL, just within WASM. All my code is C++ and normally I have bridge code, with assembler to implement the function call /RTTI and exception semantics of the specified platform. Code is at https://cgit.freedesktop.org/libreoffice/core/tree/bridges/source/cpp_uno. I've read a bit about call_indirect and stuff, but I don't have yet a good idea, how I could implement this (and there is an initial feature/wasm branch for the interested). I probably need some fixed lookup table, like on iOS, because AFAIK you can't dynamically generate code in WASM. So any pointers or ideas for an implementation? I can disassemble some minimalistic WASM example and read clang code for WASM_EmscriptenInvoke, but if there were some standalone code or documentation I'm missing, that would be nice to know."

We basically would go the same way then the other backends. Write the bridge in C++, which is probably largely boilerplate code, but the function call in WAT (https://github.com/WebAssembly/wabt) based on the LLVM WASM calling conventions in WASM_EmscriptenInvoke. I didn't get a reply to that question for hours. Maybe I'll open an Emscripten issue, if we really have to implement this.

WASM dynamic dispatch:

UNO bindings with Embind

Right now there's a very rough implementation in place. With lots of different bits unimplemented. And it might be leaking memory. i.e. Lots of room for improvement! ;)

Some usage examples through javascript of the current implementation:

// inserts a string at the start of the Writer document.
Module.initUno();
const css = Module.uno.com.sun.star;
const xModel = Module.getCurrentModelFromViewSh();
const xTextDocument = css.text.XTextDocument.query(xModel);
const xText = xTextDocument.getText();
const xTextCursor = xText.createTextCursor();
xTextCursor.setString("string here!");

// changes each paragraph of the Writer document to a random color.
Module.initUno();
const css = Module.uno.com.sun.star;
const xModel = Module.getCurrentModelFromViewSh();
const xTextDocument = css.text.XTextDocument.query(xModel);
const xText = xTextDocument.getText();
const xEnumAccess = css.container.XEnumerationAccess.query(xText);
const xParaEnumeration = xEnumAccess.createEnumeration();
while (xParaEnumeration.hasMoreElements()) {
    const next = xParaEnumeration.nextElement();
    const xParagraph = css.text.XTextRange.query(next.get());
    const xParaProps = css.beans.XPropertySet.query(xParagraph);
    const color = new Module.uno_Any(Module.uno_Type.Long(), Math.floor(Math.random() * 0xFFFFFF));
    xParaProps.setPropertyValue("CharColor", color);
    next.delete();
    color.delete();
}

Tools for problem diagnosis

  • nm -s should list the symbols in the archive, based on the index generated by ranlib. If you get linking errors that archive has no index.

Emscripten filesystem access with threads

This is closed, but not really fixed IMHO:

Dynamic libraries / modules in emscripten

There is a good summary in:

Summary: you can't use modules and threads.

This is mentioned at the end of:

The usage of MAIN_MODULE and SIDE_MODULE has other problems, a major one IMHO is symbol resolution at runtime only. So this works really more like plugins in the sense of symbol resolution without dependencies / rpath.

There is some clang-level dynamic-linking in progress (WASM dlload). The following link is already a bit old, but I found it a god summary of problems to expect:

Mixed information, links, problems, TODO

More info on Qt WASM emscripten pthreads:

WASM needs -pthread at compile, not just link time for atomics support. Alternatively you can provide -s USE_PTHREADS=1, but both don't seem to work reliable, so best provide both. https://github.com/emscripten-core/emscripten/issues/10370

The output file must have the prefix .o, otherwise the WASM files will get a node.js shebang (!) and ranlib won't be able to index the library (link errors).

Qt with threads has further memory limit. From Qt configure:

Project MESSAGE: Setting PTHREAD_POOL_SIZE to 4
Project MESSAGE: Setting TOTAL_MEMORY to 1GB

You can actually allocate 4GB:

LO uses a nested event loop to run dialogs in general, but that won't work, because you can't drive the browser event loop. like VCL does with the system event loop in the various VCL backends. Changing this will need some major work (basically dropping Application::Execute).

But with the know problems with exceptions and threads, this might change:

We're also using emconfigure at the moment. Originally I patched emscripten, because it wouldn't create the correct a.out file for C++ configure tests. Later I found that the emconfigure sets EMMAKEN_JUST_CONFIGURE to work around the problem.

ICU bug:

Alternative, probably:

There is a wasm64, but that still uses 32bit pointers!

Old outdated docs:

Reverted patch:

Generally https://emscripten.org/docs/porting:

This will be interesting:

This didn't help much yet:

Emscripten supports standalone WASI binaries:

Building headless LibreOffice as WASM for use in another product

Set up Emscripten

Follow the instructions in the first part of this document.

No Qt needed.

You don't need any dependencies other than those that normally are downloaded and compiled when building LibreOffice.

Set up LO

For instance, this autogen.input works for me:

--disable-debug --enable-sal-log --disable-crashdump --host=wasm32-local-emscripten --disable-gui --with-main-module=writer

For building LO core for use in COWASM, it is known to work to use Emscripten 3.1.30 (and not just 2.0.31 which is what the LO+Qt5 work has been using in the past).

That's all

After all, in this case you are building LO core headless for it to be used by other software.

Note that a soffice.wasm will be built, but that is just because of how the makefilery has been set up. We do need the soffice.data file that contains the in-memory file system needed by the LibreOffice Technology core code during run-time, though. That is at the moment built as a side-effect when building soffice.wasm.