Skip to content

Windows build fixes for Bazel ≥ 9.0.1#22

Open
walter-zeromatter wants to merge 50 commits into
hermeticbuild:mainfrom
walter-zeromatter:user/wgray/win-fixes
Open

Windows build fixes for Bazel ≥ 9.0.1#22
walter-zeromatter wants to merge 50 commits into
hermeticbuild:mainfrom
walter-zeromatter:user/wgray/win-fixes

Conversation

@walter-zeromatter
Copy link
Copy Markdown

@walter-zeromatter walter-zeromatter commented May 11, 2026

  • bootstrap_process_wrapper.cc: rewrite @response-files to convert the rustc --sysroot= value to its 8.3 short form via GetShortPathNameA before spawning rustc. The MSVC link.exe shipped with VS 2022 BuildTools lacks longPathAware, so the resolved <sysroot>/lib/rustlib/.../libstd-*.rlib paths overflow MAX_PATH (260) and fail with LNK1181. Also restrict get_output_base to follow external/ only when it's a symlink (Bazel ≥ 9.0.1 made it a real directory on Windows).

  • options.rs: same external-is-real-dir fix for the regular process_wrapper. Falls back to deriving output_base from current_dir's grandparent when external/ is not a symlink.

  • main.rs: skip *.rcgu.{o,bc,bc.z} codegen intermediates when building the unified -Ldependency dir, and tolerate NotFound on the copy fallback. Pipelined sibling actions can garbage-collect those temporaries between read_dir and the hardlink/copy.

  • cargo_build_script_runner/bin.rs: track exec_root symlinks for cleanup only when symlink_if_not_exists actually created the symlink (was hitting DirectoryNotEmpty trying to remove_dir real dirs created by runfiles staging, e.g. .github). Also skip .git from execroot symlinks so ring ≤ 0.17.x doesn't enter its "local hacking" mode that hardcodes ./target/tools/windows/nasm/nasm.

  • cargo_build_script_runner/cargo_manifest_dir.rs: use symlink_metadata in remove_symlink so dangling directory symlinks (Bazel's local-spawn-runner.* sandboxes cleaned up before this runs) get remove_dir'd instead of failing remove_file with ACCESS_DENIED.

UebelAndre and others added 30 commits May 11, 2026 10:25
Add support for tier 3 targets bpfeb-unknown-none and bpfel-unknown-none
(see

https://github.com/rust-lang/rust/blob/f5e2df7/src/doc/rustc/src/platform-support.md?plain=1#L311-L312).

This is modeled after bazelbuild#3507
and
should probably be updated if/when
bazelbuild/platforms#131 is merged.

(please use rebase merge when landing this as the proper commit message
is in the commit, rather than the PR description)

/cc @avrabe
…hollow rlibs: the RustcMetadata action runs rustc to completion with -Zno-codegen, emitting a .rlib archive. This approach mirrors the one used by buck2 and avoids needing to kill rustc mid-output in order to produce metadata.

While not fixing problems with SVH mismatches when non-determinism, this does simplify the codepath and uses a production tested technique that doesn't have any of the dangers associated with killing the rustc process while it's still active.
Port the sharding wrapper feature from bazelbuild#3774 into the hermeticbuild fork. The implementation wraps rust_test executables when experimental_enable_sharding is set while keeping rustc_compile_action's existing provider-list API for internal and extension callers.

rust_test now scans the returned providers to replace DefaultInfo for the wrapper, so extensions such as prost and wasm-bindgen continue to consume rustc_compile_action without API churn.

Co-authored-by: Brian Duff <bduff@linkedin.com>

Co-authored-by: Codex <noreply@openai.com>
Rustc emits GNU-like Windows staticlibs as lib<crate>.a, but rules_rust was stripping the lib prefix for all Windows non-rlib library outputs. Keep the prefix for staticlib outputs when the target ABI is gnu or gnullvm so declared outputs match rustc.
dzbarsky and others added 19 commits May 11, 2026 10:25
Keep cargo build script inputs available to their direct dependent without propagating them through transitive Rustc inputs. compile_data remains reserved for files that need to stay available to downstream consumers, such as generated link inputs.

Update the existing cargo_build_script propagation test to distinguish the direct-dependent case from the transitive case.

Co-authored-by: Codex <noreply@openai.com>
)

* When '--strategy=Rustc=local' is used with pipelined compilation, rustc's -Ldependency scan can pick up the _meta.rlib alongside the full '.rlbi', producing undefined-symbol link errors. To avoid this, emit the hollow metadata rlibs into a '_meta/' subdirectory, and add an explicit link to that directory for any metadata-consuming actions. (bazelbuild#17)

* Add RUSTC_BOOTSTRAP guardrail for -Zno-codegen pipelining

Injects RUSTC_BOOTSTRAP=1 + -Zallow-features= (empty list) on the metadata
and full compile actions when pipelined_compilation is enabled on a
stable/beta toolchain. The bootstrap env is required for -Zno-codegen on
non-nightly rustc and must match across both actions for SVH compatibility.
The empty allow-features list prevents the bootstrap env from silently
unlocking #![feature(...)] in user code.

Nightly toolchains skip both injections: unstable features are already
allowed, and the guardrail would break legitimate #![feature(...)] usage.

Escape hatch: if the user sets RUSTC_BOOTSTRAP in rustc_env / shell env or
passes -Zallow-features=... via rustc_flags / extra_rustc_flag, rules_rust
treats their configuration as authoritative and skips both injections.

