%1 = mhlo.multiply %0, %cst1 : tensor<1x1x2x4xf32>
  // CHECK:      return %[[RES]] : tensor<1x1x2x4xf32>
  func.return %1 : tensor<1x1x2x4xf32>
}

// CHECK-LABEL: @foldBroadcastInDimBeforeMulOp_bcast_dim_2D_float
func.func @foldBroadcastInDimBeforeMulOp_bcast_dim_2D_float() -> (tensor<1x2x2x3xf32>) {

  %6 = "tfl.broadcast_to"(%arg1, %4) : (tensor<8x7x6x5x?x3x2x1xf32>, tensor<8xi64>) -> tensor<8x7x6x5x?x3x2x1xf32>
  %7 = "tfl.broadcast_to"(%arg2, %4) : (tensor<?x3x2x1xf32>, tensor<8xi64>) -> tensor<8x7x6x5x?x3x2x1xf32>
  %8 = "tfl.select_v2"(%5, %6, %7) : (tensor<8x7x6x5x?x3x2x1xi1>, tensor<8x7x6x5x?x3x2x1xf32>, tensor<8x7x6x5x?x3x2x1xf32>) -> tensor<8x7x6x5x?x3x2x1xf32>
  func.return %8 : tensor<8x7x6x5x?x3x2x1xf32>
}

// -----

// Contains the `stablehlo.transpose` op of the arg (e.g. [b, f, 0, 1] to
// [b, 0, 1, f]). The weight constant is folded into [0, 1, i, o] format.
// CHECK-DAG: %[[CST:.+]] = stablehlo.constant dense<3.000000e+00> : tensor<3x3x8x8xf32>
// CHECK: %[[TRANSPOSE_1:.+]] = stablehlo.transpose %arg0, dims = [0, 2, 3, 1] : (tensor<1x8x4x4xf32>) -> tensor<1x4x4x8xf32>

  %1 = "tf.Identity"(%0) {device = ""} : (tensor<1x3x4x4xf32>) -> tensor<*xf32>
  %2 = "tf.Identity"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>
  return %2 : tensor<*xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_0(%arg0: tensor<1x3x6x6xf32>) -> tensor<1x3x4x4xf32> {
  %0 = mhlo.constant dense<1.000000e+00> : tensor<f32>

// fallback lowering is preferred for static shaped operands when available.

// CHECK-LABEL: batchmatmulv2
func.func @batchmatmulv2(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
  // NO_FALLBACK: mhlo.dynamic_broadcast_in_dim
  // NO_FALLBACK: mhlo.dot_general

  // SUPPORTED_FALLBACK_DEVICE: mhlo.reduce
  // SUPPORTED_FALLBACK_DEVICE: mhlo.dot_general

// CHECK-LABEL: strips_attributes
// CHECK-NOT: tf
func.func @strips_attributes(%arg0: tensor<32x28x28x1xf32> {tf._user_specified_name = "x"},
                             %arg1: tensor<3x3x1x5xf32> {tf._user_specified_name = "w1"},
                             %arg2: tensor<5xf32> {tf._user_specified_name = "b1"},
                             %arg3: tensor<3920x10xf32> {tf._user_specified_name = "w2"},

// CHECK: %[[CONV:.+]] = stablehlo.convolution(%[[TRANSPOSE_0]], %[[TRANSPOSE_1]]) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = {{\[\[}}1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x4x4x8xf32>, tensor<3x3x8x8xf32>) -> tensor<1x4x4x8xf32>

    %2 = "quantfork.stats"(%1) {layerStats = dense<[4.00000000e-6, 9.80000000e-1]> : tensor<2xf32>} : (tensor<2x3x2x2xf32>) -> tensor<2x3x2x2xf32>
    return %2 : tensor<2x3x2x2xf32>
  }
// Checks that the quantized XlaCallModule has been replaced by a CallOp, which
// calls the quantized entry function.

            index_vector_dim = 2>,
          slice_sizes = array<i64: 1, 2, 2>,
          indices_are_sorted = false
        } : (tensor<3x4x2xf32>, tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32>
        func.return %2 : tensor<2x3x2x2xf32>
      }
    }
  )mlir";
  OwningOpRef<ModuleOp> module_op = ParseModuleOpString(kModuleGather);
  ASSERT_TRUE(module_op);

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

  %w = arith.constant dense<127.0> : tensor<3x3x3x3xf32>
  %b = arith.constant dense<0.0> : tensor<3xf32>
  %conv = "tfl.conv_2d"(%arg0, %w, %b) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x5x5x3xf32>, tensor<3x3x3x3xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  func.return %conv : tensor<1x5x5x3xf32>