bazel-zig-cc and llvm

bazel-zig-cc has a performance issue when compiling many files. This repository reproduces that.

The test

Compiles 64 small binaries with bazel-zig-cc and llvm14. The tests were run on an x86_64 8-core machine running Ubuntu 22.04.

Steps to reproduce

Baseline is llvm14: 12-13 seconds:

bazel --batch clean; bazel --batch build --extra_toolchains=@llvm_toolchain//:cc-toolchain-x86_64-linux ...
INFO: Elapsed time: 12.454s, Critical Path: 1.11s

zig cc without the sandbox. Which means different invocations of zig cc will see that all files in zig_lib_dir are the same file. 17 seconds:

bazel --batch clean; bazel --batch build --spawn_strategy=local  --platforms=@zig_sdk//libc_aware/platform:linux_amd64_gnu.2.28 ...
INFO: Elapsed time: 17.021s, Critical Path: 1.67s

zig cc plain: 61 seconds. Sandbox is on a real disk, which means it will take even longer to re-hash all its dependencies:

bazel --batch clean; bazel --batch build --platforms=@zig_sdk//libc_aware/platform:linux_amd64_gnu.2.28 ...
INFO: Elapsed time: 61.128s, Critical Path: 8.35s

zig cc with hardlinks: --config=hermetic-sandbox. This uses hardlinks to zig_sdk instead of symlinks:

bazel --batch clean; bazel --batch build --config=hermetic-sandbox --platforms=@zig_sdk//libc_aware/platform:linux_amd64_gnu.2.28 ...
INFO: Elapsed time: 45.886s, Critical Path: 9.56s

llvm with sysroot (this sysroot is about half the size of zig):

bazel --batch clean; bazel --batch build --config=hermetic-sandbox --extra_toolchains=@llvm_toolchain_with_sysroot//:cc-toolchain-x86_64-linux ...
INFO: Elapsed time: 25.644s, Critical Path: 3.03s

Flame graphs and discussion

Flame graphs are in results/. As of 2022-12-11 most of the overhead comes from creating and deleting the sandboxes. Time for a new sandboxfs!