// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>
// CHECK:  [[MUL:%.*]] = tfl.mul(%arg0, [[CST]]) <{fused_activation_function = "NONE"}> : (tensor<3xbf16>, tensor<3x3xbf16>) -> tensor<3x3xbf16>
// CHECK:  return [[MUL]] : tensor<3x3xbf16>

  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  %3 = "tfl.div"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("div")

  %2 = "tfl.mul"(%0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  func.return %2 : tensor<4xf32>

// CHECK-LABEL: main
// CHECK-NOT: tfl.squared_difference
// CHECK: tfl.mul %[[CONST:.*]], %arg0

// conditions are also hoisted (i.e., `tf.Add` and `tf.Mul`).

// CHECK-LABEL: readvariableop_is_hoisted_if_readonly
// CHECK:       [[CST_0:%.*]] = "tf.Const"
// CHECK:       [[VAR:%.*]] = "tf.VarHandleOp"
// CHECK:       [[CST_1:%.*]] = "tf.Const"
// CHECK:       [[VAR_VAL:%.*]] = "tf.ReadVariableOp"
// CHECK:       [[RES_1:%.*]] = "tf.Add"([[VAR_VAL]], [[CST_0]])
// CHECK:       [[RES_2:%.*]] = "tf.Mul"([[RES_1]], [[CST_1]])

  // CHECK:  %[[IS_ZERO:.*]] = "tf.Equal"(%[[Y]], %[[ZERO]]) <{incompatible_shape_error = true}> : (tensor<3xf32>, tensor<f32>) -> tensor<3xi1>
  // CHECK:  %[[MUL:.*]] = "tf.Mul"(%[[X]], %[[Y]]) : (tensor<2x3xf32>, tensor<3xf32>) -> tensor<2x3xf32>
  // CHECK:  %[[RESULT:.*]] = "tf.SelectV2"(%[[IS_ZERO]], %[[ZERO]], %[[MUL]]) : (tensor<3xi1>, tensor<f32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  %0 = "tf.MulNoNan"(%arg0, %arg1) : (tensor<2x3xf32>, tensor<3xf32>) -> tensor<2x3xf32>

	// X=Z1; Y=Z1; MUL; T-   // T1 = Z1*Z1
	// X-  ; Y=T ; MUL; R=T  // R  = Z1*T1
	// X=X2; Y-  ; MUL; H=T  // H  = X2*T1
	// X=Z2; Y=Z2; MUL; T-   // T2 = Z2*Z2
	// X-  ; Y=T ; MUL; S1=T // S1 = Z2*T2
	// X=X1; Y-  ; MUL; U1=T // U1 = X1*T2
	// SUB(H<H-T)            // H  = H-U1
	// X=Z1; Y=Z2; MUL; T-   // Z3 = Z1*Z2

 	// X=Z1; Y=Z1; MUL; T-   // T1 = Z1*Z1
	// X-  ; Y=T ; MUL; R=T  // R  = Z1*T1
	// X=X2; Y-  ; MUL; H=T  // H  = X2*T1
	// X=Z2; Y=Z2; MUL; T-   // T2 = Z2*Z2
	// X-  ; Y=T ; MUL; S1=T // S1 = Z2*T2
	// X=X1; Y-  ; MUL; U1=T // U1 = X1*T2
	// SUB(H<H-T)            // H  = H-U1
	// X=Z1; Y=Z2; MUL; T-   // Z3 = Z1*Z2

  rewriter.clearInsertionPoint();
  return new_func;
}

// Outlines non-approximate GELU into a stablehlo composite.
//
//    -> mul 1/sqrt(2) -> erf -> add 1 ->
// in                                    mul
//    ---------> mul 0.5 --------------->
//
// This pattern assumes all binary ewise ops with one constant argument
// have that constant argument as the second operand. It works by

  %cst = arith.constant dense<[5, 7]> : tensor<2xi32>
  %0 = "tf.BroadcastTo"(%arg1, %cst) : (tensor<7xf32>, tensor<2xi32>) -> tensor<5x7xf32>
  %1 = "tf.Mul"(%arg0, %0) : (tensor<5x7xf32>, tensor<5x7xf32>) -> tensor<5x7xf32>
  func.return %1 : tensor<5x7xf32>
  // CHECK: %[[V0:.*]] = "tf.Mul"(%arg0, %arg1) : (tensor<5x7xf32>, tensor<7xf32>) -> tensor<5x7xf32>
  // CHECK: %[[V0]] : tensor<5x7xf32>
}

// CHECK-LABEL: @broadcast_mul1

TEXT time·now(SB),NOSPLIT,$0-24
	MOVD	$_INTERRUPT_TIME, R3
	MOVD	time_lo(R3), R0
	MOVD	$100, R1
	MUL	R1, R0
	MOVD	R0, mono+16(FP)

	MOVD	$_SYSTEM_TIME, R3
	MOVD	time_lo(R3), R0
	// convert to Unix epoch (but still 100ns units)
	#define delta 116444736000000000
	SUB	$delta, R0
	// Convert to nSec
	MOVD	$100, R1
	MUL	R1, R0

	// Code stolen from compiler output for:
	//
	//	var x uint64