CodeDepencyGraph

Instructions to make a code dependency graph.

These instructions describe how to run code-dep-graph, a tool that creates a graph of compile time dependence relationships between packages. The tools starts from the current working directory and works down from there. This tool has it's own notion of a package which is slightly different from other notions in common use. Roughly speaking any directory that contains code is its own package; if a directory has a subdirectories named inc and src, those names are elided. Here are some examples of packages taken from a recent version of Mu2e Offline: RecoDataProducts, ExtinctionMonitorFNAL/Utilities, ExtinctionMonitorFNAL/Reconstruction; the slash is part of the package name.

code-dep-graph walks the directory tree looking for C++ header and implementation files. It parses these files and extracts the #include directives. It builds the package dependency tree based on the #include directives.

Here is a cookbook to run the tool.

1. Make a clean working directory and cd to it.
mkdir work cd work
2. Clean checkout. This will be a throw-away copy of the code; don't use a copy with changes you intend to commit and push.
git clone ssh://p-mu2eofflinesoftwaremu2eoffline@cdcvs.fnal.gov/cvs/projects/mu2eofflinesoftwaremu2eoffline/Offline.git
3. Optional but highly recommended: remove files that unnecessarily complicate the picture. This includes all modules, which must sit at the top of the dependence hierarchy, and a few directories like Sandbox. This step deletes files, hence the warning about using a throw-away copy.
find Offline -name \*_module.cc -exec rm -f {} \; find Offline -name \*_source.cc -exec rm -f {} \; rm -r Offline/Sandbox Offline/HelloWorld rm -rf Offline/.git
4. Discover what versions of cetbulidtools are available. Choose a recent version. As of October 22, 2013 version v7_04_00 is known to work.
ups -aK+ cetbuildtools
5. Setup an appropriate version of cetbuildtools.
setup cetbuildtools v7_04_00
6. Run code-dep-graph from the top directory of Offline. The output is a dot file that contains dependency pairs; you can read this file with a text editor; the syntax is intuitive and you can find documentation online - google "dot file format". The text represents a directed graph. The log file is verbose and I have not found anything useful in it if the code is working properly.
cd Offline time ${CETBUILDTOOLS_DIR}/bin/code-dep-graph -v -o ../trimmed.dot >& ../trimmed.log
7. The previous step should take several minutes on an unloaded machine. The names trimmed.dot and trimmed.log are illustrative; they have no meaning.
8. Do a transitive reduction of the directed graph and create the reduced directed graph as a png file. A transitive reduction means the following. Suppose that we have 3 packages, A, B, C; also suppose the A depends directly on both B and C, while B depends on C. In the transitive reduction the dependence of A on C will be elided since it is implied by the dependence of A on B. If you don't do this step on Offline, the resulting graph is not readable.
tred trimmed.dot | dot -Tpng -o trimmed.png
9. If there are loops in the directed graph tred will issue a diagnostic like the following:
warning: %1 has cycle(s), transitive reduction not unique cycle involves edge RecoDataProducts -> BTrkData
10. Additional comments on tred
    1. If there are several loops in the graph, tred will only report the first one that it finds. So you need to fix the problem and iterate until there are no more diagnostics.
    2. The packages mentioned in the "involves the edge" can sometimes be a few hops from the real problem.
    3. The output of tred is just another dot file; you can capture it to a file instead of piping to dot.
    4. The commands dot and tred are part of the graphviz package: https://www.graphviz.org
    5. If you prefer a format different from png, dot has a lot of options: dot --help
    6. Hint: when you are in the process of discovering loops, you may be able to hand edit the dot file to remove the offending edge and rerun tred; this is faster than rerunning code-dep-graph.
11. View the graph using your favorite png browser.
12. Comments on the graph.
    1. If a package name contains a underscore character code-dep-graph gets confused and forgets to draw a box around it.
    2. Services are colored blue.
13. Optional: create the visualization of the graph without transitive reduction - but it's impossible to read
dot trimmed.dot -Tpng -o trimmed_notred.png