func.return %1 : tensor<16x128xf32>
  // CHECK: return %0 : tensor<16x128xf32>
}

// CHECK-LABEL: FuseTransposeFCLhsToBatchMatmul
func.func @FuseTransposeFCLhsToBatchMatmul(%arg0: tensor<1024x4xf32>, %arg1: tensor<8x1024xf32>, %arg2: tensor<4x256xf32>) -> tensor<8x256xf32> {

  %1 = "tfl.arg_max"(%0, %cst) : (tensor<16x1024xf32>, tensor<1xi32>) -> tensor<16xi32>
  func.return %1 : tensor<16xi32>
  // CHECK-DAG: %[[CST:.*]] = arith.constant dense<-1> : tensor<1xi32>
  // CHECK: %[[SOFTMAX:.*]] = "tfl.softmax"(%arg0) <{beta = -1.000000e+00 : f32}> : (tensor<16x1024xf32>) -> tensor<16x1024xf32>

    %1 = "tf.Identity"(%0#3) {device = ""} : (tensor<1x1024xf32>) -> tensor<1x1024xf32>
    func.return %1 : tensor<1x1024xf32>
  }

  func.func @main(%arg0: tensor<128x10xf32>, %arg1: tensor<10x1024xf32>, %arg2: tensor<128x1024xf32>) -> (tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, tensor<128x1024xf32> {mhlo.sharding = ""}) {
    func.return %arg0, %arg1, %arg2 : tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>
  }
}

// The following xla::OpSharding protos are used:

    %cst_0 = "tf.Const"() <{value = dense<0> : tensor<i32>}> : () -> tensor<i32>
    %12:4 = "tf.Split"(%cst_0, %4) {num_split = 4 : i32} : (tensor<i32>, tensor<128x1024xf32>) -> (tensor<32x1024xf32>, tensor<32x1024xf32>, tensor<32x1024xf32>, tensor<32x1024xf32>)

  // CHECK: "tf.Identity"(%[[SHARDED_ARG1]])
  %0 = "tf.Identity"(%arg1) : (tensor<10x1024xf32>) -> tensor<10x1024xf32>

  // CHECK: "tf.Identity"(%arg2)
  %1 = "tf.Identity"(%arg2) : (tensor<128x1024xf32>) -> tensor<128x1024xf32>
  func.return %arg0, %0, %1 : tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>
}

// CHECK-LABEL: @RewriteHostComputeMlirOp

    %0 = "tf.Div"(%arg0, %cst_3) {device = ""} : (tensor<1x1024xf32>, tensor<f32>) -> tensor<1x1024xf32>
    %1 = "tf.AddV2"(%0, %cst_0) {device = ""} : (tensor<1x1024xf32>, tensor<f32>) -> tensor<1x1024xf32>
    %2 = "tf.Floor"(%1) {device = ""} : (tensor<1x1024xf32>) -> tensor<1x1024xf32>
    %3 = "tf.ClipByValue"(%2, %cst_1, %cst_5) {device = ""} : (tensor<1x1024xf32>, tensor<f32>, tensor<f32>) -> tensor<1x1024xf32>

module attributes {tf.versions = {producer = 179 : i32}} {
  func.func @main(%arg0: tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, %arg1: tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, %arg2: tensor<128x1024xf32> {mhlo.sharding = ""}) {
    func.return
  }
}

// The following xla::OpSharding protos are used:
//  Serialized string:
//   "\08\03\1A\02\01\02\22\02\00\01"

constexpr absl::string_view kModuleStatic = R"mlir(
  module {
    func.func @main(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

// -----

// Tests composite tf.XlaCallModule is converted to func.call.

module {
  // CHECK-LABEL: func.func @main
  func.func @main(%arg0: tensor<1x1024xf32>) -> tensor<1x3xf32> {
    // CHECK: call @composite_dot_general_fn_1
    // CHECK-SAME: (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
    // CHECK-NOT: tf.XlaCallModule
    %0 = "tf.Const"() <{value = dense<0.5> : tensor<1024x3xf32>}> : () -> tensor<1024x3xf32>