mwhicks1 / 3c-actions Goto Github PK

I think this is what's happening anyways. See this failed workflow run where the Test 3C job fails with ninja: error: loading 'build.ninja': No such file or directory.

Some ideas to make the YAML generation in generate-workflow.py better after #7:

• From John: "The python script could include a call to yamllint to sanity check the generated workflow. Easy enough to run manually, but it might catch trivial errors in future changes." This would require everyone who runs generate-workflow.py to install yamllint, so we'd have to think about whether it's worth it.
• We could generate the output using a YAML library rather than a bunch of manual string manipulation. (Or maybe there is even a library specifically for generating GitHub workflow files?) If there are hard-coded chunks of the workflow file that are easier to understand in YAML format than as hard-coded calls to the YAML library to build the corresponding data structures, we can probably just have the YAML library parse hard-coded YAML strings.

From #13 (review):

It looks like this is skipping the test.c files (which contain duplicate mains) without an explicit --skip flag. This behavior is fine for now since the duplicate mains would cause 3c to raise an error, but we will probably want to include them when correctcomputation/checkedc-clang#505 is merged.

The reason these files are not being included now appears to be that they are not built by the default ./configure && make, so Bear doesn't add them to the compilation database. We'd have to figure out the correct arguments to ./configure and/or make to build the files.

#33 added a ton of duplicate code to generate-workflow.py for different versions of the source code for each benchmark (original, manual port, C3 reversion of manual port). When we have time, we should clean this up (unless we decide to delete the extra versions anyway once we're done with the paper).

With #7, the workflow takes over an hour to run. Some ideas to speed it up:

• (Done in #10) Use a release build of 3C and Clang, or RelWithDebInfo if having the debug info is useful (e.g., if it shows up in stack traces for assertion failures: I'd need to re-test that).
• (Done in #15) For the CMake projects, replace make with ninja. For the non-CMake projects, manually pass the appropriate flags to make for a parallel build.
• Generate the compilation database without actually building the unconverted code.
  • For the CMake projects, we should just be able to use CMAKE_EXPORT_COMPILE_COMMANDS, although I don't know if convert_project handles this correctly; it may need minor changes.
  • For the non-CMake projects, we could try to use Bear's "wrapper" mode that uses a fake compiler that doesn't actually perform compilation, but this appears to require a newer version of Bear than is currently available via Ubuntu; we'd have to manually build our own version on gamera.