case len(x) == 0:
		return z.set(y)
	case len(y) == 0:
		return z.set(x)
	}
	return z.mul(x, y)
}

// scaleDenom sets z to the product x*f.
// If f == 0 (zero value of denominator), z is set to (a copy of) x.
func (z *Int) scaleDenom(x *Int, f nat) {
	if len(f) == 0 {
		z.Set(x)
		return
	}
	z.abs = z.abs.mul(x.abs, f)
	z.neg = x.neg
}

// Cmp compares x and y and returns:
//

  // CHECK: _arithmetic_count = 1000 : i64
  %0 = "tfl.mul"(%arg0, %arg1) {fused_activation_function = "NONE"} : (tensor<10x10x10xf32>, tensor<10x10x10xf32>) -> tensor<10x10x10xf32>
  func.return %0 : tensor<10x10x10xf32>
}

func.func @testMulBroadcast(%arg0: tensor<10x10x10xf32>, %arg1: tensor<10xf32>) -> tensor<10x10x10xf32> {
  // CHECK: _arithmetic_count = 1000 : i64

(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))])

  func.return %5 : tensor<!tf_type.variant>
}

func.func @"__inference_Dataset_map_<lambda>_80"(%arg0: tensor<*xi32>) -> tensor<*xi32> {
  %0 = "tf.Const"() {value = dense<2> : tensor<i32>} : () -> tensor<i32>
  %1 = "tf.Mul"(%arg0, %0) {device = ""} : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
  %2 = "tf.Identity"(%1) {device = ""} : (tensor<*xi32>) -> tensor<*xi32>
  func.return %2 : tensor<*xi32>
}

(Select1 (Sub32carry <t> x y)) => (SGTU <typ.Bool> (SUB <t.FieldType(0)> x y) x)
(Sub32withcarry <t> x y c) => (SUB (SUB <t> x y) c)

(Mul(32|16|8) ...) => (MUL ...)
(Mul(32|64)F ...) => (MUL(F|D) ...)

(Hmul(32|32u) x y) => (Select0 (MUL(T|TU) x y))
(Mul32uhilo ...) => (MULTU ...)

(Div32 x y) => (Select1 (DIV x y))
(Div32u x y) => (Select1 (DIVU x y))

    return success();
  }
};

// Fuse Mul with proceeding FullyConnected.
// Replace ..
// Mul(FC(input, filter, bias), rhs)
// .. with ..
// FC(lhs, Mul(filter, rhs), bias)
// .. if rhs, filter, and bias are all constants.
// The generated Mul will be constant folded to a single matrix using TF::Mul.
// TODO(b/136285429): Move to tablegen when variadic is supported

	AndAnd // &&

	// precCmp
	Eql // ==
	Neq // !=
	Lss // <
	Leq // <=
	Gtr // >
	Geq // >=

	// precAdd
	Add // +
	Sub // -
	Or  // |
	Xor // ^

	// precMul
	Mul    // *
	Div    // /
	Rem    // %
	And    // &
	AndNot // &^
	Shl    // <<
	Shr    // >>
)

// Operator precedences
const (
	_ = iota
	precOrOr
	precAndAnd
	precCmp

}

Status MulModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs) {
  return ops::Mul(ctx, inputs[0], inputs[1], &outputs[0], "Mul");
}

Status Log1pModel(AbstractContext* ctx,
                  absl::Span<AbstractTensorHandle* const> inputs,
                  absl::Span<AbstractTensorHandle*> outputs) {

func (a *int64Amount) Sub(b int64Amount) bool {
	return a.Add(int64Amount{value: -b.value, scale: b.scale})
}

// Mul multiplies the provided b to the current amount, or
// returns false if overflow or underflow would result.
func (a *int64Amount) Mul(b int64) bool {
	switch {
	case a.value == 0:
		return true
	case b == 0:
		a.value = 0
		a.scale = 0
		return true
	case a.scale == 0:

    $t, $m, $v, $beta, $gamma, F32Attr:$variance_epsilon, ConstBoolAttrTrue:$scale_after_normalization),
  (TF_AddOp
    (TF_MulOp $t, (TF_MulOp:$mul (TF_RsqrtOp (TF_AddOp $v, (TF_ConstOp $variance_epsilon))), $gamma)),
    (TF_SubOp $beta, (TF_MulOp $m, $mul)))>;

class TFi32<int v> : ConstantAttr<I32ElementsAttr, !cast<string>(v)>;

// Matmul without transpose on b to matmul with explicit transpose op and
// transposed b.