[patrickmscott]

all_files carried the LLVM distribution's whole bin/, and every file group carried the binary behind every wrapped tool (clang-tidy, clangd, ...), so remote executors staged gigabytes of unused inputs per action.

• all_files takes the wrapped tools' binaries instead of all of bin/.
• Other file groups keep only the wrappers and runtime_shared_libraries; each already lists the binaries its actions run.
• Darwin-to-Darwin archive actions take llvm-libtool-darwin, which the libtool wrapper runs.

[helly25]

@patrickmscott please follow the instructions to sign your commits: https://github.com/NHSDigital/software-engineering-quality-framework/blob/main/practices/guides/retrospective-commit-signing.md

[patrickmscott]

@helly25 updated

[helly25]

@patrickmscott: The PR indeed reduces unnecessary compiler/linker inputs introduced by #914 while retaining the shared libraries required on remote workers.

Correctness issue remain: all_files drops llvm-as, although as_files still includes it through the distribution’s :as filegroup. Bazel requires all_files to be a superset of the action-specific sets. Ordinary compilation uses Clang’s integrated assembler, so existing tests can pass despite this omission. Changed input sets (https://github.com/bazel-contrib/toolchains_llvm/blob/01796b567d37d6e649f594d6d7129e9d12920c4a/toolchain/internal/configure.bzl#L900-L916), Bazel’s contract (https://bazel.build/reference/be/c-cpp#cc_toolchain.all_files).

The rest of the change looks good:

• Compiler inputs retain Clang and the module dependency scanner.
• Shared runtime libraries remain available.
• The Darwin archiver change correctly supplies llvm-libtool-darwin, matching the existing tool selection.
• All 99 CI checks passed.
• The linker override remains available through both input paths:
  • Downloaded linkers such as mold: included through bin/linkers/ with directory inputs, or explicitly through linker_file with legacy inputs.
  • auto and absolute paths: resolution is unchanged; these still require the linker installed on the execution machine.
• The PR’s mold, native-linker, feature-override, and Darwin ThinLTO CI tests all passed. The llvm-as omission I identified is separate from linker overrides.

Remaining work:

1. Restore the all_files superset invariant, preferably by composing it from the action-specific groups.
2. Add regression coverage checking both required inputs and exclusion of unrelated binaries.
3. Provide one remote execution—or equivalent isolated staging—result using shared-library LLVM tools. The PR adds no tests demonstrating that scenario.

[patrickmscott]

@helly25 I addressed all_files and added a test