Add a browser-compatible Worker polyfill (dedicated workers, structured clone, transfer lists, focused WPT regressions) - #258
Open
matthargett wants to merge 201 commits into
Conversation
…ction when the receiver was undefined, so the VM forcibly substituted the global object even in strict mode. The new implementation always routes through Function.prototype.call, preserving the exact thisArg. This only affected JSC: Chakra already pushes recv onto the argv array before invoking JsCallFunction, and V8 hands the raw recv value to Function::Call. Neither engine coerces in strict mode, so no additional fixes were required.
…bug fix in strict mode was actually found by the suite! The failing behavior was exercised by Tests/NodeApi/test/js-native-api/3_callbacks/test.js. New cmake targets emit a node-lite binary and a NodeApiTests binary, all currently enabled tests for currently supported NAPI v5 pass on Mac. Next step is to enable them for running in Android simulator.
…x handles during instrumentation, causing crashes, and now route console output through the new NodeLiteRuntime::Callbacks. On Android we forward stdout/stderr to logcat via callbacks to work around this for now. Added Android-specific shims (node_lite_android.cpp, child_process_android.cpp) so native module loading uses dlopen and JS child_process.spawnSync safely reports “unsupported”. Extended the Node‑API harness to allow in-process execution: RunNodeLiteScript captures output, SetNodeApiTestEnvironment lets the JNI layer provide a base directory and asset manager, and the GTest registration path uses that configuration instead of shelling out to the node_lite executable
…-- a use-after-free. Will check sanitizers under Android next
…ned into our copy of the JSI code.
… since they found bugs on macOS + JSC
Two build-restoration fixes (no behavior/impl or NAPI-version changes), needed
after rebasing onto upstream HEAD and building with the current Xcode/libc++:
- node_lite: NodeApi::CallFunction took std::span<napi_value> but is only ever
called with braced-init-lists ({a,b,c}). Newer libc++ correctly rejects
constructing a non-const std::span from an initializer_list (that ctor is
C++26). Switch the parameter to std::initializer_list<napi_value> (begin()
yields the const napi_value* napi_call_function wants).
- Tests/NodeApi: the POST_BUILD copy_directory of each .node runs as an Xcode
script phase BEFORE Xcode's implicit CodeSign phase signs the original, so the
copied addons that node_lite/NodeApiTests dlopen are unsigned on a clean build
and macOS refuses to load them. Ad-hoc sign the copies directly (APPLE only).
…mulator)
Build-restoration fixes for the Android in-process NodeApi harness after rebasing
onto upstream HEAD (no impl/NAPI-version changes). Each was a latent break in the
napi-tests Android integration, surfaced by a clean build on a current toolchain:
- CMakeLists.txt: drop the AndroidExtensions Globals.cpp 'patch' step. It file(COPY)'d
patches/AndroidExtensions/Globals.cpp, which was never committed in any ref (author's
local-only file). Upstream uses a newer AndroidExtensions pin and needs no patch.
- build.gradle: bump default ndkVersion 23.1.7779620 -> 28.2.13676358 (matches CI's
NDK_VERSION). NDK 23's libc++ can't compile googletest 1.17.0's <=> usage. Also map the
Android sanitizer flag JSR_ENABLE_ASAN -> ENABLE_SANITIZERS (the upstream option kept
during the rebase).
- Tests/NodeApi/CMakeLists.txt: use ${JsRuntimeHost_SOURCE_DIR} instead of
${CMAKE_SOURCE_DIR} for Core/Node-API include paths. On Android JsRuntimeHost is added
as a subdirectory of the app, so CMAKE_SOURCE_DIR was the app dir (headers not found);
the project-scoped var is correct in both standalone (macOS) and nested (Android) builds.
- Tests/NodeApi/CMakeLists.txt: allow the .node modules to link with unresolved napi_*
symbols on Android (-Wl,--unresolved-symbols=ignore-all), the ELF equivalent of Apple's
-undefined dynamic_lookup; they bind at dlopen time from the host (UnitTestsJNI).
- Shared.cpp: gate the Android NodeApi-harness block on NODE_API_AVAILABLE_NATIVE_TESTS
(defined only by UnitTestsJNI) so the standalone UnitTests target -- built but unused on
Android -- doesn't try to compile AndroidExtensions/NodeApi code it doesn't link.
