// CHECK: %[[LOCAL_BARRIER:.*]] = tf_executor.island(%[[CONTROL_ASSIGN_VAR_RES_0_0]], %[[CONTROL_ASSIGN_VAR_RES_1_0]], %[[CONTROL_ADA_GRAD]]) wraps "tf.NoOp"() : () -> ()
    %local_barrier = tf_executor.island(%assign_control_0, %assign_control_1, %apply_grad_control) wraps "tf.NoOp"() : () -> ()

operations in the while loop body are dependent on which inputs, it captures
inter iteration parallelism in while loop. Control dependencies on the other
hand create a barrier at the end of while loop body thus blocking any
parallelism across iterations.

For example, the following while loop body has a `%barrier` at the end.
Although there is no data/control dependency between `tf.AssignVariableOp`

        //   ApplyWeightUpdates_1 -> Computation_B
        //   ApplyWeightUpdates_2 -> Computation_C
        //   Computation_A -> NoOp
        //   Computation_B -> NoOp
        //   Computation_C -> NoOp
        //   "iteration++" -> NoOp
        // }
        //
        // In the graph above we can't cluster iteration++ with any of the

		// generates STXV instructions on
		// power9. The Short variation is used
		// if no loop is generated.

		// sizes >= 64 generate a loop as follows:

		// Set up loop counter in CTR, used by BC
		// XXLXOR clears VS32
		//       XXLXOR VS32,VS32,VS32
		//       MOVD len/64,REG_TMP
		//       MOVD REG_TMP,CTR
		//       loop:
		//       STXV VS32,0(R20)
		//       STXV VS32,16(R20)

  // Fourth, add all the inputs to sc_forward(). These all come from the
  // while loop operands or the non_tpu() call that's in the loop body. The
  // loop operands need to be pulled from the "i" (or "2") version of the
  // inputs. The inputs coming from non_tpu() are from the same loop iteration
  // (non_tpu_res_i).
  std::vector<Value> f_operands;

	bucketMax_8_64(text, freq, bucket)

	numLMS := 0
	lastB := int64(-1)
	bucket = bucket[:256] // eliminate bounds check for bucket[c1] below

	// The next stanza of code (until the blank line) loop backward
	// over text, stopping to execute a code body at each position i
	// such that text[i] is an L-character and text[i+1] is an S-character.
	// That is, i+1 is the position of the start of an LMS-substring.

              "{#true,&,*iv0/cond:0}<loop>");
    EXPECT_EQ(predicate_map[ControlOutputFor(dependent_iv0)],
              "{#true,&,(iv0/iv:0 & *iv0/cond:0)}<loop>");
    EXPECT_EQ(predicate_map[ControlOutputFor(dependent_iv1)],
              "{#true,&,(iv0/iv:0 & *iv0/cond:0)}<loop>");
    EXPECT_EQ(predicate_map[ControlOutputFor(add0)],
              "{#true,&,(iv0/iv:0 & *iv0/cond:0)}<loop>");
  }
}

  TF_Output cond_output;

  // The loop body graph. The inputs are the current values of the loop
  // variables. The outputs are the updated values of the loop variables.
  TF_Graph* const body_graph;
  const TF_Output* const body_inputs;
  TF_Output* const body_outputs;

  // Unique null-terminated name for this while loop. This is used as a prefix
  // for created operations.
  const char* name;
} TF_WhileParams;

		//		nxt = inc + ind
		//		goto loop
		//
		//	exit_loop:
		//
		// is rewritten to:
		//
		//	loop:
		//		ind = (Phi end nxt)
		//		if (Const [x]) < ind
		//		then goto enter_loop
		//		else goto exit_loop
		//
		//	enter_loop:
		//		do something without using ind nor nxt
		//		nxt = ind - inc
		//		goto loop
		//
		//	exit_loop:
		//

//
// `X` may contain symbolic predicates and the operations corresponding to these
// symbolic predicates are either in frame `loop` or outside it.  The symbols
// that are inside frame `loop` are loop variant (i.e. can have different
// liveness in each loop iteration) and the symbols that are outside frame
// `loop` are loop invariant (i.e. have the same liveness across all
// iterations).
class AndRecurrencePredicate : public Predicate {
 public: