StringRef reduce_op = all_reduce.getReduceOp();

    StringRef merge_op, final_op;
    if (reduce_op == "Add") {
      merge_op = "Add";
      final_op = "Id";
    } else if (reduce_op == "Mul") {
      merge_op = "Mul";
      final_op = "Id";
    } else if (reduce_op == "Min") {
      merge_op = "Min";
      final_op = "Id";
    } else if (reduce_op == "Max") {
      merge_op = "Max";
      final_op = "Id";

import org.jetbrains.kotlin.lexer.KtTokens.GTEQ
import org.jetbrains.kotlin.lexer.KtTokens.LBRACKET
import org.jetbrains.kotlin.lexer.KtTokens.LT
import org.jetbrains.kotlin.lexer.KtTokens.LTEQ
import org.jetbrains.kotlin.lexer.KtTokens.MUL
import org.jetbrains.kotlin.lexer.KtTokens.NULL_KEYWORD
import org.jetbrains.kotlin.lexer.KtTokens.OROR
import org.jetbrains.kotlin.lexer.KtTokens.QUEST
import org.jetbrains.kotlin.lexer.KtTokens.RBRACKET

	IDENT  // main
	INT    // 12345
	FLOAT  // 123.45
	IMAG   // 123.45i
	CHAR   // 'a'
	STRING // "abc"
	literal_end

	operator_beg
	// Operators and delimiters
	ADD // +
	SUB // -
	MUL // *
	QUO // /
	REM // %

	AND     // &
	OR      // |
	XOR     // ^
	SHL     // <<
	SHR     // >>
	AND_NOT // &^

	ADD_ASSIGN // +=
	SUB_ASSIGN // -=
	MUL_ASSIGN // *=
	QUO_ASSIGN // /=

    func.return %0 : tensor<4xf32>
    // CHECK:  [[VAL_0:%.*]] = tfl.sub %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<4xf32>
    // CHECK:  [[VAL_1:%.*]] = tfl.mul [[VAL_0]], [[VAL_0]] {fused_activation_function = "NONE"} : tensor<4xf32
  }

  // CHECK-LABEL: pack
  func.func @pack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> {

	return e
}

// Sub sets e = t1 - t2, and returns e.
func (e *P224Element) Sub(t1, t2 *P224Element) *P224Element {
	p224Sub(&e.x, &t1.x, &t2.x)
	return e
}

// Mul sets e = t1 * t2, and returns e.
func (e *P224Element) Mul(t1, t2 *P224Element) *P224Element {
	p224Mul(&e.x, &t1.x, &t2.x)
	return e
}

// Square sets e = t * t, and returns e.
func (e *P224Element) Square(t *P224Element) *P224Element {

    %outputs_11, %control_12 = tf_executor.island wraps "tf.Mul"(%outputs, %outputs_9) {_replication_info = "cluster", _xla_compile_device_type = "TPU", device = ""} : (tensor<f32>, tensor<*xf32>) -> tensor<*xf32>

		println(s[0])
		panic("s[0] != 1")
	}
	if s[1] != 2+3i {
		println(s[1])
		panic("s[1] != 2+3i")
	}
	if s[2] != 4+5i {
		println(s[2])
		panic("s[2] != 4+5i")
	}

	x := 1 + 2i
	y := a.Mul(x)
	if y != (1+2i)*(3+4i) {
		println(y)
		panic("y != (1+2i)*(3+4i)")
	}

	_ = x[OrOr-5]
	_ = x[AndAnd-6]
	_ = x[Eql-7]
	_ = x[Neq-8]
	_ = x[Lss-9]
	_ = x[Leq-10]
	_ = x[Gtr-11]
	_ = x[Geq-12]
	_ = x[Add-13]
	_ = x[Sub-14]
	_ = x[Or-15]
	_ = x[Xor-16]
	_ = x[Mul-17]
	_ = x[Div-18]
	_ = x[Rem-19]
	_ = x[And-20]
	_ = x[AndNot-21]
	_ = x[Shl-22]
	_ = x[Shr-23]
}

const _Operator_name = ":!<-~||&&==!=<<=>>=+-|^*/%&&^<<>>"

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU"} : tensor<100xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU"} : tensor<100xf32>
    func.return %1 : tensor<100xf32>
  }

	//	}
	ADD	R1, R2

	//	LADDW imm ',' rreg
	//	{
	//		outcode(int($1), &$2, 0, &$4);
	//	}
	ADD	$4, R1

	//	LMUL rreg ',' rreg
	//	{
	//		outcode(int($1), &$2, 0, &$4);
	//	}
	MUL	R1, R2

	//	LSHW rreg ',' sreg ',' rreg
	//	{
	//		outcode(int($1), &$2, int($4), &$6);
	//	}
	SLL	R1, R2, R3

	//	LSHW rreg ',' rreg
	//	{
	//		outcode(int($1), &$2, 0, &$4);
	//	}