[llvm] Reduce llvm-gsymutil memory usage (PR #91023)

[Kevin Frei via llvm-commits]

================
@@ -658,6 +711,9 @@ void DWARFUnit::clearDIEs(bool KeepCUDie) {
   // It depends on the implementation whether the request is fulfilled.
   // Create a new vector with a small capacity and assign it to the DieArray to
   // have previous contents freed.
+  llvm::sys::ScopedWriter FreeLock(CUDieFreeMutex);
+  llvm::sys::ScopedWriter CULock(CUDieArrayMutex);
+  llvm::sys::ScopedWriter AllLock(AllDieArrayMutex);
----------------
kevinfrei wrote:

I can't remove the other two locks. They have to stay, as the freeing mechanism requires it. The third lock (Free) is specifically to deal with the issue that @ayermolo pointed out a couple revisions ago. (Freeing could free just the CU DIE while the non-CU DIE stuff was being initialized, resulting in an invalid state of no CU DIE, but non-CU DIE's are parsed). The entire process is kind of a mess, but I need a "CU DIE initialization" RWLock, plus the "Non-CU DIE initialization" lock as two separate entities, as initializing the non-CU DIE's can trigger recursion (you have to be able to grab the reader lock of the CU DIE init to tell that it's already been initialized, then bail)

https://github.com/llvm/llvm-project/pull/91023