The instrumented run aborted with 'use of deleted global reference': the harness fell back to android::global::GetAppContext() (GetFilesDir -> GetObjectClass) whose JNI global ref is not valid during the instrumented run. JNI.cpp now computes a writable base dir from the still-valid instrumentation Context and passes it plus the native AAssetManager to SetNodeApiTestEnvironment() before RunTests() -- the wiring the harness was designed for (see e1fce6b) but which was never actually connected. This removes the crash and lets ConfigureNodeApiTests run. NOTE: on-device execution of the NodeApi conformance tests is still not achieved -- CopyAssetsRecursive relies on AAssetManager subdirectory enumeration (AAssetDir_getNextFileName lists files only, not dirs) so the nested test tree isn't copied, and the native .node modules are neither packaged nor loadable from an app-writable dir on API 29+. Tracked as follow-up.
Before this, the instrumented run passed vacuously -- no NodeApi tests ran. Several layered fixes get them executing on the emulator (macOS path unchanged: still 12/12): #1 Asset enumeration: AAssetManager can't list subdirectories, so CopyAssetsRecursive copied nothing. copyNodeApiTests now emits a file manifest (manifest.txt -- not a dotfile, which aapt would drop) and Shared.cpp copies each listed file. #2 Native module packaging/loading: build each addon as lib<name>.so on Android so AGP packages it into lib/<abi>/ (nativeLibraryDir, the only dlopen-able location on API 29+); node_lite_android loads it by soname; ResolveModulePath resolves the (on-disk absent) .node so LoadNativeModule runs. V8 lifecycle (in-process node_lite shares the host's V8): reuse the host's already- initialized V8 platform (fixes 'Wrong initialization order'); hold a Locker + Isolate::Scope so multi-isolate access is locked (fixes 'Entering the V8 API without proper locking'). KNOWN REMAINING (tracked): node_lite calls Node-API outside any napi callback during NodeLiteRuntime::Initialize/script execution, which on V8 needs a live HandleScope + current Context. v8::HandleScope/Context::Scope are stack-only (operator new is private) so they can't be held across the holder; this needs a scope-wrapping rework of node_lite's V8 entry points (or napi_open_handle_scope + context enter). Until then the on-device native tests segfault in napi_create_object.
…eate_object segfault) NodeApiEnvScope -> jsr_open_napi_env_scope was a no-op stub: it allocated a scope struct but never entered the env's V8 isolate/context. On JSC that's fine (the env carries its context explicitly), but on V8 node_lite then calls Node-API outside any napi callback with no *current context*, so napi_create_object -> v8::Object::New(isolate) segfaulted during NodeLiteRuntime::Initialize. Enter the env's context on open and exit it on close (Android only). The in-process V8 runtime now initializes and runs tests.
… error Step toward in-process error handling: ExitOnException was noexcept, but the in-process runner installs a fatal handler that throws NodeLiteFatalError (rather than std::exit) so the harness can turn a JS error into a ProcessResult. Throwing from the noexcept function std::terminate'd the test process. Dropped noexcept so it propagates to RunNodeLiteScript. (Partial: other noexcept teardown paths -- NodeApiHandleScope/NodeApiEnvScope dtors calling NODE_LITE_CALL, and the env-holder dtor's onUnhandledError -> ExitWithJSError -- can still throw during unwinding when a test errors. Full in-process error-path exception-safety is the remaining Android item.)
NodeApiHandleScope/NodeApiEnvScope destructors used the throwing NODE_LITE_CALL, and the JsRuntimeHostEnvHolder destructor's onUnhandledError can invoke the throwing in-process fatal handler -- both std::terminate if they fire while a NodeLiteFatalError is unwinding. Make the scope dtors ignore the close status and wrap onUnhandledError in try/catch. Correct robustness fixes, but they do NOT yet resolve the remaining in-process failure: when a test errors, a *second* NodeLiteFatalError is thrown during unwinding (double-exception -> std::terminate). The escaping throw site isn't visible in the tombstone (stack already unwound) and needs on-device lldb to pinpoint. macOS unaffected (12/12).
…winding Don't re-throw NodeLiteFatalError from the in-process fatal handler when std::uncaught_exceptions() > 0, to avoid a double-exception std::terminate. (Correct hardening, but the remaining in-process abort is a *single* uncaught NodeLiteFatalError escaping RunNodeLiteScript's catch -- a scope-exit destructor throw on a test that leaves a pending exception; needs on-device lldb to pinpoint.)
