Stephan Bergmann
3645088025
With Qt5, the (enabled by default) --enable-emscripten-proxy-to-pthread (i.e., leveraging Emscripten's -sPROXY_TO_PTHREAD) works reasonably well with the handful of hacks we added to our Qt5 5.15.2+wasm branch. However, for Qt6 any necessary hacking there would need to be excessive and looks unrealistic (see the mailing list thread starting at <https://lists.qt-project.org/pipermail/development/2024-December/045960.html> "[Development] Wasm: Support for Emscripten PROXY_TO_PTHREAD?" and the presentation at <https://fosdem.org/2025/schedule/event/fosdem-2025-5169-lowa-in-need-of-a-vcl-plug/>). But when leveraging the upcoming browser support for JSPI (through Emscripten's -sJSPI, Qt6's -feature-wasm-jspi, and our --enable-emscripten-jspi), it looks feasible to leave the Qt6 code alone and instead of proxying the complete application main thread (including the Qt6 event handling) off the browser main thread to an additional thread (automatically set up through Emscripten -sPROXY_TO_PTHREAD), to just proxy some of the LO event handling off the browser main thread to an additional thread (set up through a new comphelper::emscriptenthreading::setUp). That way, problematic LO code that is triggered from event handling (like "Tools - Extensions..." wanting to spawn additional threads, which requires the browser main thread to be unblocked) can be offloaded to the additional thread and get the browser main thread unblocked. Thus there is a new "Emscripten Qt6 JSPI/non-PROXY_TO_PTHREAD" mode now (via --enable-qt6 --enable-emscripten-jspi --disable-emscripten-proxy-to-pthread). It is still experimental and known to occasionally hang and crash, but has been seen to generally work at least with recent emsdk 4.0.3 and a recent Qt dev branch (towards Qt 6.10, built with -feature-wasm-jspi) and running on recent Chrome 132 with "Experimental WebAssembly JavaScript Promise Integration (JSPI)" enabled under <chrome://flags>. There are two places where code is proxied off the browser main thread to the additional thread: One is QtInstance::ProcessEvent, the other is invoking certain kinds of tasks (just "vcl::Dialog maLayoutIdle", for now) in Scheduler::CallbackTaskScheduling (unconditionally proxying off all kinds of tasks there appeared to cause more issues than it would solve). For the latter, the concept of "transferability" has been added to class Task; the new Emscripten mode is the only case making use of that concept. Also for the latter, it is important that the browser main thread does not get blocked at the > SolarMutexGuard aGuard; in QtTimer::timeoutActivated when such a timer event happens while the additional thread has the SolarMutex locked and tries to do something that requires the browser main thread to not be blocked (like spawning additional threads upon "Tools - Extensions..."). Therefore, as a bad hack, and only for the new Emscripen mode, that SolarMutexGuard has been moved from QtTimer::timeoutActivated down to the call of pTask->Invoke() in Scheduler::CallbackTaskScheduling (see the two TODO comments added to the code). This appears to work relatively well, but is still a bad hack that should rather be done properly. And then there's some code across vcl/qt5 that may now end up running proxied off the browser main thread, but which actually requires to be run on the browser main thread (because, e.g., it calls into Qt6 code that accesses JS entities that are only available there). Those places use a newly introduced QtInstance::EmscriptenLightweightRunInMainThread that takes some code via a lambda: For the new Emscripten mode, the code is explicitly proxied back to the browser main thread, while in all other cases it is run directly in place. Some further notes: * -sPTHREAD_POOL_SIZE needed to be bumped by one, to accommodate for the additional new thread. * The code requires two additional JSPI "entry points", _emscripten_check_mailbox (to complement the suspension in QtInstance::ProcessEvent) and the invocation of the Qt6-internal qstdweb::EventListener::handleEvent (to complement a Qt6-internal suspension point hit when e.g. doing "Tools - Extension Manager..." and clicking "Add"). * The new Emscripten mode requires -fexperimental-library (for std::jthread and std::stop_token) when building at least with emsdk 4.0.3, but it shouldn't hurt to have that switch enabled unconditionally for Emscripten builds (which are self-contained, so can't be hit by any compatibility issues that switch might cause). Change-Id: Id3b13e2cc3c304b072f388c6d736e7e663b36c58 Reviewed-on: https://gerrit.libreoffice.org/c/core/+/181424 Tested-by: Jenkins Reviewed-by: Stephan Bergmann <stephan.bergmann@allotropia.de> Reviewed-by: Michael Weghorn <m.weghorn@posteo.de>
LibreOffice
LibreOffice is an integrated office suite based on copyleft licenses and compatible with most document formats and standards. Libreoffice is backed by The Document Foundation, which represents a large independent community of enterprises, developers and other volunteers moved by the common goal of bringing to the market the best software for personal productivity. LibreOffice is open source, and free to download, use and distribute.
