// CHECK: %[[RESHAPE:.*]] = stablehlo.reshape %[[CALL]]

    // CHECK: %[[REDUCE:.*]] = "stablehlo.reduce_window"(%[[RESHAPE]], %[[Q0]])
    // CHECK{LITERAL}: padding = dense<[[0, 0], [1, 1], [0, 0]]> : tensor<3x2xi64>
    // CHECK-SAME: window_dimensions = array<i64: 1, 3, 1>
    // CHECK: %[[ARG1:.*]]: tensor<!quant.uniform<i8:f32, 3.000000e-01:1>>, %[[ARG2:.*]]: tensor<!quant.uniform<i8:f32, 3.000000e-01:1>>

// HISTOGRAM-PERCENTILE-CHECK-NEXT:  [[lhs:%.*]], {{.*}}, {{.*}}, {{.*}} = "tf.CustomAggregator"(%arg0) <{calibration_method = 3 : i32, id = "composite_conv2d_with_relu6_fn_arg_0_calibration_method_3", max_percentile = 9.999900e+01 : f32, min_percentile = 1.000000e-03 : f32, num_bins = 512 : i32}> : (tensor<*xf32>) -> (tensor<*xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)

  %8 = "mhlo.broadcast_in_dim"(%7) <{broadcast_dimensions = dense<0> : tensor<1xi64>}> : (tensor<32xi64>) -> tensor<32x32xi64>
  %9 = mhlo.compare  LT, %8, %6 : (tensor<32x32xi64>, tensor<32x32xi64>) -> tensor<32x32xi1>
  %10 = mhlo.add %8, %4 : tensor<32x32xi64>
  %11 = mhlo.select %9, %10, %8 : tensor<32x32xi1>, tensor<32x32xi64>
  %12 = mhlo.reshape %11 : (tensor<32x32xi64>) -> tensor<32x32x1xi64>

// CHECK: %[[CUSTOM_AGGREGATOR_1:.*]], %[[MIN_1:.*]], %[[MAX_1:.*]], %[[HISTOGRAM_1:.*]] = "tf.CustomAggregator"
// CKECK-SAME: <{calibration_method = 5 : i32, id = "1", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x3x4x3xf32>) -> (tensor<1x3x4x3xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)

  %cst_0 = arith.constant dense<[2, -1]> : tensor<2xi64>
  %0 = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2) {num_bits = 5, narrow_range = false} : (tensor<1x2xf32>, tensor<f32>, tensor<f32>) -> tensor<1x2xf32>
  %1 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

  %cst_0 = arith.constant dense<[2, -1]> : tensor<2xi64>
  %0 = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2) {num_bits = 3, narrow_range = false} : (tensor<1x2xf32>, tensor<f32>, tensor<f32>) -> tensor<1x2xf32>
  %1 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

  // CHECK-SAME: %[[ARG0:.*]]: tensor<3x4x5xf32>
  // CHECK-SAME: %[[ARG1:.*]]: tensor<3x5x2xf32>
  // CHECK-SAME: %[[ARG2:.*]]: tensor<2x3x2xi64>
  func.func private @composite_and_gather(%arg0: tensor<3x4x5xf32>, %arg1: tensor<3x5x2xf32>, %arg2: tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32> {
    // CHECK: %[[Q1:.*]] = "quantfork.qcast"(%[[ARG0]]) {volatile} : (tensor<3x4x5xf32>) -> tensor<3x4x5x!quant.uniform<i8:f32, 5.000000e-03>>

// -----

// Test that q/dq pair is not inserted between constant and XlaCallModule op
// whose entry function name does not include conv nor dot_general.

func.func @no_qdq_except_conv_and_dot_general(%arg0: tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32> attributes {tf._original_func_name = "main_0"} {
  %cst = "tf.Const"() {value = dense<3.000000e-01> : tensor<3x4x2xf32>} : () -> tensor<3x4x2xf32>
  %0 = "tf.XlaCallModule"(%cst, %arg0) {

  func.func @main() -> () {
    %cst0 = "tf.Const"(){ value = dense<0> : tensor<3x5xi1>} : () -> tensor<3x5xi1>
    %0 = "tf.Where"(%cst0) : (tensor<3x5xi1>) -> tensor<?x2xi64>
    func.return
  }
})";

absl::StatusOr<XlaCompiler::CompilationResult> CompileMlirModule(
    const char* mlir_module_str,
    ConfigProto::Experimental::MlirBridgeRollout rollout_state) {

      stablehlo.return %3 : tensor<f32>
  }) {
    window_dimensions = array<i64: 1, 1, 2, 2>,
    window_strides = array<i64: 1, 1, 2, 2>,
    padding = dense<[[0, 0], [0, 0], [1, 1], [1, 1]]> : tensor<4x2xi64>
  } : (tensor<1x4x16x16xf32>, tensor<f32>) -> tensor<1x4x9x9xf32>
  return %2 : tensor<1x4x9x9xf32>
}
// CHECK-SAME: %[[ARG:.+]]: tensor<1x16x16x4xf32>
// CHECK-DAG: stablehlo.constant