// CHECK: [[ACCUM_MOMENTUM:%.*]] = "tf.Mul"([[ACCUM]], [[MOMENTUM]])
    // CHECK: [[ACCUM_NEW:%.*]] = "tf.AddV2"([[ACCUM_MOMENTUM]], [[GRAD]])
    // CHECK: "tf.AssignVariableOp"([[ACCUM_HANDLE]], [[ACCUM_NEW]])
    // CHECK: [[GRAD_LR:%.*]] = "tf.Mul"([[GRAD]], [[LR]])
    // CHECK: [[MOMENTUM_LR:%.*]] = "tf.Mul"([[MOMENTUM]], [[LR]])
    // CHECK: [[ACCUM_NEW_MOMENTUM_LR:%.*]] = "tf.Mul"([[ACCUM_NEW]], [[MOMENTUM_LR]])

                        ZerosLike(scope, x));
    auto gy_1 = Mul(scope, Mul(scope, grad, z), log_x);
    return BinaryGradCommon(scope, op, grad_outputs, gx_1, gy_1);
  } else {
    // There's no sensible real value to return if x < 0, so return 0
    auto log_x = Where3(scope, Greater(scope, x, zero), Log(scope, x),
                        ZerosLike(scope, x));
    auto gy_1 = Mul(scope, Mul(scope, grad, z), log_x);

  // 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

    %5 = "tf.AddV2"(%4, %arg2) : (tensor<1x3x2x2xi32>, tensor<2xi32>) -> tensor<1x3x2x2xi32>
    %6 = "tf.Mul"(%arg3, %arg5) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %7 = "tf.Div"(%6, %arg9) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %8 = "tf.Cast"(%5) {Truncate = false} : (tensor<1x3x2x2xi32>) -> tensor<1x3x2x2xf32>
    %9 = "tf.Mul"(%7, %8) : (tensor<f32>, tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

    %add, %add_control = tf_executor.island wraps "tf.Add"(%arg2, %arg3) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    // CHECK: %[[MUL:.*]], %{{.*}} = tf_executor.island wraps "tf.Mul"(%[[ARG_2]], %[[ARG_3]]) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %mul, %mul_control = tf_executor.island wraps "tf.Mul"(%arg2, %arg3) : (tensor<f32>, tensor<f32>) -> tensor<f32>

// so it must be used with care.
(MUL(Q|L)const [-9] x) => (NEG(Q|L) (LEA(Q|L)8 <v.Type> x x))
(MUL(Q|L)const [-5] x) => (NEG(Q|L) (LEA(Q|L)4 <v.Type> x x))
(MUL(Q|L)const [-3] x) => (NEG(Q|L) (LEA(Q|L)2 <v.Type> x x))
(MUL(Q|L)const [-1] x) => (NEG(Q|L) x)
(MUL(Q|L)const [ 0] _) => (MOV(Q|L)const [0])
(MUL(Q|L)const [ 1] x) => x
(MUL(Q|L)const [ 3] x) => (LEA(Q|L)2 x x)
(MUL(Q|L)const [ 5] x) => (LEA(Q|L)4 x x)

		{0, 1, 2, -1, '-', (*Float).Sub},
		{1, 2, 0, 2, '-', (*Float).Sub},
		{2, 0, 1, -1, '-', (*Float).Sub},

		{0, 0, 0, 0, '*', (*Float).Mul},
		{0, 1, 2, 2, '*', (*Float).Mul},
		{1, 2, 0, 0, '*', (*Float).Mul},
		{2, 0, 1, 0, '*', (*Float).Mul},

		// {0, 0, 0, 0, '/', (*Float).Quo}, // panics
		{0, 2, 1, 2, '/', (*Float).Quo},
		{1, 2, 0, 0, '/', (*Float).Quo}, // = +Inf
		{2, 0, 1, 0, '/', (*Float).Quo},

(MOVWnop (MOVWconst [c])) => (MOVWconst [c])

// mul by constant
(MUL x (MOVWconst [c])) && int32(c) == -1 => (RSBconst [0] x)
(MUL _ (MOVWconst [0])) => (MOVWconst [0])
(MUL x (MOVWconst [1])) => x
(MUL x (MOVWconst [c])) && isPowerOfTwo32(c) => (SLLconst [int32(log32(c))] x)
(MUL x (MOVWconst [c])) && isPowerOfTwo32(c-1) && c >= 3 => (ADDshiftLL x x [int32(log32(c-1))])