12 KiB
Troubleshooting
General issues
clang-uml crashes when generating diagram
If clang-uml crashes with a segmentation fault, it is possible to trace the
exact stack trace of the fault using the following steps:
First, build clang-uml from source in debug mode, e.g.:
make debug
Then run clang-uml, preferably with -vvv for verbose log output. If your
.clang-uml configuration file contains more than 1 diagram, specify only
a single diagram to make it easier to trace the root cause of the crash, e.g.:
debug/src/clang-uml -vvv -n my_diagram
After clang-uml crashes again, detailed backtrace (generated using
backward-cpp) should be visible
on the console.
If possible, create an issue and paste the stack trace and few last logs from the console.
Diagram generation is very slow
clang-uml uses
Clang's RecursiveASTVisitor,
to
traverse the source code. By default, this visitor is invoked on every
translation unit (i.e. each entry in your compile_commands.json), including
all of their header dependencies recursively. This means, that for large code
bases with hundreds or thousands of translation units, traversing all of them
will be slow (think clang-tidy slow...).
Fortunately, in most practical cases it is not necessary to traverse the entire source code for each diagram, as all the information necessary to generate a single diagram usually can be found in just a few translation units, or even a single one.
This is where the glob configuration parameter comes in. It can be used to
limit the number of translation units to visit for a given diagram, for
instance:
diagrams:
ClassAContextDiagram:
type: class
# ...
glob:
- src/classA.cpp
# ...
InterfaceHierarchyDiagram:
type: class
# ...
glob:
- src/interfaces/*.cpp
# ...
This should improve the generation times for individual diagrams significantly.
Furthermore, diagrams are generated in parallel if possible, by default using
as many threads as virtual CPU's are available on the system, however it can
be adjusted also manually using -t command line option.
Diagram generated with PlantUML is cropped
When generating diagrams with PlantUML without specifying an output file format, the default is PNG. Unfortunately PlantUML will not check if the diagram will fit in the default PNG size, and often the diagram will be incomplete in the picture. A better option is to specify SVG as output format and then convert to PNG, e.g.:
plantuml -tsvg mydiagram.puml
convert +antialias mydiagram.svg mydiagram.png
Clang produces several warnings during diagram generation
During the generation of the diagram clang may report a lot of warnings, which
do not occur during the compilation with other compiler (e.g. GCC). This can be
fixed easily by using the add_compile_flags config option. For instance,
assuming that the warnings are as follows:
... warning: implicit conversion from 'int' to 'float' changes value from 2147483647 to 2147483648 [-Wimplicit-const-int-float-conversion]
... warning: declaration shadows a variable in namespace 'YAML' [-Wshadow]
simply add the following to your .clang-uml configuration file:
add_compile_flags:
- -Wno-implicit-const-int-float-conversion
- -Wno-shadow
Alternatively, the same can be passed through the clang-uml command line, e.g.
clang-uml --add-compile-flag -Wno-implicit-const-int-float-conversion \
--add-compile-flag -Wno-shadow ...
Please note that if your compile_commands.json already contains - for instance
-Wshadow - then you also have to remove it, i.e.:
add_compile_flags:
- -Wno-implicit-const-int-float-conversion
- -Wno-shadow
remove_compile_flags:
- -Wshadow
Cannot generate diagrams from header-only projects
Currently, in order to generate UML diagrams using clang-uml it is necessary
that at least one translation unit (e.g. one cpp) exists and it is included
in the generated compile_commands.json database.
However, even if your project is a header only library, first check if the
generated compile_commands.json contains any entries - if yes you should be
fine - just make sure the glob pattern in the configuration file matches
any of them. This is due to the fact that most header only projects still have
test cases, which are compiled and executed, and which include the headers.
These are perfectly fine to be used as translation units to generate the
diagrams.
In case, the code really does not contain any translation units, you will have
to create one, typically a basic main.cpp which includes the relevant headers
should be fine.
Also, it's possible to simply create a separate project, with a single
translation unit, which includes the relevant headers and create diagrams
from there.
In the future there might be a workaround for this in clang-uml.
