}
    }
  });
}

// tf.NoOp islands are used to combine multiple control dependencies into one.
// These islands have a single tf.NoOp inside them and consume multiple control
// outputs to generate a single control output.
//
// For example,
// ```
// %merged_control = "tf_executor.island"(%control_a, %control_b) ({
//   "tf.NoOp"() : () -> ()
//   "tf_executor.yield"() : () -> ()

  }];
  let constructor = "TF::CreateRemoveUnusedWhileResultsPass()";
}

def HoistLoopInvariantPass : Pass<"tf-hoist-loop-invariant", "mlir::func::FuncOp"> {
  let summary = "Hoists loop invariant ops to the outside of the loop";
   let description = [{
      Hoists loop invariant to the outside of the loop. The pass is similar to
      LoopInvariantCodeMotion pass, but it also hoists ReadVariableOps,
      if the variable is read only.

package ssa

// loopRotate converts loops with a check-loop-condition-at-beginning
// to loops with a check-loop-condition-at-end.
// This helps loops avoid extra unnecessary jumps.
//
//	 loop:
//	   CMPQ ...
//	   JGE exit
//	   ...
//	   JMP loop
//	 exit:
//
//	  JMP entry
//	loop:
//	  ...
//	entry:
//	  CMPQ ...
//	  JLT loop
func loopRotate(f *Func) {
	loopnest := f.loopnest()

  for (int i = 0; i < n; ++i) result.push_back(ToOutput(output_tensors[i]));
  return result;
}

// The backprop loop counter and main backprop loop run in their own execution
// frame (conceptually, the main forward loop and forward loop counter run
// together in a frame, then the backprop loop counter and backprop loop run
// together in a different frame). This returns the frame name to use for the
// backprop while loops.

it does not call the loop body again after it has returned false or
after the entire loop has exited (it might retain a copy of the body
function, or pass it to another goroutine) -- each generated loop has
its own #stateK variable that is used to check for permitted call
patterns to the yield function for a loop body.

The state values are:

abi.RF_DONE = 0      // body of loop has exited in a non-panic way

				}
				i++
			}
			if i != 4 {
				t.Fatalf("should loop four times")
			}
		}},
		{"int8", ValueOf(int8(4)), func(t *testing.T, s iter.Seq[Value]) {
			i := int8(0)
			for v := range s {
				if v.Interface().(int8) != i {
					t.Fatalf("got %d, want %d", v.Int(), i)
				}
				i++
			}
			if i != 4 {
				t.Fatalf("should loop four times")
			}
		}},

// For best code generation, RF_DONE and RF_READY should be 0 and 1.
const (
	RF_DONE          = RF_State(iota) // body of loop has exited in a non-panic way
	RF_READY                          // body of loop has not exited yet, is not running  -- this is not a panic index
	RF_PANIC                          // body of loop is either currently running, or has panicked
	RF_EXHAUSTED                      // iterator function return, i.e., sequence is "exhausted"

	// R5 = b_len
	// R6 = b_cap (unused)
	// R7 = byte to find
	AND	$0xff, R7
	MOVV	R4, R6		// store base for later
	ADDV	R4, R5		// end
	ADDV	$-1, R4

	PCALIGN	$16
loop:
	ADDV	$1, R4
	BEQ	R4, R5, notfound
	MOVBU	(R4), R8
	BNE	R7, R8, loop

	SUBV	R6, R4		// remove base
	RET

notfound:
	MOVV	$-1, R4
	RET

TEXT ·IndexByteString<ABIInternal>(SB),NOSPLIT,$0-32
	// R4 = s_base
	// R5 = s_len

	CMP   R5, $0
	MOVD  R5, CTR         // Set up loop counter
	BEQ   tail            // If R5 = 0, we can't use the loop

	// Process 4 elements per iteration. Unrolling this loop
	// means a performance trade-off: we will lose performance
	// for small values of z_len (0.90x in the worst case), but
	// gain significant performance as z_len increases (up to
	// 1.45x).

	PCALIGN $16
loop:
	MOVD  8(R8), R11      // R11 = x[i]

    %control_2 = tf_executor.island wraps "tf.NoOp"() {_pivot_for_cluster = "cluster", device = ""} : () -> ()
    // CHECK: %control_3 = tf_executor.island(%control_2) wraps "tf.NoOp"() {_tpu_replicate = "cluster", device = ""} : () -> ()
    %control_3 = tf_executor.island(%control_2) wraps "tf.NoOp"() {_replication_info = "cluster", _xla_compile_device_type = "TPU", device = ""} : () -> ()