Tests:
- //test/pipelining_bootstrap_gate: manual target that #![feature(trait_alias)]
  to validate E0725 fires on stable under the guardrail.
- //test/unit/pipelined_compilation: guardrail_{baseline,user_env,user_flag,
  space_form,global_env,global_flag}_optout_test variants.
- bootstrap_process_wrapper.cc: rewrite `@response-files` to convert the
  rustc `--sysroot=` value to its 8.3 short form via `GetShortPathNameA`
  before spawning rustc. The MSVC `link.exe` shipped with VS 2022
  BuildTools lacks `longPathAware`, so the resolved
  `<sysroot>/lib/rustlib/.../libstd-*.rlib` paths overflow MAX_PATH (260)
  and fail with LNK1181. Also restrict `get_output_base` to follow
  `external/` only when it's a symlink (Bazel ≥ 9.0.1 made it a real
  directory on Windows).

- options.rs: same external-is-real-dir fix for the regular
  process_wrapper. Falls back to deriving output_base from `current_dir`'s
  grandparent when `external/` is not a symlink.

- main.rs: skip `*.rcgu.{o,bc,bc.z}` codegen intermediates when building
  the unified `-Ldependency` dir, and tolerate `NotFound` on the copy
  fallback. Pipelined sibling actions can garbage-collect those
  temporaries between `read_dir` and the hardlink/copy.

- cargo_build_script_runner/bin.rs: track exec_root symlinks for cleanup
  only when `symlink_if_not_exists` actually created the symlink (was
  hitting DirectoryNotEmpty trying to `remove_dir` real dirs created by
  runfiles staging, e.g. `.github`). Also skip `.git` from execroot
  symlinks so `ring` ≤ 0.17.x doesn't enter its "local hacking" mode
  that hardcodes `./target/tools/windows/nasm/nasm`.

- cargo_build_script_runner/cargo_manifest_dir.rs: use `symlink_metadata`
  in `remove_symlink` so dangling directory symlinks (Bazel's
  `local-spawn-runner.*` sandboxes cleaned up before this runs) get
  `remove_dir`'d instead of failing `remove_file` with ACCESS_DENIED.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@dzbarsky
Copy link
Copy Markdown
Member

@ArchangelX360 can you take a look at these and try them out?

@walter-zeromatter
Copy link
Copy Markdown
Author

Triage: which of these patches are MSVC link.exe–specific vs. Bazel-9.0.1/pipelining-specific

Tested 2026-05-14 on Windows by swapping in the toml2json-style hermetic toolchain (ArchangelX360's toolchains_llvm_bootstrapped fork + windows_support@0.1.1 + rules_cc PR bazelbuild#561, --repo_env=BAZEL_DO_NOT_DETECT_CPP_TOOLCHAIN=1, --extra_toolchains=@llvm//toolchain:all). With that toolchain, C builds go through clang and linking through lld-link / rust-lld.exe — MSVC link.exe is never invoked. Full bazel build sdk:sdk succeeded with the 8.3-path conversion reverted.

So the patches split into two independent groups:

MSVC-link.exe-specific (obsoleted by hermetic LLVM/lld):

  • bootstrap_process_wrapper.cc: the to_short_path / make_absolute / rewrite_response_file / rewrite_response_files logic and the <fstream> include. Premise was VS 2022 BuildTools' link.exe lacking longPathAware; lld-link and rust-lld handle long paths fine, so this is unneeded under a hermetic toolchain. Still needed for anyone resolving the auto-detected MSVC cc_toolchain.

Bazel ≥9.0.1 / pipelining-specific (still required regardless of C++ toolchain):

  • bootstrap_process_wrapper.cc: get_output_base — only follow external/ when it's a symlink. Bazel ≥9.0.1 made it a real directory on Windows; canonicalizing external/.. otherwise yields the execroot, not the output base.
  • util/process_wrapper/options.rs: same external/-is-real-dir fallback for the regular process_wrapper, deriving output_base from current_dir's grandparent.
  • util/process_wrapper/incremental_cache.rs: find_output_base preferring an ancestor whose basename is not execroot. Bazel writes DO_NOT_BUILD_HERE in both the real output_base and execroot/ on Windows ≥9.0.1; without this, the path doubles to …/execroot/execroot/_main/….
  • util/process_wrapper/main.rs: skip *.rcgu.{o,bc,bc.z} codegen intermediates and tolerate NotFound on the copy fallback. Pipelined sibling actions GC those temporaries between read_dir and hardlink/copy.
  • cargo/private/cargo_build_script_runner/bin.rs: track exec_root symlinks for cleanup only when symlink_if_not_exists actually created them (was failing remove_dir on real dirs like .github from runfiles staging); skip .git so ring ≤ 0.17.x doesn't enter its "local hacking" hardcoded-nasm-path mode.
  • cargo/private/cargo_build_script_runner/cargo_manifest_dir.rs: symlink_metadata in remove_symlink so dangling directory symlinks (Bazel's local-spawn-runner.* sandboxes cleaned up before this runs) get remove_dir'd instead of failing remove_file with ACCESS_DENIED.

If reviewers want, I can split the first bullet into its own PR and rebase this one to contain only the Bazel-9.0.1 fixes. Open question on whether the 8.3 path conversion is still worth keeping behind a feature flag for users on auto-detected MSVC.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

8 participants