[conrade-ctc]

CppInterOp's Bazel build depends on a shared module, cppyy_bazel, that lives
in compiler-research/cppyy at cppyy/bazel/. That repo is archived and the
four Python repos merged into this one, so the module has no home in the
surviving stack. CppInterOp's bazel (best-effort) job still clones cppyy to
find it, and both bazel build //... and bazel test //:tests fail there
today.

This PR gives the module a home here and adds a Bazel build for cppjit.

What it contains

• bazel/ — the shared module, moved out of the archived cppyy repo and
  renamed cppjit_bazel. It holds the copts and linkopts, the @llvm module
  extension, the wheel-layout staging rule and the test macros. It is its own
  Bazel module, so cppjit and CppInterOp can both depend on it without a
  bzlmod cycle.
• A root Bazel build. One merged cc_shared_library, mirroring CMake:
  every src/cpyrt/*.cxx plus the two interop translation units link into
  python/cppjit/libcppjit.so. stage_files recreates the wheel layout
  beside it, which is what cppinterop_paths() anchors at its own dladdr
  location.
• A standalone CppInterOp pin. The build fetches CppInterOp from a
  pinned archive, so a fresh clone needs no sibling checkout.
  Only LLVM_DIR comes from the host, as with CMake. A commented
  local_path_override keeps the sibling dev loop available — the Bazel
  equivalent of CMake's CPPINTEROP_SOURCE_DIR.
• ${ORIGIN} expansion in interpreter args. No installer knows its
  absolute prefix at build time. libcppjit.so now rewrites
  CPPINTEROP_EXTRA_INTERPRETER_ARGS, so a consumer passes the token through
  verbatim.
• The best-effort Bazel CI job, restored from the archived cppyy repo.
  Non-gating (continue-on-error), on LLVM 22 and 23; the standalone pin
  removes the sibling clones the old job needed.
• extra_tags on cppinterop_cc_test. A consumer can select or exclude
  one test with --test_tag_filters.

Standalone constraint

The build must work from a fresh clone with nothing beside it. Verified: this
branch cloned into an empty directory, no sibling checkout, a fresh
--output_base, and --nosystem_rc --nohome_rc so no host bazelrc can
supply a registry.

LLVM_DIR=/path/to/llvm bazelisk test //...
Executed 26 out of 26 tests: 26 tests pass.

test_memoryanalysis, added since, passes in a normal checkout.

//:test_selflocation is the layout gate: it imports and JITs with cwd /
and no path environment variables, so a regression in the dladdr-relative
resource lookup fails here instead of only in an installed wheel.

Every commit in the series builds on its own.

CppInterOp pin

Pinned at f8f0689a, the CppInterOp commit with the Bazel fix for the
generated BakedArgs.h (compiler-research/CppInterOp#1162). CPPINTEROP_GIT_TAG
names d3c8cf7a, which predates that fix, so the two pins differ until the
CMake pin moves to f8f0689a or later. The in-tree patch only
renames the shared module in the fetched tree; it goes away once the pinned
commit carries the rename.

Parity with pip install .

//:site_packages reproduces the pip install . payload class for class —
one unversioned libclangCppInterOp.so, the CppInterOp headers, the clang
builtin resource dir, libcppjit.so with a matching soname, and the
pure-Python sources — with a single remaining difference: Bazel ships the
library unstripped (181 MB against 156 MB stripped). .dynsym survives
stripping, so the dlsym-based dispatch resolves either way. Happy to add a
strip step if you want exact bytes.

Follow-up

A companion PR in CppInterOp repoints its bazel_dep and
local_path_override at cppjit_bazel, updates the CI clone, wires the test
sources CMake already lists, and tags VTableOverlayCrossTUBench as a
benchmark. It needs this PR first.

[conrade-ctc]

@aaronj0, here's the bazel infra added back in for cppjit. I'll open another PR against CppInterOp to adjust properly once this is in, and we should be cleanly bazel capable again!

[conrade-ctc]

@mcbarton / @aaronj0, all updated now! thanks.

[conrade-ctc]

One more thing the 3.14 bump surfaced: a latent bug. bazel/sitecustomize.py repoints sysconfig's INCLUDEPY — python-build-standalone bakes it as /install/include/pythonX.Y, which does not exist at runtime — and anchors the runfiles-relative LD_LIBRARY_PATH segments to absolute paths. It relied on site auto-importing the magic module name. Under rules_python 1.8.5 the py_test bootstrap builds sys.path in-process, after site runs, so the module was never imported and both fixups were silently inert. test_regression catches the first; the second would only have shown up as a dlopen failure on GitHub Actions. test/test_main.py now imports the module explicitly, guarded by try/except ImportError so the pip/CMake path is unaffected. 26/26 green on 3.14.

[conrade-ctc]

@aaronj0, can this merge now? it'd be great to get the bazel stuff settled and off the docket.

[aaronj0]

Thanks @conrade-ctc! This LGTM % minor comment on the patch file

[aaronj0]

Do we still need this patch file?

[conrade-ctc]

Yes, for now. CppInterOp main still declares bazel_dep(name = "cppyy_bazel") and loads from @cppyy_bazel. The patch renames these in the fetched tree.

The order is fixed: this PR gives cppjit_bazel a home. Right after this merges, a CppInterOp PR switches to it, and its CI clones cppjit instead of cppyy. After that merges, I bump the pin here and delete the patch.