Code Linting#

Clang-Tidy#

Running Tidy#

Use the aggregate make tidy target to lint the C++ codebase.

  1. Configure the repo (tidy targets are always generated):

./configure --cmake-generator=Ninja

  1. Run clang-tidy across all sources:

make tidy -j $(nproc)

Running Tidy With CUDA (Optional)#

When you also want to lint CUDA kernels (.cu files):

  1. Configure with CUDA and clang-as-CUDA so tidy analyzes .cu files with a clang-compatible toolchain:

./configure --with-cuda \
            --with-cuda-dir="${CONDA_PREFIX}" \
            --with-fake-fatbins-for-tidy \
            --with-cudac=clang++ \
            --CUDAFLAGS "-O3" \
            --cuda-arch=90

  1. Run all tidy targets (C++ and CUDA) via the aggregate target:

make tidy -j $(nproc)

CUDA options#

  • --with-cuda: enable CUDA toolchain so .cu files are analyzed.

  • --with-cuda-dir=...: point to the CUDA toolkit when using a conda toolchain.

  • --with-cudac=clang++: use clang for CUDA (required for CUDA tidy).

  • --with-fake-fatbins-for-tidy: generate stub fatbins so host code that embeds fatbins can compile under tidy without building fatbins/Legion.

  • --CUDAFLAGS "-O3": minimal CUDA flags for tidy’s compile database.

  • --cuda-arch=...: target GPU architecture for device-only analysis. clang-tidy supports only a single architecture at a time – pass one (e.g., 90).

NOLINT guidance#

  • Prefer narrow, inline suppressions (// NOLINT(<check>)) at the exact diagnostic site.

  • Inside macro definitions with line continuations, use block comments (/* NOLINT(...) */) before the trailing \ or wrap the macro with a narrow NOLINTBEGIN/END for bugprone-macro-parentheses.