[Mlir-commits] [mlir] [mlir][bufferization] Hoist allocations through exactly-once regions (PR #224921)

[Federico Bruzzone]

================
@@ -367,12 +369,24 @@ struct BufferAllocationLoopHoistingState : BufferAllocationHoistingStateBase {
     return dependencyBlock ? dependencyBlock : nullptr;
   }
 
-  /// Returns true if the given operation represents a loop with sequential
-  /// execution semantics and one of the aliases caused the
-  /// `aliasDominatorBlock` to be "above" the block of the given loop operation.
-  /// If this is the case, it indicates that the allocation is passed via a back
-  /// edge.
-  bool isLegalPlacement(Operation *op) {
+  /// Returns true if the placement search can cross the given parent operation.
+  /// Non-loop regions may be crossed if the current region is executed exactly
+  /// once. Loops must have sequential execution semantics and must not violate
+  /// the alias back-edge constraint.
+  bool isLegalPlacement(Operation *op, Region *region) {
+    if (!isLoop(op)) {
+      auto branchOp = dyn_cast<RegionBranchOpInterface>(op);
+      if (!branchOp)
+        return false;
----------------
FedericoBruzzone wrote:

Prob we can just use something like:
```suggestion
      auto branchOp = cast<RegionBranchOpInterface>(op);
```

https://github.com/llvm/llvm-project/pull/224921