…(fixes terminate) THE fix for the in-process abort. ExitWithJSError / ExitWithJSAssertError / ExitWithMessage were declared noexcept. With the default fatal handler they call std::exit (never throw), but the in-process runner installs a handler that *throws* NodeLiteFatalError (caught by RunNodeLiteScript and turned into a ProcessResult). A throw crossing a noexcept boundary is an immediate std::terminate -- so when any test errored (e.g. the expected-error basics tests throw_string/mustcall_failure), the whole instrumented run aborted instead of reporting a result. Removing noexcept lets the throw unwind to the catch. Confirmed on the emulator via a temporary _Unwind_Backtrace probe (now removed): the throw stack was HandleFatalError <- ExitWithMessage(noexcept!) <- ExitWithJSError <- RunTestScript <- RunNodeLiteScript. Net effect: the in-process Android run no longer aborts; the js-native-api v5 tests (2-5) pass; the remaining failures are the basics harness self-tests, run through the generic fixture rather than the specialized test_basics.cpp path macOS uses. macOS unaffected (still 12/12).
… napi The js-native-api conformance addons are dlopen'd in-process by the Android harness and import napi_* from the host (libUnitTestsJNI.so). The host is loaded RTLD_LOCAL by System.loadLibrary, and bionic's linker-namespace model does not surface its statically-linked (but exported) napi_* symbols to a dlopen'd module -- so the addon cannot bind them at load time. Post-hoc RTLD_GLOBAL promotion of the host is a no-op on bionic (confirmed on device: the module dlopen still returns NULL with the host re-opened RTLD_GLOBAL). Making these tests runnable on Android requires building napi as a shared library (libnapi.so) depended on by both the host and the addons -- a packaging change affecting every Android consumer, deferred to a separate change per the v5-suite-in-place scope. Until then, skip the in-process addon tests on Android with a clear reason; macOS runs the full v5 addon suite (12/12) as the reference. This unblocks the Android suite: it now builds, the in-process harness runs without aborting, and the suite passes (addon tests reported SKIPPED).
…ared-lib napi follow-up
… in-process) Dynamic .node loading is never shipped to the Play / Quest stores, and bionic won't resolve a dlopen'd addon's napi_* imports against the System.loadLibrary-loaded host anyway (the addon carries no DT_NEEDED for napi; RTLD_GLOBAL host promotion is a no-op on bionic). Rather than make napi a shared library for every Android consumer (tracked separately, task #9), compile the conformance addons directly into the host (UnitTestsJNI) and resolve them in-process. To link several addons into one binary without symbol clashes: - node_api.h: make NODE_API_MODULE_REGISTER_FUNCTION / _GET_API_VERSION_FUNCTION overridable. - entry_point.h (JSR_NODE_API_STATIC_LINK): give Init internal linkage and emit a per-addon load-time constructor that self-registers its uniquely-suffixed registrar/version functions with the host. - The Android CMakeLists compiles each addon as an OBJECT library with per-module unique entry-point names and links them into UnitTestsJNI. - node_lite_android LoadFunction resolves entry points from the in-process static registry by module name instead of dlopen+dlsym. Removes the Android GTEST_SKIP. The 4 v5 js-native-api conformance tests now execute in-process and PASS on Android (2_function_arguments, 3_callbacks, 4_object_factory, 5_function_factory). macOS is unchanged (the desktop dynamic .node path uses the #else branches).
The conformance suite runs gtest in-process; its results (RUN/OK/FAILED and failure file:line:message) went to stdout, which Android discards -- leaving only the JUnit "expected 0, was 1" with no detail. Pump stdout/stderr to logcat (tag NodeApiTests) so test output and any pre-crash native context are visible via `adb logcat -s NodeApiTests`.
…(supersedes skip)
…by static linking) The conformance addons are statically linked into the in-process Android test host (f32130e), so the standalone SHARED .so + -Wl,--unresolved-symbols=ignore-all + lib<name>.so naming that the old dlopen-on-Android path needed are dead. add_node_api_module now early-returns on Android and is a clean desktop-only MODULE .node helper. Also fixes a stale node_lite comment describing the abandoned soname-dlopen path. No functional change on desktop (MODULE .node, -undefined dynamic_lookup, POST_BUILD staging, codesign all preserved); macOS still 12/12.
V8Platform::EnsureInitialized() became a no-op once we found the host AppRuntime already initializes V8's process-global platform; the class and its unused init_flag_/platform_ members were left over from the abandoned platform-init attempt. Fold the (still-important) "don't re-init the platform" rationale into a comment at the isolate-creation site, and drop the dead class plus the now-unused <mutex> / <libplatform> includes. No behavior change; Android still 4/4 js-native-api.
