func.func @testBitcast_v5(%arg0: tensor<3x2xi16>) -> tensor<2xi32> {
  // expected-error @+1 {{the 0th dim of output tensor is 2. It is not equal to the one in input tensor, which is 3}}
  %0 = "tf.Bitcast"(%arg0) : (tensor<3x2xi16>) -> tensor<2xi32>
  func.return %0 : tensor<2xi32>
}

// -----

// CHECK-LABEL: func @testBitcast_v6
func.func @testBitcast_v6(%arg0: tensor<3x2xi32>) -> tensor<3x2x2xi16> {

  // Build the graph:
  // "add" = "arg0" + "arg1"
  // "ret0" = "add"
  // "ret1" = "arg1"
  tensorflow::Scope s = tensorflow::Scope::NewRootScope();
  Output arg0 = ops::_Arg(s.WithOpName("arg0"), DT_INT32, 0);
  Output arg1 = ops::_Arg(s.WithOpName("arg1"), DT_FLOAT, 1);
  Output arg2 = ops::_Arg(s.WithOpName("arg2"), DT_INT32, 2);
  Output add = ops::Add(s.WithOpName("add"), arg0, arg0);

  auto arg0 = ops::_Arg(scope.WithOpName("a_0_arg"), DT_INT32, 0);
  auto arg1 = ops::_Arg(scope.WithOpName("b_0_arg"), DT_FLOAT, 1);
  auto arg2 = ops::_Arg(scope.WithOpName("c_0_arg"), DT_INT32, 2);
  auto arg3 = ops::_Arg(scope.WithOpName("d_0_arg"), DT_FLOAT, 3);

  auto arg4 = ops::_Arg(scope.WithOpName("u_0_arg"), DT_RESOURCE, 4);
  auto arg5 = ops::_Arg(scope.WithOpName("v_0_arg"), DT_RESOURCE, 5);

		// Atomic compare and swap.
		// arg0 = pointer, arg1 = old value, arg2 = new value, arg3 = memory.
		// if *arg0 == arg1 {
		//   *arg0 = arg2
		//   return (true, memory)
		// } else {
		//   return (false, memory)
		// }
		// MOV  $0, Rout
		// LR	(Rarg0), Rtmp
		// BNE	Rtmp, Rarg1, 3(PC)
		// SC	Rarg2, (Rarg0), Rtmp
		// BNE	Rtmp, ZERO, -3(PC)
		// MOV  $1, Rout

	// can be executed on g0. Second, it is called frequently, so will
	// benefit from this fast path.
	CMPQ	RARG0, $0
	JNE	rest
	get_tls(RARG0)
	MOVQ	g(RARG0), RARG0
	MOVQ	g_m(RARG0), RARG0
	MOVQ	m_p(RARG0), RARG0
	MOVQ	p_raceprocctx(RARG0), RARG0
	MOVQ	RARG0, (RARG1)
	RET

rest:
	// Transition from C ABI to Go ABI.
	PUSH_REGS_HOST_TO_ABI0()
	// Set g = g0.
	get_tls(R12)

}

TEST(IsSupportedByReplicatedBridge, ReplicatedModule) {
  const char* const code = R"mlir(
func.func @entry_func_1(%arg0: tensor<i32>) -> tensor<i32> attributes {tf.entry_function = {}} {
  %0 = "tf.Identity"(%arg0) {_tpu_replicate = "cluster"} : (tensor<i32>) -> (tensor<i32>)
  func.return %0 : tensor<i32>
}
)mlir";
  mlir::MLIRContext context;

//   id0 = arg0
//   if-then-branch: [u0,   arg0, arg0]
//   if-else-branch: [arg0, arg0, arg1]
//     => first result is unknown, second and third is passthrough
//   if results    : [*,    arg0, {arg0, arg1}[
//   ID #2: read (unknown)         -> succ {5, 6)
//   ID #3: read (arg0)            -> succ {5}
//   ID #4: read({arg0,arg1})      -> succ {5,6}
//   ID #5: write(arg0)

^bb0(%arg0: tensor<? x f32>):
  // CHECK: "tfl.log"(%arg0)
  %0 = "tfl.log"(%arg0): (tensor<? x f32>) -> tensor<? x f32>
  func.return %0 : tensor<? x f32>
}

// CHECK-LABEL: testLogStaticShapeInputAndDynamicShapeOutput
func.func @testLogStaticShapeInputAndDynamicShapeOutput(tensor<8 x f32>) -> tensor<? x f32> {
^bb0(%arg0: tensor<8 x f32>):
  // CHECK: "tfl.log"(%arg0)

CHECK:  %[[LSTM:[a-z0-9]*]] = "tfl.unidirectional_sequence_lstm"(%arg2, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg1, %arg1, %arg1, %arg1, %arg1, %arg1, %arg1, %arg0, %arg1, %[[CST_0]], %[[CST_1]], %arg0, %arg0, %arg0, %arg0) <{fused_activation_function = "NONE", time_major = true}> : (tensor<4x4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>,...

func.func @test_batch_matmul_to_einsum(%arg0: tensor<1x2x3xf32>, %arg1: tensor<3x4xf32>) -> tensor<1x2x4xf32> {
  // CHECK-LABEL: test_batch_matmul_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<1x2x3xf32>, tensor<3x4xf32>) -> tensor<1x2x4xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<1x2x3xf32>, tensor<3x4xf32>) -> tensor<1x2x4xf32>