// ScalarMult, but in this case the doublings are precomputed. The value
	// [windowValue]G added at iteration k would normally get doubled
	// (totIterations-k)×4 times, but with a larger precomputation we can
	// instead add [2^((totIterations-k)×4)][windowValue]G and avoid the
	// doublings between iterations.
	t := NewP224Point()
	p.Set(NewP224Point())
	tableIndex := len(tables) - 1
	for _, byte := range scalar {

	// ScalarMult, but in this case the doublings are precomputed. The value
	// [windowValue]G added at iteration k would normally get doubled
	// (totIterations-k)×4 times, but with a larger precomputation we can
	// instead add [2^((totIterations-k)×4)][windowValue]G and avoid the
	// doublings between iterations.
	t := NewP521Point()
	p.Set(NewP521Point())
	tableIndex := len(tables) - 1
	for _, byte := range scalar {

    }

  // Sort matched operations by their position in the block.
  llvm::sort(cluster.operations, [](Operation *a, Operation *b) -> bool {
    return a->isBeforeInBlock(b);
  });

  // Create tf_device::ClusterOp before the last operation in the block that
  // is a part of a match set.
  auto back = cluster.operations.back();
  auto loc = back->getLoc();
  OpBuilder builder(back);

  for (int i = 0; i < 1024; ++i) {
    RandomRemat remat(kNumOperations, kNumTensors, kNumUses, rng);
    // Worst-case scenario: we might double the length of the computation
    // schedule each time we remat, so only a few iterations...
    std::array<int, 3> randos;
    const auto& [begin, end, insert] = randos;
    for (int i = 0, num_operations = kNumOperations; i < 4;
         ++i, num_operations += end - begin) {

func MergePackageFiles(pkg *Package, mode MergeMode) *File {
	// Count the number of package docs, comments and declarations across
	// all package files. Also, compute sorted list of filenames, so that
	// subsequent iterations can always iterate in the same order.
	ndocs := 0
	ncomments := 0
	ndecls := 0
	filenames := make([]string, len(pkg.Files))
	var minPos, maxPos token.Pos
	i := 0
	for filename, f := range pkg.Files {

	numas := acc.sortAvailableNUMANodes()

	// Calculate the minimum and maximum possible number of NUMA nodes that
	// could satisfy this request. This is used to optimize how many iterations
	// of the loop we need to go through below.
	minNUMAs, maxNUMAs := acc.rangeNUMANodesNeededToSatisfy(cpuGroupSize)

	// Try combinations of 1,2,3,... NUMA nodes until we find a combination

//       (m[16:32]r² + m[48:64])r    <- lane 1
//                ^             ^
//                |             coefficients for second iteration
//                coefficients for first iteration
//
// So in this case we would have two iterations. In the first
// both lanes are multiplied by r². In the second only the
// first lane is multiplied by r² and the second lane is

  context:
    cluster: foobar
    user: a name
current-context: default
`,
			wantErr: true,
		},
	}

	for _, tt := range tests {
		// Use a closure so defer statements trigger between loop iterations.
		err := func() error {
			tempfile, err := ioutil.TempFile("", "")
			if err != nil {
				return err
			}
			p := tempfile.Name()
			defer utiltesting.CloseAndRemove(t, tempfile)

	{{break}}
		The innermost {{range pipeline}} loop is ended early, stopping the
		current iteration and bypassing all remaining iterations.

	{{continue}}
		The current iteration of the innermost {{range pipeline}} loop is
		stopped, and the loop starts the next iteration.

	{{template "name"}}
		The template with the specified name is executed with nil data.

	{{template "name" pipeline}}

	// ScalarMult, but in this case the doublings are precomputed. The value
	// [windowValue]G added at iteration k would normally get doubled
	// (totIterations-k)×4 times, but with a larger precomputation we can
	// instead add [2^((totIterations-k)×4)][windowValue]G and avoid the
	// doublings between iterations.
	t := NewP256Point()
	p.Set(NewP256Point())
	tableIndex := len(tables) - 1
	for _, byte := range scalar {