… track, don't adopt yet
…d libnapi.so Replaces the interim static-link-into-host approach with the dynamic .node model used by nodejs/node-api-cts (add_node_api_cts_addon), so the Android suite and a future node-api-cts migration share one addon model. - Core/Node-API: build napi as a SHARED library (libnapi.so) on Android. It exports all 106 napi_* (default visibility -- no global -fvisibility=hidden), and the host plus every addon depend on the one libnapi.so via a real DT_NEEDED, so there is a single napi instance. Static elsewhere. - The conformance addons are again standalone SHARED lib<name>.so (packaged into nativeLibraryDir), now linking napi (DT_NEEDED libnapi.so) instead of -Wl,--unresolved-symbols=ignore-all. - node_lite_android resolves entry points via dlopen(soname)+dlsym again; the addon's napi_* bind from libnapi.so at load. Reverts the static-link infra (entry_point.h JSR_NODE_API_STATIC_LINK branch, node_api.h overridable registrar macros, the host's per-addon OBJECT libraries). Verified on device: lib2_function_arguments.so has DT_NEEDED [libnapi.so], its napi_* are imports, libnapi.so exports the 106 napi_*, and all 4 v5 js-native-api tests pass in-process. macOS unchanged (12/12; desktop keeps static napi + dlopen'd MODULE .node).
… with node-api-cts)
Replace the hand-rolled pipe+thread stdout pump (added while bringing up the in-process Node-API harness) with android::StdoutLogger::Start()/Stop() from AndroidExtensions, which the rest of the UnitTests host already uses. Same effect -- the in-process gtest output (incl. failure file:line:message) is visible in logcat (tag StdoutLogger) -- with less bespoke code. Verified on emulator: 8/8 UnitTests pass incl. 4/4 js_native_api, gtest output present in logcat.
…not a hand-rolled pump
…PI_SHARED option)
…on merge left unbalanced lines)
… adopt the per-feature test layout Upstream BabylonJS#257 split Tests/UnitTests/Scripts/tests.ts and Shared/Shared.cpp by feature and packages fixtures under Assets. Everything this stack added to the monolithic files moves accordingly: - Scripts: tests.streams.ts, tests.compression.ts, tests.headers.ts, tests.response.ts and tests.indexedDB.ts hold the suites that were new blocks; the Blob, File, fetch, AbortController and URL.createObjectURL changes are merged into the split files that already hold those blocks (upstream's app:///Assets paths kept). - Native: Tests.Streams.cpp, Tests.Compression.cpp, Tests.Fetch.cpp, Tests.IndexedDB.cpp, Tests.Performance.cpp and Tests.Worker.cpp; BabylonJS#239's primitive-value tests join Tests.NodeApi.cpp, the Apple run-loop regression joins Tests.AppRuntime.cpp, and TestAssetRoot() (the executable directory, shared by the JavaScript suite's Worker ScriptRoot and the Worker tests) becomes TestAssetRoot.h. - Fixtures: the focused Worker regressions move to Assets/WebPlatformTests and are staged as a directory next to the bundled assets (Assets/WebPlatformTests/runner.js). - The Worker WPT workflow that ran on PRs against napi-v7 is dropped: ci.yml's Linux JSC, sanitizer and ThreadSanitizer jobs cover the same builds on a PR against main. Verified locally on macOS (JavaScriptCore, RelWithDebInfo): 39/39 native tests including Worker.WebPlatformTests, and the 322-test JavaScript suite; NDK syntax checks of the Android-compiled sources and a Gradle configuration pass for the Android test app.
This was referenced Sep 26, 2026
…abylonJS#257) The Streams suite moves from the monolithic tests.ts to Source/Scripts/tests.streams.ts, the native idempotency test to Source/Tests.Streams.cpp, and the JavaScript suite initializes Streams before Blob/File (Blob.stream() and Response bodies build on it) in Source/Tests.JavaScript.cpp.
…abylonJS#257) The Streams and Compression suites move from the monolithic tests.ts to Source/Scripts/tests.streams.ts and tests.compression.ts, the native idempotency tests to Source/Tests.Streams.cpp and Source/Tests.Compression.cpp, and Source/Tests.JavaScript.cpp initializes Streams before Blob/File and Compression before Fetch.
