}
  // CHECK: func.func private @f_batch_callee(%[[ARG_0:.*]]: tensor<?xf32>) -> (tensor<?xf32>, tensor<0xf32>)
  // CHECK-DAG: %[[CST_0:.*]] = "tf.Const"() <{value = dense<> : tensor<0xf32>}> : () -> tensor<0xf32>
  // CHECK: return %[[ARG_0]], %[[CST_0]] : tensor<?xf32>, tensor<0xf32>

  func.func @f(%arg: tensor<1xf32>) -> (tensor<*xf32>, tensor<*xf32>) {

  func.return %2, %3: tensor<i32>, tensor<i32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() <{value = dense<2> : tensor<i32>}> : () -> tensor<i32>
// CHECK-DAG: %[[CST1:.*]] = "tf.Const"() <{value = dense<1> : tensor<i32>}> : () -> tensor<i32>
// CHECK-NEXT: return %[[CST]], %[[CST1]]
}

// Tests ops that variable shapes are correctly evaluated on static types.

constexpr absl::string_view kModuleConstantAdd = R"mlir(
  module {
    func.func @constant_add() -> (tensor<3x2xf32>) {
      %cst1 = stablehlo.constant dense<2.4> : tensor<3x2xf32>
      %cst2 = stablehlo.constant dense<5.7> : tensor<3x2xf32>
      %add = stablehlo.add %cst1, %cst2 : (tensor<3x2xf32>, tensor<3x2xf32>) -> tensor<3x2xf32>
      func.return %add : tensor<3x2xf32>
    }
  }
)mlir";

  return %1: tensor<1x3xf32>
}
// CHECK-DAG: %[[CST:.*]] = "tf.Const"
// CHECK-SAME: dense<0>
// CHECK-DAG: %[[CST_0:.*]] = "tf.Const"
// CHECK-SAME: dense<1>
// Check that there are no extra "tf.Const"s existing in this function.
// CHECK-NOT: "tf.Const"

// Check that the usages of %[[CST]] and %[[CST_0]] are untouched.
// CHECK: %[[ADD:.*]] = "tf.AddV2"(%[[CST]], %[[CST_0]])
// CHECK: "tf.Max"({{.*}}, %[[ADD]])

// -----

  "tf_device.launch"() ({
    %cst_0 = "tf.Const"() {value = dense<""> : tensor<1x!tf_type.string>} : () -> tensor<1x!tf_type.string>
    "tf.OpA"(%cst_0) { mini_batch_splits = ""} : (tensor<1x!tf_type.string>) -> ()
    "tf.OpB"(%cst_0) { mini_batch_in_csr = ""} : (tensor<1x!tf_type.string>) -> ()
    tf_device.return

// RUN: tf-opt %s --tf-hoist-loop-invariant | FileCheck %s --dump-input=fail

// CHECK-LABEL: hoist_loop_invariant
// CHECK:       [[CST_0:%.*]] = "tf.Const"
// CHECK-NEXT:  [[CST_1:%.*]] = "tf.Const"
// CHECK:       [[RES_1:%.*]] = "tf.Add"([[CST_1]], [[CST_0]])
// CHECK:       [[RES_2:%.*]] = "tf.Mul"([[RES_1]], [[CST_1]])
// CHECK:       tf.WhileRegion
// CHECK:       ^bb0
// CHECK:       tf.OpA
// CHECK:       ^bb0([[ARG_2:%[a-zA-Z0-9_]+]]

func.func @decompose_batch_norm(%arg0: tensor<*xf32>) -> (tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_0 = "tf.Const"() {value = dense<0.500000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

module {
  func.func @restore_single() {
    %cst = "tf.Const"() <{value = dense<"restore_ariables"> : tensor<!tf_type.string>}> : () -> tensor<!tf_type.string>
    %cst_0 = "tf.Const"() <{value = dense<""> : tensor<1x!tf_type.string>}> : () -> tensor<1x!tf_type.string>
    %cst_1 = "tf.Const"() <{value = dense<"y"> : tensor<1x!tf_type.string>}> : () -> tensor<1x!tf_type.string>

  %cst = arith.constant dense<1.0> : tensor<5x2x3x4xf32>
  %1 = tfl.add %0, %cst { fused_activation_function = "NONE" } : tensor<5x2x3x4xf32>
  func.return %1 : tensor<5x2x3x4xf32>
}

// CHECK: %cst = arith.constant dense<[3, 0, 1, 2]> : tensor<4xi32>
// CHECK: %cst_0 = arith.constant dense<1.000000e+00> : tensor<2x3x4x5xf32>

}

// -----

module {
  func.func @not_propagate_last_op(%arg0: tensor<10x2xi32>) -> tensor<1x300x10xf32> {
    %cst = "tf.Const"() {value = dense<[1, 1, 300]> : tensor<3xi64>} : () -> tensor<3xi64>
    %cst_0 = "tf.Const"() {value = dense<127> : tensor<200x100x300xi8>} : () -> tensor<200x100x300xi8>
    %0 = "tf.Identity"(%cst_0) : (tensor<200x100x300xi8>) -> tensor<200x100x300xi8>