%0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ijk,j->i"}: (tensor<4x5x6xf32>, tensor<5xf32>) -> tensor<4xf32>
  func.return %0 : tensor<4xf32>
// CHECK-LABEL: einsum_no_match
// CHECK: %[[v0:.*]] = "tf.Einsum"(%arg0, %arg1) <{equation = "ijk,j->i"}> {T = "tfdtype$DT_FLOAT"} : (tensor<4x5x6xf32>, tensor<5xf32>) -> tensor<4xf32>
// CHECK: return %[[v0]]
}

    %1 = stablehlo.get_dimension_size %0, dim = 0 : (tensor<?x3x4x2xf32>) -> tensor<i32>
    %2 = stablehlo.reshape %1 : (tensor<i32>) -> tensor<1xi32>
    %3 = stablehlo.concatenate %2, %cst_0, %cst_1, %cst_2, dim = 0 : (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<4xi32>
    %4 = stablehlo.dynamic_broadcast_in_dim %arg2, %3, dims = [0, 1, 2, 3] : (tensor<1x1x1x2xf32>, tensor<4xi32>) -> tensor<?x3x4x2xf32>

  func.return %add : tensor<*xf32>
}
// CHECK: func @decompose_batch_norm
// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<2.49743462E-5> : tensor<2xf32>}> : () -> tensor<2xf32>

// CHECK-LABEL: import_stats_skip
func.func @import_stats_skip(%arg0: tensor<4xf32>, %cst: tensor<i32>) -> (tensor<2xf32>,tensor<2xf32>) {
  %0:2 = "tfl.split"(%cst, %arg0) {num_splits = 2 : i32} : (tensor<i32>, tensor<4xf32>) -> (tensor<2xf32>, tensor<2xf32>)
    loc(fused["skip1", "skip2.cc":10:8, callsite("op" at "skip3.cc":10:8)])
  func.return %0#0, %0#1 : tensor<2xf32>, tensor<2xf32>

// CHECK-NEXT: "tfl.split"
// CHECK-NEXT: return
}

  // CHECK-DAG: %[[IND_SLICE1_START:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK-DAG: %[[IND_SLICE1_SIZE:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK: %[[IND_SLICE1:.*]] = "tf.Slice"(%[[INDS]], %[[IND_SLICE1_START]], %[[IND_SLICE1_SIZE]]) : (tensor<2xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  // CHECK: %[[SLICE1_START:.*]] = "tf.ConcatV2"(%[[IND_SLICE1]],

  %cst_2 = "tf.Const"() {value = dense<3> : tensor<1xi32>} : () -> tensor<1xi32>
  %cst_3 = "tf.Const"() {value = dense<4> : tensor<1xi32>} : () -> tensor<1xi32>
  %0 = "tf.XlaReduceWindow"(%arg0, %arg1, %cst_0, %cst_1, %cst_2, %cst_3, %cst) {computation = @sum_reducer} : (tensor<7xf32>, tensor<f32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1x2xi32>) -> tensor<10xf32>

      Targs = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>],
      Tresults = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>]}
      : (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>, tensor<2x!tf_type.string>)
        -> (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>)
    tf_device.return %0#0, %0#1, %0#2 : tensor<32xf32>, tensor<64xf32>, tensor<8xf32>

// CHECK-NEXT:  %[[quant:.*]] = "tfl.quantize"(%[[split]]#0) <{qtype = tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>}> {volatile} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>
// CHECK-NEXT:  return %[[quant]] : tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>

    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()
    %cst_1 = "tf.Const"() {value = dense<2.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>