src,tools: initialize cppgc

[dharesign]

This patch:

• Initializes cppgc in InitializeOncePerProcess() when
  kNoInitializeCppgc is not set
• Create a CppHeap and attach it to the Isolate when
  there isn't one already during IsolateData initialization.
  The CppHeap is detached and terminated when IsolateData
  is freed.
• Publishes the cppgc headers in the tarball.

This allows C++ addons to start using cppgc to manage objects.

A helper node::SetCppgcReference() is also added to help addons
enable cppgc tracing in a user-defined object.

Refs: #40786
Refs: https://docs.google.com/document/d/1ny2Qz_EsUnXGKJRGxoA-FXIE2xpLgaMAN6jD7eAkqFQ/edit

[nodejs-github-bot]

Review requested:

• @nodejs/startup

[dharesign]

In Electron, the cppgc heap will be initialized by Chromium. I don't think anything in this PR clashes with what Electron will want.

When they initialize node, they can use kNoInitializeCppgc here:
https://github.com/electron/electron/blob/4e661842879ae0cb01a9fc251d56c8cabc958466/shell/app/node_main.cc#L161-L165

They don't use NodeMainInstance or CommonEnvironmentSetup.

CC: @codebytere / @nornagon for awareness

[nornagon]

FYI I believe we don't currently initialize cppgc in the main process in Electron (though Blink does of course initialize it). So we would want to be able to initialize Node with cppgc initialization in the main process (either by doing it ourselves or through this initialization option), and without cppgc initialization in the renderer process (since that would clash with Blink's initialization).

[dharesign]

Erp, I hit the close button by mistake. Please re-approve running the workflows >_<

[dharesign]

The previous test runs seemed to fail, and actually running my local build of node likewise fails. It seems to be due to a bug in v8:

https://github.com/v8/v8/blob/main/src/heap/cppgc/heap-base.cc#L104

I believe this should be delgate_ not delegate: the latter is moved from.

CC: @mlippautz

[dharesign]

The previous test runs seemed to fail, and actually running my local build of node likewise fails. It seems to be due to a bug in v8:

https://github.com/v8/v8/blob/main/src/heap/cppgc/heap-base.cc#L104

I believe this should be delgate_ not delegate: the latter is moved from.

CC: @mlippautz

Yep, changing that fixes my local build.

[dharesign]

The previous test runs seemed to fail, and actually running my local build of node likewise fails. It seems to be due to a bug in v8:
https://github.com/v8/v8/blob/main/src/heap/cppgc/heap-base.cc#L104
I believe this should be delgate_ not delegate: the latter is moved from.
CC: @mlippautz

Yep, changing that fixes my local build.

I added a second commit to this PR which fixes the above. We can switch the commit to cherry-pick the v8 commit once it gets fixed there. Hopefully the CI checks will proceed now.

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/48450/

[dharesign]

Node is still segfaulting when I run it locally:

* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x0)
    frame #0: 0x0000000100a5c14c node`cppgc::internal::MarkerBase::EnterAtomicPause(cppgc::EmbedderStackState) [inlined] cppgc::internal::SingleThreadedHandle::Cancel(this=<unavailable>) at task-handle.h:27:20 [opt]
   24
   25  	  void Cancel() {
   26  	    DCHECK(is_cancelled_);
-> 27  	    *is_cancelled_ = true;
   28  	  }
   29
   30  	  bool IsCanceled() const {
Target 0: (node) stopped.
(lldb) frame sel 1
frame #1: 0x0000000100a5c144 node`cppgc::internal::MarkerBase::EnterAtomicPause(this=0x0000000106808200, stack_state=kNoHeapPointers) at marker.cc:260:33 [opt]
   257 	    // Cancel remaining incremental tasks. Concurrent marking jobs are left to
   258 	    // run in parallel with the atomic pause until the mutator thread runs out
   259 	    // of work.
-> 260 	    incremental_marking_handle_.Cancel();
   261 	    heap().stats_collector()->UnregisterObserver(
   262 	        incremental_marking_allocation_observer_.get());
   263 	    incremental_marking_allocation_observer_.reset();
(lldb) print incremental_marking_handle_
(cppgc::internal::MarkerBase::IncrementalMarkingTaskHandle) $0 = {
  is_cancelled_ = nullptr {
    __ptr_ = 0x0000000000000000
  }
}

It looks like foreground_task_runner_ is nullptr, so MarkerBase::ScheduleIncrementalTaskRunner exits early before initializing incremental_marking_handle_. Later, MarkerBase::EnterAtomicPause assumes that incremental_marking_handle_ is set such that it can call .Cancel(), but it's not.

[dharesign]

Looks like there's no isolate_ set in the CppgcPlatformAdapter, which implies CppHeap::AttachIsolate has not been called, though it has.

After digging a bit, it's because CppHeap::AttachIsolate does static_cast<CppgcPlatformAdapter*>(platform()), where platform() in this case is actually PlatformWithPageAllocator. The static_cast is therefore invalid.

CC: @mlippautz

[dharesign]

Submitted the following v8 bugs:

• https://bugs.chromium.org/p/v8/issues/detail?id=13589
• https://bugs.chromium.org/p/v8/issues/detail?id=13590

[dharesign]

I added a commit to add SetIsolate as a virtual method on the cppgc::Platform protocol. Not sure if that's the direction v8 will go to fix the issue, but it should help get us past that issue and check the rest of the CI passes.

[dharesign]

Please re-approve running the workflows 😅

[nodejs-github-bot]

Landed in 7f2c810

[joyeecheung]

Note to backporters: this depends on #48660

[UlisesGascon]

This PR didn't land cleanly because of #48660 in Node 20.x. Can you backport it?

[glaubitz]

This change broke NodeJS on big-endian ppc64 on Linux for me. (see: #58277)

[joyeecheung]

I don't think V8 tests for ppc64be on their CI, nor does Node.js support ppc64be + Linux, from the stack trace it looks like a V8 support issue and the fix probably won't be in Node.js.