[Mlir-commits] [mlir] [mlir][bufferization] Add `BufferizableOpInterface::hasTensorSemantics` (PR #75273)

[Matthias Springer]

================
@@ -989,3 +995,22 @@ bufferization::detail::unknownGetAliasingValues(OpOperand &opOperand) {
           r.addAlias({bbArg, BufferRelation::Unknown, /*isDefinite=*/false});
   return r;
 }
+
+static bool isaTensor(Type t) { return isa<TensorType>(t); }
+
+bool bufferization::detail::defaultHasTensorSemanticsForBufferization(
+    Operation *op) {
+  bool hasTensorBlockArgument = any_of(op->getRegions(), [](Region &r) {
+    return any_of(r.getBlocks(), [](Block &b) {
+      return any_of(b.getArguments(), [](BlockArgument bbArg) {
+        return isaTensor(bbArg.getType());
+      });
+    });
+  });
+  if (hasTensorBlockArgument)
+    return true;
+
+  bool hasTensorResult = any_of(op->getResultTypes(), isaTensor);
+  bool hasTensorOperand = any_of(op->getOperandTypes(), isaTensor);
----------------
matthias-springer wrote:

In practice, results are usually sufficient. But we could have a hypothetical op such as `test.dump_tensor`, which takes a tensor operand and has no result.


https://github.com/llvm/llvm-project/pull/75273