component-model-gradle.png...

		"minusdu;":                         '\U00002A2A',
		"mlcp;":                            '\U00002ADB',
		"mldr;":                            '\U00002026',
		"mnplus;":                          '\U00002213',
		"models;":                          '\U000022A7',
		"mopf;":                            '\U0001D55E',
		"mp;":                              '\U00002213',
		"mscr;":                            '\U0001D4C2',

    // Match Add.
    DenseElementsAttr added_value;
    Value constant_val = add_op.getRhs();
    if (!matchPattern(constant_val, m_Constant(&added_value))) {
      // The constant may be preceded by QDQs in models with QDQ format, so we
      // should set it to the real constant.
      auto dq = dyn_cast_or_null<DequantizeOp>(constant_val.getDefiningOp());
      if (!dq) return failure();

             << padding.getShapedType().getShape();
    }
  }

  auto module = op->getParentOfType<mlir::ModuleOp>();
  auto func = dyn_cast_or_null<mlir::func::FuncOp>(
      SymbolTable::lookupSymbolIn(module, op.getComputation()));
  if (!func) {
    return op.emitOpError() << "has no reduction function specified";
  }

  //   %pack   = tf.Pack(dim0, ...) { axis = 0 }   // [? x ...]
  //
  // Where `...` are some statically known dimensions. In this case %pack can be
  // replace with a %shape. This is a common pattern in models with a dynamic
  // batch size.

  // Test Rank 2
  // CHECK: %[[SHAPE0:.*]] = "tf.Shape"
  %3 = "tf.Shape"(%arg0) : (tensor<?x1xf32>) -> tensor<2xi32>

    this will happen in one of the last steps before emitting the flatbuffer model
    in order to enable optimizations that rely on a fixed order of operations (such
    as rematerialization.)
    The flatbuffer exporter will unwrap the wrapped region and annotate the generated
    model with metadata such that any runtime reorderings will respect the order
    given by the control dependencies.
  }];