[llvm] [Offload] Make olLaunchKernel test thread safe (PR #149497)

[Ross Brunton via llvm-commits]

================
@@ -1329,12 +1329,15 @@ Error PinnedAllocationMapTy::unlockUnmappedHostBuffer(void *HstPtr) {
   return eraseEntry(*Entry);
 }
 
-Error GenericDeviceTy::synchronize(__tgt_async_info *AsyncInfo) {
+Error GenericDeviceTy::synchronize(__tgt_async_info *AsyncInfo,
+                                   bool RemoveQueue) {
+  std::lock_guard<std::mutex> AllocationGuard{AsyncInfo->AllocationsMutex};
----------------
RossBrunton wrote:

The use case I was seeing in the tests was that multiple threads were calling `olWait()` concurrently on different threads. They would all go into `synchronize` at the same time and start double-freeing and clobbering the vector.

As far as I can tell, it should also solve the problem you've describe though.

https://github.com/llvm/llvm-project/pull/149497