Bazel: align compiler and Windows toolchain behavior with Make

[napetrov]

Description

This PR closes seven reproducible compiler/toolchain parity gaps found while auditing the Bazel path against the existing Make behavior for issue #3530.

The changes are intentionally kept as seven focused commits. Together they align x86-64 ISA selection across GCC, Clang, ICX, and the MSVC fallback; fix Windows CPU detection and CRT selection; derive the DPC++ archiver from the selected compiler; and make the Windows nightly BAZEL_SH requirement explicit and fail-fast.

This is progress toward #3530, but it does not claim that every Make-only platform or feature is ready for deprecation.

What changed

ISA parity

• Use -march=skylake-avx512 for GCC SKX objects in both Make and Bazel.
• Define the missing Clang CPU matrix:
  • SSE2: -march=nocona
  • AVX2: -march=haswell
  • AVX-512: -march=skylake-avx512
• Propagate ISA flags to materialized MSVC fallback actions while retaining Make parity:
  • SSE2: no /arch option
  • AVX2: /arch:AVX2
  • SKX/AVX-512 dispatch objects: /arch:AVX2

Toolchain/runtime behavior

• Derive llvm-ar from the selected DPC++ compiler path instead of slicing the unrelated host compiler path; fail configuration when the companion archiver is absent.
• Build the Windows CPU detector with MSVC and use the .exe output name expected by Windows.
• Select the Windows CRT through a toolchain feature:
  • optimized/non-debug actions: /MD
  • debug actions: /MDd
• Remove the former unconditional -MD from the shared Windows ICX flags.

Windows nightly reproducibility

• Set step-scoped BAZEL_SH to Git for Windows Bash.
• Fail before Bazel with a clear diagnostic if that executable is missing.
• Document that Windows Bazel build, test, and analysis commands require BAZEL_SH.

Commit structure

1. build: enable AVX-512 for GCC skx targets
2. bazel: define Clang CPU ISA flags
3. bazel: derive DPC++ archiver from compiler path
4. bazel: compile CPU detector with MSVC on Windows
5. bazel: select Windows CRT by build mode
6. bazel: add MSVC CPU ISA flags
7. ci: require Bazel shell on Windows nightly

The CRT removal/addition is atomic in commit 5; the workflow and documentation change are paired in commit 7.

Validation

Linux — exact clean PR head

Validated on clean head f7ab546fb85dd1c41098f4132c01060ecd7fbced with Bazel 9.2.0.

• GCC 15.2, Clang 21.1.8, and ICX 2026.0 ISA action matrices passed.
• SSE2/AVX2/AVX-512 actions materialized the expected compiler options.
• GCC, ICX, and ICPX archive actions selected:
  • /opt/intel/oneapi/compiler/2026.0/bin/compiler/llvm-ar
• Full host release:

PATH=/opt/intel/oneapi/compiler/2026.0/bin:$PATH \
CC=icx bazelisk \
  --output_base=<isolated-output-base> \
  build //:release --release_dpc=false --cpu=all --jobs=2

Result: rc 0, 3,188 actions, 17m50s.

Produced-release gates also passed:

• ISA coverage: 592 dispatch symbols for each SSE2/AVX2/AVX-512 tier
• release structure and SONAME checks
• package metadata: 9 passed, 0 failed
• MKL linkage probe: no bundled MKL archive objects or exported mkl_ symbols

Windows — cumulative source-equivalent patch

Validated on Windows Server 2025 with:

• MSVC 19.44.35227
• Intel oneAPI DPC++/C++ Compiler 2026.0.0
• Bazel 9.2.0
• Git for Windows Bash 5.3.9

Source integrity checks passed for the exact eight-file cumulative diff: git diff --check, reverse patch application, file inventory, and SHA-256 preservation.

Materialized MSVC parameters:

Dispatch tier	params	expected /arch	violations
SSE2	390	none	0
AVX2	390	/arch:AVX2	0
SKX/AVX-512	390	/arch:AVX2	0

Materialized CRT parameters:

mode	/MD	/MDd
opt	1,533	0
dbg	0	1,533

Additional Windows gates:

• generated CPU detector compiled and executed successfully; output: avx512
• valid BAZEL_SH guard: rc 0
• deliberately missing BAZEL_SH guard: rc 1 with the expected diagnostic
• broad MSVC host build of //examples/daal/cpp:library_version_info_host: rc 0, 1,933 actions
• full ICX/DPC release:

bazelisk --output_user_root=<isolated-root> build //:release \
  --config=release-dpc --cpu=avx2 --jobs=2 \
  --disk_cache= --remote_upload_local_results=false --verbose_failures

Result: rc 0, 3,294 actions, 49m43s.

