Ideas for things to do:

- more projects (compiler-rt's asan/ubsan/msan/tsan, lldb)
  - once there are more projects, have an llvm_enable_projects arg, modeled
    after llvm_targets_to_build in the GN build
  - a check-all build target that runs test of all projects
- more targets (AMDGPU ARC AVR MSP430 XCore)
  - example: https://reviews.llvm.org/D56416
- investigate feasibility of working `gn check`

- "optimized tblgen" mode
  - either just always build tablegen and support with opt config
  - or use opt toolchain and build tablegen twice in debug builds, like cmake

- cross builds using GN's toolchain feature

- move clang_tablegen into lib/ for private files
- add dead code stripping
- move run_tablegen.py from build to tablegen folder
- figure out why -Iclang/Support gets added so often
- plugin() template with working rpath, exports thingers
  - then port clang_build_examples and enable by default so that clang
    plugin tests run by default
- plugin_host() template