[lldb-dap] Simplify DAPSessionManager GetInstance

[da-viper]

We don't need the std::call_once as this guaranteed to be thread safe and initialised once since c++11.
I also don't see any issue with the destructor chain since the class doesn't own any data that can be affected.

[llvmorg-github-actions]

@llvm/pr-subscribers-lldb

Author: Ebuka Ezike (da-viper)

Changes

We don't need the std::call_once as this guaranteed to be thread safe and initialised once since c++11.
I also don't see any issue with the destructor chain since the class doesn't own any data that can be affected.

---

Full diff: https://github.com/llvm/llvm-project/pull/209264.diff

1 Files Affected:

• (modified) lldb/tools/lldb-dap/DAPSessionManager.cpp (+2-8)

diff --git a/lldb/tools/lldb-dap/DAPSessionManager.cpp b/lldb/tools/lldb-dap/DAPSessionManager.cpp
index 580d4fa102a89..91c0538e63d5f 100644
--- a/lldb/tools/lldb-dap/DAPSessionManager.cpp
+++ b/lldb/tools/lldb-dap/DAPSessionManager.cpp
@@ -33,14 +33,8 @@ ManagedEventThread::~ManagedEventThread() {
 }
 
 DAPSessionManager &DAPSessionManager::GetInstance() {
-  static std::once_flag initialized;
-  static DAPSessionManager *instance =
-      nullptr; // NOTE: intentional leak to avoid issues with C++ destructor
-               // chain
-
-  std::call_once(initialized, []() { instance = new DAPSessionManager(); });
-
-  return *instance;
+  static DAPSessionManager instance;
+  return instance;
 }
 
 void DAPSessionManager::RegisterSession(lldb_private::MainLoop *loop,

[JDevlieghere]

Dropping call_once is fine, the statics handles thread-safe init. But this also drops the intentional leak which I think was load bearing.

Now the destructor will run at program exit. Client sessions run on detached threads and UnregisterSession ends with notify_all_at_thread_exit. The main thread only waits for the sessions map to empty, which happens before the detached thread finishes its exit tail. So there's a window where ~DAPSessionManager tears down m_sessions_mutex/m_sessions_condition while a detached thread is still notifying them. That's what the old comment was about.

Let's keep the leak but drop the call_once:

DAPSessionManager &DAPSessionManager::GetInstance() {
  // Intentionally leaked: detached client threads may still be notifying on
  // m_sessions_condition at exit, so it has to outlive them.
  static DAPSessionManager *instance = new DAPSessionManager();
  return *instance;
}