chromium-include-cleanup

Scripts to help guide cleanup of #include lines in a codebase, using clangd

Scripts

• apply_include_changes.py - Apply include changes to files in the source tree
• filter_include_changes.py - Filter include changes output
• post_process_compilation_db.py - Post-process the clang compilation database for analysis
• set_edge_weights.py - Set edge weights in include changes output
• suggest_include_changes.py - Suggests includes to add and remove

Prerequisites

To use these scripts, you'll need:

• A release of clangd which has "IncludeCleaner" with support for missing includes (17.0.0+)
• The full output of //tools/clang/scripts/analyze_includes.py, see discussion on the mailing list for how to generate it
• A compilation database for clangd to use, which can be generated with gn gen . --export-compile-commands in the Chromium output directory
  • The generated compile_commands.json should be post-processed with the post_process_compilation_db.py script for best results

Install Dependencies

$ pip install -r ~/chromium-include-cleanup/requirements.txt

Patching clangd

To get suggestions for includes to add, and other tweaks, clangd needs to be patched with the patches in clangd_patches and built from source.

clangd Configuration

You need to enable MissingIncludes and UnusedIncludes diagnostics in a clangd config file:

Diagnostics:
  MissingIncludes: Strict
  UnusedIncludes: Strict

Finding Unused Includes

These instructions assume you've already built and processed the build log with //tools/clang/scripts/analyze_includes.py, if you haven't, see the link above under "Prerequisites". It assumes the output is at ~/include-analysis.js, so adjust to taste.

This also assumes you have clangd on your $PATH.

$ cd ~/chromium/src/out/Default
$ gn gen . --export-compile-commands
$ python3 ~/chromium-include-cleanup/post_process_compilation_db.py compile_commands.json > compile_commands-fixed.json
$ mv compile_commands-fixed.json compile_commands.json
$ cd ../../
$ python3 ~/chromium-include-cleanup/suggest_include_changes.py --compile-commands-dir=out/Default ~/include-analysis.js > ~/unused-edges.csv
$ python3 ~/chromium-include-cleanup/set_edge_weights.py ~/unused-edges.csv ~/include-analysis.js --config ~/chromium-include-cleanup/configs/chromium.json > ~/weighted-unused-edges.csv

Another useful option is --filename-filter=^base/, which lets you filter the files which will be analyzed, which can speed things up considerably if it is limited to a subset of the codebase.

Edge weights are set in a separate script to allow quick iteration, since suggest_include_changes.py takes many hours to run. The default metric for edge weights pulls the "Added Size" metric from the include analysis output. This means new weights can be easily be applied to the output of suggest_include_changes.py by downloading the latest hosted include analysis output at https://commondatastorage.googleapis.com/chromium-browser-clang/include-analysis.js, but mileage may vary since you're combining output from your local build and the hosted build.

Performance

For a full codebase run of the suggest_include_changes.py script on Ubuntu, it takes 7 hours on a 4 core, 8 thread machine. clangd is highly parallel though, and the script is configured to use all available logical CPUs, so it will scale well on beefier machines.

Current Limitations

Currently the suggest_include_changes.py script has problems with suggesting includes to remove when the filename in the #include line does not match the filename in the include analysis output, which could happen for includes inside third-party code which is including relative to itself, not the source root.

When suggesting includes to add, clangd will sometimes suggest headers which are internal to the standard library, like <__hash_table>, rather than the public header. Unfortunately these cases can't be disambiguated by this script, since there's not enough information to work off of.

Accuracy of Output

These scripts rely on clangd and specifically the "IncludeCleaner" feature to determine which includes are unused, and which headers need to be added. With the Chromium codebase, there are many places where clangd will return false positives, suggesting that an include is not used when it actually is. As such, the output is more of a guide than something which can be used as-is in an automated situation.

Known situations in Chromium where clangd will produce false positives:

• When an include is only used for a friend class declaration
• When the code using an include is inside an #ifdef not used on the system which built the codebase
• Macros in general are often a struggle point
• Umbrella headers
• Certain forward declarations seem to be flagged incorrectly as the canonical location for a symbol, such as "base/callback_forward.h"
• Forward declarations in the file being analyzed
  • clangd won't consider an include unused even if forward declarations exist which make it unnecessary
  • clangd will still suggest an include even if a forward declaration makes it unnecessary