Earlier Windows LNK1180/server-reset attempts were traced to a user-level C:\b-cache upload duplicating the multi-gigabyte onedal_core.lib archive until the disk reached zero free bytes. The same source passed after removing only that disposable cache and explicitly disabling cache upload. No compiler/linker source failure remained.

Current main applicability

All seven commits were cherry-picked, in order and without conflicts, onto main@9d920262384e681331fca9c3c4ed481e477ae81b in a disposable checkout.

• clean status
• unchanged cumulative diff: 8 files, 69 insertions, 15 deletions
• workflow YAML assertions passed
• //:release --release_dpc=false --cpu=all --nobuild passed with ICX/Bazel 9.2.0: rc 0, 4,641 configured targets

Independent review

Two independent passes reviewed the cumulative diff:

• compiler/toolchain semantics: approve-with-nits; critical 0, should-fix 0
• CI/integration: approve-with-nits for opening one PR; no code-level blocker

The recommendation was to keep this as one PR with the seven logical commits rather than squash or split it.

Draft / merge gates

This PR intentionally starts as a draft.

Before marking ready/merging:

• Run and inspect the actual GitHub-hosted Windows Bazel workflow for this pushed head.
• Confirm all required CI checks are green.
• Resolve commit sign-off/DCO requirements if reported by CI (no DCO failure was reported).
• Optional follow-up: an additional exact-head Windows opt/debug runtime smoke was not rerun manually after the validation VM account lock; exact-head hosted Windows CI passed.

Scope boundary

This series does not by itself complete global Make deprecation. Remaining broader items include non-x86 platforms, macOS, OpenRNG and other Make-only switches, real DPC CPU/GPU execution coverage, sanitizer execution, downstream-consumer gates, ABI/release-comparison hardening, performance/build-time/hermeticity gates, and Bazel-backed nupkg/documentation deliverables.

---

Checklist

Completeness and readability

• Hard-to-understand behavior is commented, including the intentional MSVC SKX-to-/arch:AVX2 mapping.
• Windows Bazel setup documentation is updated.
• No commit sign-off/DCO failure was reported by CI.
• The series applies to the current base without conflicts.

Testing

• The changes were tested extensively on Linux and Windows as described above.
• Required hosted CI passed on the exact PR head.
• Validation covers materialized compiler parameters, full host/release builds, CPU detection, workflow guards, package structure, and metadata.

Performance

• No algorithm, numerical path, or dispatch policy is changed; this PR corrects compiler/toolchain selection for existing dispatch tiers. Runtime benchmark changes are not expected.
• No new measurable algorithm functionality is introduced, so no benchmark-suite extension is required.

[Copilot]

Pull request overview

Aligns Bazel’s compiler/toolchain behavior with the existing Make-based contract to reduce parity gaps during the ongoing Make→Bazel migration work (issue #3530), with a focus on x86-64 ISA flag matrices, Windows toolchain/runtime behavior, and Windows CI reproducibility.

Changes:

• Updates GCC/Clang ISA flag selection in Bazel (and GCC SKX in Make) to ensure distinct SSE2/AVX2/AVX-512 compilation targets (e.g., -march=skylake-avx512 for AVX-512).
• Improves toolchain correctness: derives the DPC++ archiver from the DPC++ compiler path (fail-fast if missing), fixes Windows CPU detector compilation/output naming, and selects Windows CRT flags (-MD vs -MDd) via a dedicated toolchain feature.
• Ensures Windows Bazel nightly builds are reproducible by explicitly setting and validating BAZEL_SH, and updates Windows Bazel setup documentation accordingly.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
dev/make/compiler_definitions/gnu.32e.mk	Updates GCC SKX target to -march=skylake-avx512 to match the Bazel ISA matrix.
dev/bazel/toolchains/cc_toolchain_lnx.bzl	Derives llvm-ar from the selected DPC++ compiler location and fails configuration when absent.
dev/bazel/toolchains/cc_toolchain_config_win.bzl	Adds a toolchain feature to select /MD vs /MDd behavior (via -MD/-MDd) based on build mode.
dev/bazel/README.md	Clarifies that Windows Bazel build/test/analysis require BAZEL_SH (not just bazel test).
dev/bazel/flags.bzl	Aligns GCC/Clang ISA flags (including AVX-512) and removes unconditional Windows -MD from shared ICX flags.
dev/bazel/config/config.bzl	Builds the CPU detector with MSVC on Windows and uses the expected .exe output name.
dev/bazel/cc/compile.bzl	Propagates MSVC /arch settings to CPU-tier compilation actions to match Make’s dispatch-object behavior.
.github/workflows/nightly-build.yml	Sets and validates BAZEL_SH explicitly for the Windows nightly Bazel release step (fail-fast if missing).