y1 := y[k:]       // y1 is normalized because y is
		t = t.mul(x0, y1) // update t so we don't lose t's underlying array
		addAt(z, t, k)

		// add xi*y0<<i, xi*y1*b<<(i+k)
		y0 := y0.norm()
		for i := k; i < len(x); i += k {
			xi := x[i:]
			if len(xi) > k {
				xi = xi[:k]
			}
			xi = xi.norm()
			t = t.mul(xi, y0)
			addAt(z, t, i)
			t = t.mul(xi, y1)
			addAt(z, t, i+k)
		}

		putNat(tp)

// CHECK-NEXT:     builtin_code: IF
// CHECK-NEXT:   }, {
// CHECK-NEXT:     version: 1
// CHECK-NEXT:   }, {
// CHECK-NEXT:     deprecated_builtin_code: 18,
// CHECK-NEXT:     version: 1,
// CHECK-NEXT:     builtin_code: MUL
// CHECK-NEXT:   } ],
// CHECK-NEXT:   subgraphs: [ {
// CHECK-NEXT:     tensors: [ {
// CHECK-NEXT:       shape: [ 1 ],
// CHECK-NEXT:       buffer: 1,
// CHECK-NEXT:       name: "arg0",

int64_t GetElementIndex(llvm::SmallVectorImpl<int64_t> &shape,
                        llvm::SmallVectorImpl<int64_t> &current_index) {
  int64_t ind = 0;
  int64_t mul = 1;
  for (int i = shape.size() - 1; i >= 0; --i) {
    ind += (current_index[i] % shape[i]) * mul;
    mul *= shape[i];
  }
  return ind;
}

// Helper method that increment index represented in 'current_index_ptr'
// in the shape of 'result_shape'.

  (Add32 (Const32 <t> [c*d]) (Mul32 <t> (Const32 <t> [c]) x))

// Rewrite x*y ± x*z  to  x*(y±z)
(Add(64|32|16|8) <t> (Mul(64|32|16|8) x y) (Mul(64|32|16|8) x z))
	=> (Mul(64|32|16|8) x (Add(64|32|16|8) <t> y z))
(Sub(64|32|16|8) <t> (Mul(64|32|16|8) x y) (Mul(64|32|16|8) x z))
	=> (Mul(64|32|16|8) x (Sub(64|32|16|8) <t> y z))

// rewrite shifts of 8/16/32 bit consts into 64 bit consts to reduce

  // We cannot fuse this tfl.mul into the preceding conv op because %cst2 is not broadcast-compatible to %cst0.
  %1 = "tfl.mul"(%0, %cst2) {fused_activation_function = "RELU6"} : (tensor<1x4x4x2xf32>, tensor<4x2xf32>) -> tensor<1x4x4x2xf32>

  func.return %1 : tensor<1x4x4x2xf32>

// CHECK:  %0 = "tfl.depthwise_conv_2d"(%arg0, %cst, %cst_0)
// CHECK:  %1 = tfl.mul(%0, %cst_1)
// CHECK:  return %1
}

  %5 = "tfl.mul"(%0, %1) {fused_activation_function = "NONE"} : (tensor<  f32>, tensor<  f32>) -> tensor<  f32>
  %6 = "tfl.mul"(%0, %3) {fused_activation_function = "NONE"} : (tensor<  f32>, tensor<4xf32>) -> tensor<4xf32>
  %7 = "tfl.mul"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<  f32>) -> tensor<4xf32>

//RUN: tf_tfl_translate --enable-stablehlo-conversion --input-mlir %s -o /tmp/temp.stablehlo; [ -f /tmp/temp.stablehlo ]


module {
func.func @main(%arg0: tensor<2xi32>) -> tensor<2xi32> {
  %0 = "tf.Mul"(%arg0, %arg0) : (tensor<2xi32>, tensor<2xi32>) -> tensor<2xi32>
  func.return %0 : tensor<2xi32>
}

// 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)

  %1 = "tf.BiasAdd"(%0, %bias) {T = "tfdtype$DT_FLOAT", data_format = "NHWC"}: (tensor<256x8x7x16xf32>, tensor<16xf32>) -> tensor<256x8x7x16xf32>
  %2 = "tf.Mul"(%1, %value) {T = "tfdtype$DT_FLOAT"} : (tensor<256x8x7x16xf32>, tensor<16xf32>) -> tensor<256x8x7x16xf32>
  func.return %2 : tensor<256x8x7x16xf32>

 Const 357
 Conv2D 53
 Conv2DBackpropFilter 53
 Conv2DBackpropInput 52
 DivNoNan 1
 Equal 1
 FusedBatchNorm 53
 FusedBatchNormGrad 53
 Identity 2
 MatMul 3
 MaxPool 1
 MaxPoolGrad 1
 Mean 1
 Mul 164
 Pad 1
 ReadVariableOp 646
 Relu 49
 ReluGrad 49
 Reshape 2
 ResourceApplyKerasMomentum 161
 ShapeN 50
 Softmax 1
 SparseSoftmaxCrossEntropyWithLogits 1
 Square 55
 Squeeze 1
 Sub 106
 Sum 57