//
// TF_OpDefinitionBuilder* builder = TF_NewOpDefinitionBuilder("OpName");
//
// Inputs, outputs and attributes can be added to the builder with the
// corresponding functions, e.g.
//
// TF_OpDefinitionBuilderAddInput(builder, "input1: int32");
// TF_OpDefinitionBuilderAddOutput(builder, "output1: int64");
// TF_OpDefinitionBuilderAddAttr(builder, "attr: int32");
//

                void myAction() {
                    def output = outputFile.get().asFile
                    output << launcher1.get().executablePath
                    output << launcher2.get().executablePath
                }
            }

            tasks.register("myToolchainTask", ToolchainTask) {
                outputFile = layout.buildDirectory.file("output.txt")
                launcher1 = javaToolchains.launcherFor {

For the aforementioned reasons, task output caching is automatically disabled for tasks whose output directories overlap with another task.

Build scans show tasks where caching was disabled due to overlapping outputs in the timeline:

[.screenshot]
image::build-cache/overlapping-outputs-timeline.png[]

== Reuse of outputs between different tasks

}

func.func @not_modified(%arg0: tensor<f32>, %arg1: tensor<1x224x224x3xf32>) -> (tensor<1x401408xf32>, tensor<1x224x224x3xf32>) attributes {tf.entry_function = {control_outputs = "", inputs = "input0,input1", outputs = "output0,output1"}} {
  %cst = arith.constant dense<[1, 401408]> : tensor<2xi32>

    // Convert back to a C++ Tensor and ensure we get expected output.
    Tensor output;
    ASSERT_EQ(absl::OkStatus(), TF_TensorToTensor(dst, &output)) << line;
    ASSERT_EQ(src.NumElements(), output.NumElements()) << line;
    for (int64_t i = 0; i < src.NumElements(); ++i) {
      ASSERT_EQ(data[i], output.flat<tstring>()(i)) << line;
    }

    TF_DeleteTensor(dst);
  }
}

----
<1> *Register Output*: The `process` task writes its output to a log file.
<2> *Modify Output*: The task writes to its output file as it executes.
<3> *Task Failure*: The task throws a `VerificationException` and fails at this point.
<4> *Continue to Modify Output*: This line never runs due to the exception stopping the task.

		gp21carry = regInfo{inputs: []regMask{gpg, gpg}, outputs: []regMask{gp, 0}}
		gp2flags  = regInfo{inputs: []regMask{gpg, gpg}}
		gp2flags1 = regInfo{inputs: []regMask{gp, gp}, outputs: []regMask{gp}}
		gp22      = regInfo{inputs: []regMask{gpg, gpg}, outputs: []regMask{gp, gp}}
		gp31      = regInfo{inputs: []regMask{gp, gp, gp}, outputs: []regMask{gp}}
		gp31carry = regInfo{inputs: []regMask{gp, gp, gp}, outputs: []regMask{gp, 0}}

          %0:2 = tf_executor.graph {
            %outputs, %control = tf_executor.island wraps "tf.Add"(%arg0, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
            %outputs_0, %control_1 = tf_executor.island wraps "tf.Add"(%outputs, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
            %outputs_2, %control_3 = tf_executor.island wraps "tf.Print"(%outputs_0) {message = "add result"} : (tensor<*xi32>) -> tensor<*xi32>

      %0:2 = tf_executor.graph {
        %outputs, %control = tf_executor.island wraps "tf.Add"(%arg0, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
        %outputs_0, %control_1 = tf_executor.island wraps "tf.Add"(%outputs, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
        %outputs_2, %control_3 = tf_executor.island wraps "tf.Print"(%outputs_0) {message = "add result"} : (tensor<*xi32>) -> tensor<*xi32>

becomes part of `outputs[partitions[js]]`.  The slices with `partitions[js] = i`
are placed in `outputs[i]` in lexicographic order of `js`, and the first
dimension of `outputs[i]` is the number of entries in `partitions` equal to `i`.
In detail,

```python
    outputs[i].shape = [sum(partitions == i)] + data.shape[partitions.ndim:]

    outputs[i] = pack([data[js, ...] for js if partitions[js] == i])