swift-library-variant-task-graph.png...

cpp-library-variant-task-graph.png...

  compileTestVariantCpp -> linkTestVariant -> installTestVariant -> runTestVariant [dir=back, style=dashed]
  compileTestVariantCpp[label=<compile<i>Variant</i>Cpp>]
  linkTestVariant[label=<link<i>Variant</i>>]
  installTestVariant[label=<installTest<i>Variant</i>>]
  runTestVariant[label=<runTest<i>Variant</i>>]

cpp-application-variant-task-graph.png...

swift-application-variant-task-graph.png...

cpp-unit-test-variant-task-graph.png...

As a consequence, if we want our instrumented classes to be used in place of this variant when executing tests, we need to attach similar attributes to our variant.
In fact, the attribute we care about is `org.gradle.libraryelements` which explains _what the variant contains_, so we can setup the variant this way:

.Declaring the variant attributes
====

Finally, Gradle selects the appropriate variant by looking at the variant attributes:

- the consumer wants a variant with attributes `org.gradle.usage=java-api`
- the producer has a matching variant (`apiElements`)
- the producer has a non-matching variant (`runtimeElements`)

Gradle provides the artifacts and dependencies from the `apiElements` variant to the consumer.

== A more complicated example

    "tf.Yield"(%barg1, %barg2) : (tensor<!tf_type.variant<tensor<?x1xf32>>>, tensor<!tf_type.variant<tensor<?x1xf32>>>) -> ()
  }) {is_stateless = false} : (tensor<!tf_type.variant<tensor<?x1xf32>>>, tensor<!tf_type.variant<tensor<?x1xf32>>>) -> (tensor<!tf_type.variant<tensor<?x1xf32>>>, tensor<!tf_type.variant<tensor<?x1xf32>>>)
  func.return
}

// CHECK: while_region_op_two_sep_args_empty_tensor_list

  %1 = "tf.AddN"(%arg0, %arg1) : (tensor<!tf_type.variant<tensor<*xi32>>>, tensor<!tf_type.variant<tensor<*xi32>>>) -> tensor<!tf_type.variant<tensor<*xi32>>>