YAML anchors and aliases are not fully supported
clang-uml uses yaml-cpp library, which
currently does not support
merging YAML anchor dictionaries,
e.g. in the following configuration file the main_sequence_diagram will work,
but the foo_sequence_diagram will fail with parse error:
compilation_database_dir: debug
output_directory: output
.sequence_diagram_anchor: &sequence_diagram_anchor
type: sequence
glob: [ ]
start_from:
- function: 'main(int,const char**)'
diagrams:
main_sequence_diagram: *sequence_diagram_anchor # This will work
foo_sequence_diagram:
<<: *sequence_diagram_anchor # This will not work
glob: [ src/foo.cc ]
start_from:
- function: 'foo(int,float)'
One option around this is to use some YAML preprocessor, such as
yq on such file and passing
the configuration file to clang-uml using stdin, e.g.:
yq 'explode(.)' .clang-uml | clang-uml --config -
Class diagrams
"fatal error: 'stddef.h' file not found"
This error means that Clang cannot find some standard headers in include
paths specified in the compile_commands.json. This typically happens on macOS
and sometimes on Linux, when the code was compiled with different Clang version
than clang-uml itself.
One solution to this issue is to add the following line to your CMakeLists.txt
file:
set(CMAKE_CXX_STANDARD_INCLUDE_DIRECTORIES ${CMAKE_CXX_IMPLICIT_INCLUDE_DIRECTORIES})
Another option is to provide an option (on command line or in configuration
file) called query_driver (inspired by the clangd
language server - although much less flexible), which will invoke the
provider compiler command and query it for its default system paths, which then
will be added to each compile command in the database. This is especially useful
on macOS as well as for embedded toolchains, example usage:
clang-uml --query-driver arm-none-eabi-g++
Another option is to make sure that the Clang is installed on the system (even if not used for building your project), e.g.:
apt install clang
If this doesn't help to include paths can be customized using config options:
add_compile_flags- which adds a list of compile flags such as include paths to each entry of the compilation databaseremove_compile_flags- which removes existing compile flags from each entry of the compilation database
For instance:
add_compile_flags:
- -I/opt/my_toolchain/include
remove_compile_flags:
- -I/usr/include
These options can be also passed on the command line, for instance:
clang-uml --add-compile-flag -I/opt/my_toolchain/include \
--remove-compile-flag -I/usr/include ...
Also see here.
How can I generate class diagram of my entire project
I want to generate a diagram containing all classes and relationships in my project - I don't care how big it is going to be.
Of course this is possible, the best way to do this is to specify
that clang-uml
should only include elements defined in files contained in project sources,
e.g.:
diagrams:
all_classes:
type: class
include:
paths: [ include, src ]
As the diagram will be huge for even medium-sized projects, it will likely not be readable. However, this option can be useful for cases when we want to get a complete JSON model of the codebase using the JSON generator:
clang-uml -g json -n all_classes --progress
Cannot generate classes for 'std' namespace
Currently, system headers are skipped automatically by clang-uml, due to
too many errors they produce when generating diagrams, especially when trying
to process GCC's or MSVC's system headers by Clang - not yet sure why
that is the case.
Basically it's best to either include only specific namespaces through the inclusion filters, e.g.:
include:
namespaces:
- myns1::myns12
or explicitly exclude std namespace:
exclude:
namespaces:
- std
Hopefully this will be eventually resolved.
Sequence diagrams
Generated diagram is empty
In order to generate sequence diagram the start_from configuration option must
have a valid starting point
for the diagram (e.g. function), which must match exactly the function
signature in the clang-uml model.
Look for error in the console output such as:
Failed to find participant mynamespace::foo(int) for start_from condition
which means that either you have a typo in the function signature in the
configuration file, or that the function
was not defined in the translation units you specified in the glob patterns
for this diagram. Run again the
clang-uml tool with -vvv option and look in the console output for any
mentions of the function from
which the diagram should start and copy the exact signature into the
configuration file.
Generated diagram contains several empty control blocks or calls which should not be there
Currently, the filtering of call expressions and purging empty control blocks (
e.g. loops or conditional statements),
within which no interesting calls were included in the diagram is not perfect.
In case the regular namespaces filter
is not enough, it is useful to add also a paths filter, which will only
include participants and call expressions
from files in a subdirectory of your project, e.g.:
include:
namespaces:
- myproject
paths:
- src