…abylonJS#257) The Blob streaming/slicing tests land in the split tests.blob.ts and tests.file.ts (upstream's app:///Assets paths kept), the Streams suite in tests.streams.ts with its native idempotency test in Source/Tests.Streams.cpp, and Source/Tests.JavaScript.cpp initializes Streams before Blob/File.
…abylonJS#257) The Headers and Response suites become Source/Scripts/tests.headers.ts and tests.response.ts, the fetch, Blob and File changes merge into the split files that hold those blocks (upstream's app:///Assets paths kept), the native Fetch host-class tests move to Source/Tests.Fetch.cpp and the Streams idempotency test to Source/Tests.Streams.cpp; Source/Tests.JavaScript.cpp initializes Streams before Blob/File and Fetch after them.
…abylonJS#257) The data: URL fetch tests merge into the split tests.fetch.ts alongside the Headers/Response and Streams suites this branch carries (tests.headers.ts, tests.response.ts, tests.streams.ts), the Blob and File changes into tests.blob.ts and tests.file.ts (upstream's app:///Assets paths kept), and the native tests to Source/Tests.Fetch.cpp and Source/Tests.Streams.cpp; Source/Tests.JavaScript.cpp initializes Streams before Blob/File and Fetch after them.
…abylonJS#257) The three Node-API tests (primitive exception through a native catch, coerced property receivers, references to primitives) move from Shared.cpp to Source/Tests.NodeApi.cpp with their engine guards, and the JSRUNTIMEHOST_NAPI_ENGINE_JAVASCRIPTCORE definition follows the test target into the relocated CMake lists.
…abylonJS#257) Performance.TimeOriginIsPerEnvironment moves from Shared.cpp to Source/Tests.Performance.cpp, registered in the desktop and Android test targets.
…abylonJS#257) The Blob-identity and writable-URL-statics regression tests move from the monolithic tests.ts into the URL.createObjectURL block of Source/Scripts/tests.url.ts.
…into the Worker stack
…818ee) into the Worker stack
…into the Worker stack
…62818ee) into the Worker stack
… into the Worker stack
…62818ee) into the Worker stack
…main 62818ee) into the Worker stack
This was referenced Sep 26, 2026
Open
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A browser-compatible dedicated
Workerfor JsRuntimeHost.What it does
AppRuntime, engine realm and event loop per worker.Worker,EventTarget,Event,MessageEvent,ErrorEvent,DOMException, handler attributes,postMessage,terminate.DedicatedWorkerGlobalScope:self,location,name,navigator,close, synchronousimportScripts, timers, fetch/XHR, URL, Blob/File, text codecs, WebSocket, performance, abort APIs, Streams, CompressionStream, IndexedDB.napi_detach_arraybuffer.app:///and relative paths, explicitfile://, percent-encodeddata:.Stack
Top of a stack; the diff carries the dependencies until they merge (bottom-up):
data:fetch.globalThis.performance.now()origin per environment.JsRuntime).The dependencies are merged in as their own heads, so they appear as merge commits; the Worker's own commits sit on top (
Restack Worker …throughWorker: treat { type: undefined } as the classic default), plus two small fixes to existing polyfills: primitiveAbortSignaland fetch abort reasons are preserved.Not included: #206, #217, #182 and the other standalone fixes from the same integration work (#243–#250, #252–#254). Current with
main@ 62818ee and #257's test layout.Tests
Tests.Worker.cpp:DOMExceptionpreserved across Worker initialization;WorkerOptions.type(undefinedmeans classic; any other value that is notclassic/moduleis aTypeError);Assets/WebPlatformTests, each fixture linking its source: WPTclose()and startup ordering, Firefox pre-start queue and dead-worker cleanup, WPT terminate-during-evaluation with the Chromium dead-context case, the WebKit termination-order case, corrupt gzip input, and an app-derived visualization worker that hydrates an IndexedDB cache from a Worker-relative gzip fetch.Tests.Streams.cpp,Tests.Compression.cpp,Tests.Fetch.cpp,Tests.IndexedDB.cpp) and the JavaScript suites for Streams, Compression, Headers, Response, Blob streaming and IndexedDB.Worker.WebPlatformTests, 322/322 JavaScript tests, four consecutive full runs.First-cut boundaries
type: "module"accepts self-contained script-compatible bundles; imports must be flattened (public JavaScriptCore has no module-loader hook).unhandledrejectionsynthesis,SharedArrayBuffer,MessagePorttransfer, Blob object URLs or nested workers yet.