[[sec:task_output_caching_details]]
== Cacheable tasks

Since a task describes all of its inputs and outputs, Gradle can compute a _build cache key_ that uniquely defines the task's outputs based on its inputs.
That build cache key is used to request previous outputs from a build cache or store new outputs in the build cache.

  func.func @main() attributes {tf.entry_function = {inputs = "", outputs = ""}, tf_saved_model.exported_names = ["main"]} {
    tf_executor.graph {
      tf_executor.fetch
    }
    return
  }
  // CHECK: @main(%[[ARG_0:.*]]: tensor<!tf_type.string> {tf_saved_model.bound_input = @__tf_saved_model_asset0_file.txt})
  // CHECK-SAME: tf.entry_function = {inputs = "init_op_0:0", outputs = ""}

    }
}
----
=====
====

== Understanding inputs and outputs

Task inputs and outputs are important for:

1. *Up-to-date checks* - Gradle’s incremental build feature helps your build avoid doing the same work more than once by looking at changes for task inputs and outputs.
2. *Linking task inputs and outputs* - When outputs of one task are linked to the inputs of another, Gradle can automatically create task dependencies.

            withArgument("--max-workers=$threadCount")
        }
    }

    def "overlapping outputs prevent parallel execution"() {
        given:
        withParallelThreads(2)

        and:
        buildFile << """
            aPing.outputs.dir "dir"
            bPing.outputs.file "dir/file"
        """
        expect:
        2.times {

        result.assertTasksNotSkipped(assembleAppTasks)

        outputs.deletedClasses("multiply", "sum")

        // See https://github.com/gradle/gradle-native/issues/1004
        if (toolchainUnderTest.version.major == 5) {
            outputs.recompiledClasses('renamed-sum')
        } else {
            outputs.recompiledClasses('greeter', 'renamed-sum', 'main')
        }

	k = len(i.outputs)
	for _, p := range a.abiInfo.OutParams() {
		for _, r := range p.Registers {
			m := archRegForAbiReg(r, c)
			a.reg.outputs = append(a.reg.outputs, outputInfo{idx: k, regs: (1 << m)})
			k++
		}
	}
	a.reg.outputs = append(a.reg.outputs, i.outputs...)
	a.reg.clobbers = i.clobbers
	return a.reg
}

        def clean = rootBuild.task('clean').dependsOn(cleanFoo).dependsOn(cleanBar)
        def generateFoo = foo.task('generateFoo').outputs('build/foo')
        def generateBar = bar.task('generateBar').outputs('build/bar')
        def generate = rootBuild.task('generate').dependsOn(generateBar).dependsOn(generateFoo)

        // conflicts with command line order

    // Add StopGradients according to 'stop_outputs'.
    auto out0 =
        stop_outputs[0] ? StopGradient(scope_, y.output[0]) : y.output[0];
    auto out1 =
        stop_outputs[1] ? StopGradient(scope_, y.output[1]) : y.output[1];
    auto out2 =
        stop_outputs[2] ? StopGradient(scope_, y.output[2]) : y.output[2];

    auto g0 = Const(scope_, {1, 2, 3, 4, 5, 6, 7, 8}, {2, 4});

    std::cout << '\'' << subgraph_name << "' outputs:\n";
    mlir::Operation *terminator = nullptr;
    if (subgraph->name()) {
      if (auto fn = module.lookupSymbol<FuncOp>(subgraph->name()->str()))
        terminator = fn.back().getTerminator();
    }
    i = 0;
    for (auto output : *subgraph->outputs()) {
      print_buffer(*subgraph, i, output, [&](int i) {

  return absl::OkStatus();
}

// Converts `noutputs` and `outputs` into `outputs_tensors` and does various
// checks while doing so.
Status ProcessOutputs(const TF_Graph* fn_body, const char* fn_name,
                      int noutputs, const TF_Output* outputs,
                      std::vector<OutputTensor>* output_tensors)
    TF_EXCLUSIVE_LOCKS_REQUIRED(fn_body->mu) {
  output_tensors->reserve(noutputs);
  for (int i = 0; i < noutputs; ++i) {