关联的Issue

https://gitcode.com/openharmony/arkcompiler_ets_runtime/issues/13571

修改原因（目的、解决的问题等）

ReMarkAndPrepare posted concurrent sweeper tasks (PostTask(true)) before mutator TLABs were retired (ResetTlab ran later inside PrepareForCopy). Once tasks are posted, worker threads concurrently build free objects over unmarked TLAB remains in shared non-movable regions (FillFreeObject + freeList_->Free), while the daemon still writes the same memory via ThreadLocalAllocationBuffer::Reset -> FillBumpPointer -> FillFreeObject without any lock. The unlocked write clobbers sweeper-built FreeObject nodes (next reset to INVALID_OBJECT, available overwritten), which can break the free list chain or corrupt span metadata — subsequent FreeListAllocator::Allocate may hand out memory overlapping live objects (heap corruption). All other shared GCs (SharedGC / SharedFullGC / SharedPartialGC) reset TLABs before posting sweep tasks; only SharedCC had the order inverted.

Note: the string table concurrent-sweep guard issue also tracked in #13571 is intentionally NOT addressed in this change (deferring it needs a safe protocol w.r.t. threads already inside string table critical sections); it will be handled separately.

修改描述（做了什么，变更了什么）

• Moved ResetTlab for all threads from PrepareForCopy to before Sweep(true)/PostTask(true) in ReMarkAndPrepare (same STW section, mutators suspended; SetReadBarrierState timing unchanged, still after the FROM/TO swap).
• Fixed the dead ParallelCopy skip condition (CalculateCopyThreadNum always returns >= 1); it now checks copyTasks_.empty() directly.
• Narrowed PrepareAllThreads to PrepareMainThread: only the main thread is posted before marking (the WaitMainThreadReady handshake); other threads are switched directly when their stack is empty, or posted at their next idle point from CCMarkFlipFunction during initial mark, where the flip iteration also guarantees thread liveness.
• Switched to RUNNING state (ThreadManagedScope) in JSNApi::SetPostTaskToThreadCallback before installing the callback: SharedCC sweeper/marker threads read postTaskToThreadCallback_ concurrently (shared_cc.cpp / shared_concurrent_marker.cpp), and the RUNNING/GC mutual exclusion prevents a data race on the store.

自测试项（测试结果截图，直接贴到每一个测试项底下）

jit-workload 本地执行（compiler/jit/isa/stub/ic/对象布局等修改必须执行）

• 已通过
• 不涉及，无需验证

interpreter-js_perf 性能测试（性能优化/性能敏感场景：IR化/ic/hClass/对象布局/isa/builtins等修改必须执行）

• 已验证不劣化
• 不涉及，无需验证

aot-js_perf 性能测试（性能优化/性能敏感场景：IR化/ic/hClass/对象布局/isa/builtins等修改必须执行）

• 已验证不劣化
• 不涉及，无需验证

interpreter-workload性能测试（性能优化/性能敏感场景：IR化/ic/hClass/对象布局/isa/builtins等修改必须执行）

• 已验证不劣化
• 不涉及，无需验证

aot-workload性能测试（性能优化/性能敏感场景：IR化/compiler/aot/isa/stub/ic/对象布局等修改必须执行）

• 已验证不劣化
• 不涉及，无需验证

非StableArray不能采用ElementKind读取元素（修改解释器相关IR代码需要自检）

• 已自检
• 不涉及，无需验证

在IR时添加fastpath时，需要检查是否有分支会改变非字典序模式或者修改数组长度，此时应该走slowpath（修改解释器相关IR代码需要自检）

• 已自检
• 不涉及，无需验证

IR化后的逻辑需要与与c++实现一致，不一致的地方需要有分支用例看护（修改解释器相关IR代码需要自检）

• 已自检
• 不涉及，无需验证

c++代码中检测或抛出异常的地方，IR需要同样实现（修改解释器相关IR代码需要自检）

• 已自检
• 不涉及，无需验证

IR中使用默认入参是HeapObject对象的接口时，需要在上下文中保证了参数的正确性（修改解释器相关IR代码需要自检）

• 已自检
• 不涉及，无需验证

修改c解释器，需要同时检查汇编解释器与AOT是否需要修改

• 已自检
• 不涉及，无需验证

是否已执行L0用例

• 已验证
• 不涉及。如不涉及，请写明理由
  SharedGC concurrency fix in C++ GC code, no IR/interpreter changes; to be covered by CI gate GC test suites (GC_First/Second/Third_Test, GC_SharedPartial_Test, GC_SharedHeap_Test).