A quick overview of the LibreOffice code structure.
Overview
You can develop for LibreOffice in one of two ways, one recommended and one much less so. First the somewhat less recommended way: it is possible to use the SDK to develop an extension, for which you can read the API docs and Developers Guide. This re-uses the (extremely generic) UNO APIs that are also used by macro scripting in StarBasic.
The best way to add a generally useful feature to LibreOffice is to work on the code base however. Overall this way makes it easier to compile and build your code, it avoids any arbitrary limitations of our scripting APIs, and in general is far more simple and intuitive - if you are a reasonably able C++ programmer.
The Build Chain and Runtime Baselines
These are the current minimal operating system and compiler versions to run and compile LibreOffice, also used by the TDF builds:
- Windows:
- Runtime: Windows 10
- Build: Cygwin + Visual Studio 2019 version 16.10
- macOS:
- Runtime: 10.15
- Build: 12 (13 for aarch64) + Xcode 14
- Linux:
- Runtime: RHEL 8 or CentOS 8 and comparable
- Build: either GCC 12; or Clang 12 with libstdc++ 10
- iOS (only for LibreOfficeKit):
- Runtime: 11.4 (only support for newer i devices == 64 bit)
- Build: Xcode 9.3 and iPhone SDK 11.4
- Android:
- Build: NDK r23 and SDK 30.0.3
- Emscripten / WASM:
- Runtime: a browser with SharedMemory support (threads + atomics)
- Build: Qt 5.15 with Qt supported Emscripten 1.39.8
- See README.wasm
Java is required for building many parts of LibreOffice. In TDF Wiki article Development/Java, the exact modules that depend on Java are listed.
The baseline for Java is Java Development Kit (JDK) Version 17 or later.
The baseline for Python is version 3.11. It follows the version available in SUSE Linux Enterprise Desktop and the Maintenance Support version of Red Hat Enterprise Linux.
If you want to use Clang with the LibreOffice compiler plugins, the minimal version of Clang is 12.0.1. Since Xcode doesn't provide the compiler plugin headers, you have to compile your own Clang to use them on macOS.
You can find the TDF configure switches in the distro-configs/ directory.
To setup your initial build environment on Windows and macOS, we provide the LibreOffice Development Environment (LODE) scripts.
For more information see the build instructions for your platform in the TDF wiki.
The Important Bits of Code
Each module should have a README.md file inside it which has some
degree of documentation for that module; patches are most welcome to
improve those. We have those turned into a web page here:
However, there are two hundred modules, many of them of only peripheral interest for a specialist audience. So - where is the good stuff, the code that is most useful. Here is a quick overview of the most important ones:
| Module | Description |
|---|---|
| sal/ | this provides a simple System Abstraction Layer |
| tools/ | this provides basic internal types: Rectangle, Color etc. |
| vcl/ | this is the widget toolkit library and one rendering abstraction |
| framework/ | UNO framework, responsible for building toolbars, menus, status bars, and the chrome around the document using widgets from VCL, and XML descriptions from /uiconfig/ files |
| sfx2/ | legacy core framework used by Writer/Calc/Draw: document model / load/save / signals for actions etc. |
| svx/ | drawing model related helper code, including much of Draw/Impress |
Then applications
| Module | Description |
|---|---|
| desktop/ | this is where the main() for the application lives, init / bootstrap. the name dates back to an ancient StarOffice that also drew a desktop |
| sw/ | Writer |
| sc/ | Calc |
| sd/ | Draw / Impress |
There are several other libraries that are helpful from a graphical perspective:
| Module | Description |
|---|---|
| basegfx/ | algorithms and data-types for graphics as used in the canvas |
| canvas/ | new (UNO) canvas rendering model with various backends |
| cppcanvas/ | C++ helper classes for using the UNO canvas |
| drawinglayer/ | View code to render drawable objects and break them down into primitives we can render more easily. |
Rules for #include Directives (C/C++)
Use the "..." form if and only if the included file is found next to the
including file. Otherwise, use the <...> form. (For further details, see the
mail Re: C[++]: Normalizing include syntax ("" vs
<>).)
The UNO API include files should consistently use double quotes, for the benefit of external users of this API.
loplugin:includeform (compilerplugins/clang/includeform.cxx) enforces these rules.
Finding Out More
Beyond this, you can read the README.md files, send us patches, ask
on the mailing list libreoffice@lists.freedesktop.org (no subscription
required) or poke people on IRC #libreoffice-dev on irc.libera.chat -
we're a friendly and generally helpful mob. We know the code can be
hard to get into at first, and so there are no silly questions.
SAST Tools
PVS-Studio - static analyzer for C, C++, C#, and Java code.