//
	// setup creates a fake scavengeIndex that can be mutated and queried by
	// the functions it returns. Those functions capture the testing.T that
	// setup is called with, so they're bound to the subtest they're created in.
	//
	// Tests are then organized into test cases which mark some pages as
	// scavenge-able then try to find them. Tests expect that the initial

                        ctx.ApplyConfigOrFail(ctx, nil, mycfg)
                        // Do more stuff here.
                    })
            }
        })
}
```

Under the hood, calling `subtest.Run()` delegates to `t.Run()` in order to create a child `testing.T`.

### Parallel Tests

Many tests can take a while to start up for a variety of reasons, such as waiting for pods to start or waiting

# # whether the latter is stale.
# dialectgen(
#     name = "tf-generated-ops",
#     output = "new_tf_generated_ops.td",
# )
#
# sh_test(
#     name = "tf_generated_ops_not_stale",
#     srcs = ["ir/tf_generated_ops_not_stale.sh"],
#     args = [
#         "$(location //tensorflow/compiler/mlir/tensorflow:ir/tf_generated_ops.td)",

			t.Fatal(err)
		}
		dir = d
		gopath = d
	}

	exe := buildPtrTests(t, gopath, false)
	exe2 := buildPtrTests(t, gopath, true)

	// We (TestPointerChecks) return before the parallel subtest functions do,
	// so we can't just defer os.RemoveAll(dir). Instead we have to wait for
	// the parallel subtests to finish. This code looks racy but is not: