ceb26770b3
Analysing all the plugin sources using just one process takes a lot of time, so split out a separate analyzer tool that'll analyse one source and print out the data to a .plugininfo file. The generator then will read all of these and generate sharedvisitor.cxx . This allows parallelising the expensive analysis. With this commit sharedvisitor.cxx is no longer included in the repository, as this and the next commit should make the generation fast enough. Change-Id: Idfc33c4ea6ccfd84f829b51001c8ddeb0be09961 Reviewed-on: https://gerrit.libreoffice.org/78568 Tested-by: Jenkins Reviewed-by: Luboš Luňák <l.lunak@collabora.com>
46 lines
2.6 KiB
Plaintext
46 lines
2.6 KiB
Plaintext
These tools generate another "plugin" which in fact only dispatches Visit* and Traverse*
|
|
calls to all other plugins registered with it. This means that there is just one
|
|
RecursiveASTVisitor pass for all those plugins instead of one per each, which
|
|
with the current number of plugins actually makes a performance difference.
|
|
|
|
If you work on a plugin, comment out LO_CLANG_SHARED_PLUGINS in Makefile-clang.mk in order
|
|
to disable the feature (re-generating takes time).
|
|
|
|
There are two tools:
|
|
- analyzer, which analyses one .cxx file (one plugins) and writes info about it to a .plugininfo
|
|
file, this allows parallelising this part, since it can take some time
|
|
- generator, which reads all the .plugininfo files and generates sharedvisitor.cxx
|
|
|
|
Requirements for plugins:
|
|
- Can use Visit* and Traverse* functions, but not WalkUp*.
|
|
- Visit* functions can generally remain unmodified.
|
|
- run() function must be split into preRun() and postRun() if there's any additional functionality
|
|
besides calling TraverseDecl(). The shared visitor will call the preRun() and postRun() functions
|
|
as necessary while calling its own run(). The run() function of the plugin must stay
|
|
(in case of a non-shared build) but should generally look like this:
|
|
if( preRun())
|
|
if( TraverseDecl(compiler.getASTContext().getTranslationUnitDecl()))
|
|
postRun();
|
|
- Traverse* functions must be split into PreTraverse* and PostTraverse*, similarly to how run()
|
|
is handled, the Traverse* function should generally look like this:
|
|
bool ret = true;
|
|
if( PreTraverse*(decl))
|
|
{
|
|
ret = RecursiveASTVisitor::Traverse*(decl);
|
|
PostTraverse*(decl, ret);
|
|
}
|
|
return ret;
|
|
|
|
|
|
TODO:
|
|
- Create macros for the standardized layout of run(), Traverse*, etc.?
|
|
- Possibly check plugin sources more thoroughly (e.g. that run() doesn't actually do more).
|
|
- Have one tool that extracts info from plugin .cxx files into some .txt file and another tool
|
|
that generates sharedvisitor.cxx based on those files? That would generally make the generation
|
|
faster when doing incremental changes. The .txt file could also contain some checksum of the .cxx
|
|
to avoid the analysing pass completely if just the timestamp has changed.
|
|
- Do not re-compile sharedvisitor.cxx if its contents have not actually changed.
|
|
- Is it possible to make the clang code analyze just the .cxx without also parsing all the headers?
|
|
- Instead of having to comment out LO_CLANG_SHARED_PLUGINS, implement --enable-compiler-plugins=debug .
|
|
- Try make analyzer use a precompiled header of Clang headers, for better performance.
|