Sanitizers (TZ 3.5 / 5.2)

OXL_SANITIZE builds the runtime (liboxlrt), the compiler (oxlc, oxl) and every test binary with AddressSanitizer + UndefinedBehaviorSanitizer, and makes oxl build link the sanitizer runtime into every compiled OxL program (the program itself is not instrumented, but the instrumented liboxlrt.a inside it is).

cmake -S . -B build-asan -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
      -DOXL_SANITIZE=address,undefined          # or -DOXL_SANITIZE=ON (same thing)
cmake --build build-asan -j2                    # ~25 min on 2 cores (instrumented LLVM-heavy compiler)

export ASAN_OPTIONS=detect_leaks=0 UBSAN_OPTIONS=print_stacktrace=1
ctest --test-dir build-asan -j2 --timeout 1800 --output-on-failure --repeat until-pass:2 -E lsp_latency
python3 scripts/run_conformance.py --oxl "$PWD/build-asan/tools/oxl/oxl" --stage 3 -j 2 --quiet \
        --timeout 300 --run-timeout 60 --ratchet conformance/baseline-stage3.json

How it works:

CMakeLists.txt: OXL_SANITIZE is a string (ON, or a list such as address,undefined; thread is passed through but not part of the nightly). It adds -fsanitize=<list> -fno-omit-frame-pointer to compile and link options, and -fno-sanitize-recover=undefined, so any undefined behaviour aborts the process and fails the test instead of scrolling by as a warning.
compiler/CMakeLists.txt defines OXL_SANITIZE_LINK_FLAG; compiler/src/driver/Build.cpp appends it to the link command of compiled programs (otherwise the link fails with undefined __asan_* / __ubsan_* symbols).
Pass an absolute --oxl path to run_conformance.py: oxl test kinds change the working directory.

Notes on running under sanitizers:

Time: unit tests take ~1.7 h of CPU (test_sema 15 min, test_codegen 16 min); the stage-3 conformance run 30 min at -j2.
lsp_latency measures wall-clock latency against fixed limits and always fails under ASan; it is excluded.
detect_leaks=0 for the unit tests: LeakSanitizer works, but many existing unit tests intentionally do not free what they build (hand-built descriptors without drop, etc.). The fuzz harnesses run with leak detection on, which is where leaks in the decoders are checked. Triage of test-only leaks is a known gap.
Sched.ManySuspendedTasksStayCheap skips its RSS bound under ASan (shadow memory and redzones).
shop_oxl_test (Orders.recent ...) is timing dependent: checkout clears the cart in the background (~>), so the next addToCart of the same user can be overtaken by the clear. It passes ~7 of 8 runs under ASan load and almost always without; not a sanitizer finding.

Bugs found and fixed#

Where	Finding	Fix
`runtime/src/grpc/client.cpp` `ClientCall::resolveFromStream`	heap-use-after-free (ASan, `Grpc.BadServers`): `findHeader(stream_->headers(), ...)` returned a pointer into a temporary copy of the headers	keep the copy in a local
`runtime/src/grpc/wire.cpp` `parseTimeout`	signed overflow (UBSan): peer-controlled `grpc-timeout: 99999999H` = 3.6e20 ns	saturate to `INT64_MAX`
`runtime/src/ser_json.cpp` `parseRfc3339`	signed overflow (UBSan, fuzz_ser): timestamps beyond year 2262 / before 1677 in JSON	range check, decoded as `DecodeError.Invalid`
`runtime/src/http2/h2.cpp` `Stream::endStream`	not idempotent when the stream closed cleanly between two calls (`-EPIPE` instead of `0`); made `H2.ApiMisuse` flaky under load	no-op on a cleanly closed stream
tests: `config_test.cpp` `SecretStr.NewRevealDrop`	read of a freed string (ASan)	check before the release
tests: `h2_test.cpp` `H2.WriteDeadline`	handler task read a stack variable after the test function returned (stack-use-after-scope)	state lives in a `shared_ptr`
tests: `grpc_abi_test.cpp` `UnaryAndMeta`	asserted `active_calls == 0` immediately after the client saw the status (race)	wait for the server side to finish

Regression tests: tests/unit/runtime/fuzz_regress_test.cpp, the saturation cases in grpc_test.cpp (GrpcWire.Timeout); the use-after-free is covered by Grpc.BadServers under OXL_SANITIZE.

Results#

ctest under ASan+UBSan: all tests pass except the two exclusions above (see the final report for counts).
Conformance stage <= 3 under ASan+UBSan: 1083 baseline tests pass, no ratchet regression (stage 2 is a subset).
CI: .github/workflows/nightly.yml job asan-ubsan runs exactly the commands above.

Исходник: docs/sanitizers.md