[Mlir-commits] [mlir] [mlir] Use non thread-local allocator for DistinctAttr when threading is disabled (PR #128566)

[Tobias Gysi]

================
@@ -411,12 +414,51 @@ class DistinctAttributeAllocator {
   /// Allocates a distinct attribute storage using a thread local bump pointer
   /// allocator to enable synchronization free parallel allocations.
   DistinctAttrStorage *allocate(Attribute referencedAttr) {
-    return new (allocatorCache.get().Allocate<DistinctAttrStorage>())
-        DistinctAttrStorage(referencedAttr);
+    if (!useThreadLocalAllocator && threadingIsEnabled) {
+      std::scoped_lock<std::mutex> lock(allocatorMutex);
+      return allocateImpl(referencedAttr);
+    }
+    return allocateImpl(referencedAttr);
+  }
+
+  /// Sets flags to use thread local bump pointer allocators or a single
+  /// non-thread safe bump pointer allocator depending on if multi-threading is
+  /// enabled. Use this to disable the use of thread local storage and bypass
+  /// thread safety synchronization overhead.
----------------
gysit wrote:

```suggestion
  /// Sets a flag that stores if multi-threading is enabled. The flag is used to decide if locking is needed
  /// when using a non thread-safe allocator.
```


https://github.com/llvm/llvm-project/pull/128566