Resolve circular dependencies

[rrsettgast]

Describe the issue At some point we were having problems running nsightcompute when dynamically linking our libraries, so we switched to static linking. This is undesirable for two reasons:

1. GEOSX is LGPL, so we should be dynamically linking, and
2. the size of the statically linked unit tests are very large. Each executable contains all the static objects. Not good

Proposed cleanup Switch back to dynamic linking as default. I think that this requires us to clean up our circular dependencies once and for all. We should maintain the option to statically link in case there are any problems later.

[bmhan12]

On the note of circular dependencies, I ran the tools we added last time we looked at circular dependencies in #1352:

Graphviz

Running config-build.py with the --graphviz option (code link) produces the graph of the dependencies according to CMake.

Note: In order to get this graph, I built GEOSX with shared libraries on and object libraries turned off. For whatever reason, I couldn't get CMake to detect the object libraries registered through blt to show up.

By CMake's view, we don't have any circular dependencies currently.

circular_dependencies.sh

This script produces two files: circular_dependencies.txt and dependency_matrix.txt (Excel view).

circular_dependencies.txt shows what header includes exist between two components.

Currently, there are 9:

1. common & codingUtilities
2. common & mesh
3. codingUtilities & mainInterface
4. constitutive & mesh
5. constitutive & fileIO
6. constitutive & physicSolvers
7. mesh & finiteElement
8. finiteVolume & physicsSolvers
9. physicsSolvers & mainInterface

dependency_matrix.txt shows if a component contains headers to another component (yes/no)

find_cycles.py

I did the following to run this script:

ml ninja (load the module)

./scripts/config-build.py -hc host-configs/LLNL/quartz-clang\@10.0.0.cmake -ecc -G Ninja

python3 ./scripts/find_cycles.py -c build-quartz-clang\@10.0.0-debug/compile_commands.json

Output on quartz: find_cycles_quartz_output.txt

The output shows 42 cycles, but most of them look to be outside GEOSX and in individual third-party libraries. Notably, the first cycle is a header including itself.

[TotoGaz]

This is a really fruitful analysis, thx @bmhan12

In a nutshell, is it fair to state that we cannot really find real C++ cycles, but that our C++ architecture does not respect our cmake architecture?

PS: There is an option in find_cycles .py to remove some directories. PPS: find_cycles.py could be improved to add the cmake clustering into the game! It's missing this option to find this other kind of cycles.

[rrsettgast]

The current picture from cmake not looking good: