// CHECK-DAG: %[[SCALE:.*]] = "tf.Const"() <{value = dense<0.0787401571> : tensor<f32>
}

// -----

module {
  func.func @quantize_xlagather(%arg0: tensor<10x2xi32>) -> tensor<1x300x10xf32> {
    %cst_0 = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<200x100x300xf32>} : () -> tensor<200x100x300xf32>

    %0 = "tf.BatchMatMulV2"(%arg3, %arg1) {adj_x = false, adj_y = false} : (tensor<1x3xf32>, tensor<3x4xf32>) -> tensor<1x4xf32>
    %1 = arith.constant dense<[[2.3, 3.4, 4.5, 5.5]]> : tensor<1x4xf32>
    %2 = "tf.Add"(%0, %1) : (tensor<1x4xf32>, tensor<1x4xf32>) -> tensor<1x4xf32>
    %3 = tensor.cast %2 : tensor<1x4xf32> to tensor<1x?xf32>
    func.return %3 : tensor<1x?xf32>
}

// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<-131072> : tensor<1x3xi32>}> : () -> tensor<1x3xi32>
// CHECK: %[[MATMUL:.*]] = "tf.XlaDotV2"({{.*}}, %[[WEIGHT]])
// CHECK-SAME: (tensor<1x1024xi8>, tensor<1024x3xi8>) -> tensor<1x3xi32>
// CHECK: %[[SUB:.*]] = "tf.Sub"(%[[MATMUL]], %[[CONST]]) : (tensor<1x3xi32>, tensor<1x3xi32>) -> tensor<1x3xi32>
}

// -----

module attributes {} {

  %0 = "tf.Const"() {value = dense<0.0> : tensor<1x2xf32>} : () -> tensor<1x2xf32>

  // Operand and identity shapes are broadcastable. However, we cannot fold
  // because the operand does not match the result shape.
  %1 = "tf.AddV2"(%arg0, %0) : (tensor<4x1xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  %2 = "tf.AddV2"(%0, %arg0) : (tensor<1x2xf32>, tensor<4x1xf32>) -> tensor<4x2xf32>

      tf_device.return %5 : tensor<512xi32>
    }) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : () -> (tensor<512xi32>, tensor<1024xi32>, tensor<1024xi32>, tensor<1024xf32>)
    // CHECK: %[[COMPILE_OUTPUT:[0-9]*]]:4 = "tf_device.launch"
    // CHECK: "tf._TPUCompileMlir"()
    // CHECK: is_bounded_dynamic_dim: true