func.func @slice_mhlo_sizes(%arg0: tensor<1x1024x4xf32>, %arg1: tensor<3xi32>) -> tensor<1x512x4xf32> {
  // CHECK-NOT: "tf.Slice"
  %0 = "mhlo.constant"() {value = dense<[1, 512, 4]> : tensor<3xi32>} : () -> tensor<3xi32>
  %1 = "tf.Slice"(%arg0, %arg1, %0) : (tensor<1x1024x4xf32>, tensor<3xi32>, tensor<3xi32>) -> tensor<1x512x4xf32>
  func.return %1 : tensor<1x512x4xf32>
}

// -----

// CHECK: %[[ADD:.+]] = stablehlo.add %[[CONV]], %[[BIAS_CONST]] : tensor<1x5x5x4xf32>
// CHECK: %[[MAX:.+]] = stablehlo.maximum %[[ADD]], %[[ZERO_CONST]] : tensor<1x5x5x4xf32>
// CHECK: %[[TRANSPOSE_1:.+]] = stablehlo.transpose %[[MAX]], dims = [0, 3, 1, 2] : (tensor<1x5x5x4xf32>) -> tensor<1x4x5x5xf32>
// CHECK: return %[[TRANSPOSE_1]]

// -----

func.func @add_with_activation_transpose_permutation_mismatch(
      %arg0: tensor<1x2x3x4xf32>) -> tensor<1x3x2x4xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<1x3x2x4xf32>
  %1 = stablehlo.transpose %arg0, dims = [0, 2, 1, 3] : (tensor<1x2x3x4xf32>) -> tensor<1x3x2x4xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x3x2x4xf32>
  return %2 : tensor<1x3x2x4xf32>
}
// CHECK: %[[TRANSPOSE_0:.+]] = stablehlo.transpose

func.func @mean_4d_no_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = false} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x512xf32>
  func.return %0 : tensor<1x512xf32>
}

// CHECK:       func @mean_4d_no_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

  func::FuncOp main_fn = FindMainFuncOp(*module_op);

  %flow_1 = "tf.TensorArrayWriteV3"(%handle, %index, %value, %flow) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<2x!tf_type.resource<tensor<1x512xf32>>>, tensor<i32>, tensor<1x512xf32>, tensor<f32>) -> tensor<f32>

      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

      func.return %%1 : tensor<1x4x4xf32>
    }
  })";
  std::string mat_mul_method =

  %cst1 = arith.constant dense<[[[2.0, 3.0, 4.0, 5.0]]]> : tensor<1x1x4xf32>

  %0 = "tfl.fully_connected" (%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x40x37xf32>, tensor<4x37xf32>, none) -> (tensor<1x40x4xf32>)
  %1 = "tfl.add"(%0, %cst1) {fused_activation_function = "NONE"} : (tensor<1x40x4xf32>, tensor<1x1x4xf32>) -> tensor<1x40x4xf32>

// CHECK-NEXT:    %[[BMM_0:.*]] = "tfl.batch_matmul"(%[[RESHAPED_0]], %[[RESHAPED_1]]) <{adj_x = false, adj_y = false, asymmetric_quantize_inputs = false}> : (tensor<3x5x12xf32>, tensor<3x12x4xf32>) -> tensor<3x5x4xf32>
// CHECK-NEXT:    %[[RESHAPED_BMM:.*]] = mhlo.reshape %[[BMM_0]]
// CHECK-NEXT:    return %[[RESHAPED_BMM]] : tensor<3x5x1x4xf32>
}

// CHECK:           %[[VAL_9:.*]] = "tf.Reshape"(%[[VAL_5]], %[[VAL_8]]) : (tensor<3x2x6x4xf32>, tensor<3xi64>) -> tensor<3x12x4xf32>
// CHECK:           %[[VAL_10:.*]] = "tf.BatchMatMulV3"(%[[VAL_7]], %[[VAL_9]]) <{adj_x = false, adj_y = false, grad_x = false, grad_y = false}> : (tensor<3x5x12xf32>, tensor<3x12x4xf32>) -> tensor<3x5x4xf32>
// CHECK:           %[[VAL_11:.*]] = arith.constant dense<[3, 5, 1, 4]> : tensor<4xi64>