// not be a predecessor.
//
// Note that mergePredecessors behaves slightly differently between
// first and subsequent calls for a block.  For the first call, the
// starting state is approximated by taking the state from the
// predecessor whose state is smallest, and removing any elements not
// in all the other predecessors; this makes the smallest number of

			// 2, Avoid the predecessor that contains the function call, because the predecessor that
			//    contains the function call usually generates a lot of spills and lose the previous
			//    allocation state.
			// TODO: Improve this part. At least the size of endRegs of the predecessor also has
			// an impact on the code size and compiler speed. But it is not easy to find a simple

    for (Operation* op : ops_to_process) {
      if (predecessors) {
        for (Value operand : op->getOperands()) {
          // Stop at the block boundary.
          if (mlir::isa<BlockArgument>(operand)) continue;

          Operation* predecessor = operand.getDefiningOp();
          if (!operations->contains(predecessor) &&
              !ops_to_avoid.contains(predecessor)) {
            new_ops.insert(operand.getDefiningOp());

  }) {...} : () -> tensor<4xf32>
```
### `-tf-tpu-colocate-splits`

_Colocates each Split op with its predecessor_

It is beneficial for performance to assign a `Split` op to the same device
as its predecessor. This is because the weight of cut edges is always
minimized when the `Split` is with its predecessor. This colocation
constraint will be used by the placer graph optimization to assign a device
to the op.

		// If p and p.Succs[0] are dominators it means that every path
		// from entry to b passes through p and p.Succs[0]. We care that
		// no path from entry to b passes through p.Succs[1]. If p.Succs[0]
		// has one predecessor then (apart from the degenerate case),
		// there is no path from entry that can reach b through p.Succs[1].
		// TODO: how about p->yes->b->yes, i.e. a loop in yes.

				},
				`<a href="/foo?x=foo?x=%3cbar%3e?x=baz">`,
			},
		*/
	}

	// pred is a template function that returns the predecessor of a
	// natural number for testing recursive templates.
	fns := FuncMap{"pred": func(a ...any) (any, error) {
		if len(a) == 1 {
			if i, _ := a[0].(int); i > 0 {
				return i - 1, nil
			}
		}

std::vector<int> MarkForCompilationPassImpl::FindAlternatePathForDebugging(
    int from, int to) {
  std::vector<int> rpo = cycles_graph_.AllNodesInPostOrder();
  absl::c_reverse(rpo);

  // best_pred_for_node[n] contains a predecessor of `n` that has an
  // unclusterable node in some path from `from` to itself.
  // best_pred_for_node[n] is unpopulated for nodes that are not reachable from

        succ = curr.next;
        if (curr.thread != null) { // we aren't unlinking this node, update pred.
          pred = curr;
        } else if (pred != null) { // We are unlinking this node and it has a predecessor.
          pred.next = succ;
          if (pred.thread == null) { // We raced with another node that unlinked pred. Restart.
            continue restart;
          }

        succ = curr.next;
        if (curr.thread != null) { // we aren't unlinking this node, update pred.
          pred = curr;
        } else if (pred != null) { // We are unlinking this node and it has a predecessor.
          pred.next = succ;
          if (pred.thread == null) { // We raced with another node that unlinked pred. Restart.
            continue restart;
          }

====

[[using_junit5]]
== Using JUnit 5

http://junit.org/junit5[JUnit 5] is the latest version of the well-known JUnit test framework.
Unlike its predecessor, JUnit 5 is modularized and composed of several modules:

    JUnit 5 = JUnit Platform + JUnit Jupiter + JUnit Vintage