// CHECK-LABEL: import_stats_skip
func.func @import_stats_skip(%arg0: tensor<4xf32>, %cst: tensor<i32>) -> (tensor<2xf32>,tensor<2xf32>) {
  %0:2 = "tfl.split"(%cst, %arg0) {num_splits = 2 : i32} : (tensor<i32>, tensor<4xf32>) -> (tensor<2xf32>, tensor<2xf32>)
    loc(fused["skip1", "skip2.cc":10:8, callsite("op" at "skip3.cc":10:8)])
  func.return %0#0, %0#1 : tensor<2xf32>, tensor<2xf32>

       it != dimension_to_splits_map->rend(); ++it) {
    int concat_dimension = it->first;
    int num_splits = it->second;

    llvm::SmallVector<mlir::Value, 4> new_outputs;
    new_outputs.reserve(num_splits);
    for (int i = 0, end = outputs_to_merge.size(); i < end;
         i = i + num_splits) {
      mlir::TF::ConcatOp concat_op =
          CreateConcatOp(concat_dimension, location,

            0,
            1
          ],
          "outputs": [
            2,
            3
          ],
          "builtin_options_type": "SplitOptions",
          "builtin_options": {
            "num_splits": 2
          }
        }
      ],
      "name": "main"
    }
  ],
  "description": "MLIR Converted."

  %0 = "tfl.quantize"(%arg0) {qtype = tensor<4x!quant.uniform<u8:f32, 1.0>>} : (tensor<4xf32>) -> tensor<4x!quant.uniform<u8:f32, 1.0>>
  %1:4 = "tfl.split"(%arg1, %0) {num_splits = 4 : i32} : (tensor<i32>, tensor<4x!quant.uniform<u8:f32, 1.0>>)
  -> (tensor<2x!quant.uniform<u8:f32, 1.0>>, tensor<2x!quant.uniform<u8:f32, 1.0>>,tensor<2x!quant.uniform<u8:f32, 1.0>>, tensor<2x!quant.uniform<u8:f32, 1.0>>)

// CHECK:       }

// -----

func.func @unrollSplit(%arg0: tensor<i32>, %arg1: tensor<1x8x8x1024xf32>) -> (tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>) {
  %0:4 = "tfl.split"(%arg0, %arg1) {num_splits = 4 : i32, tac.device = "CPU"} : (tensor<i32>, tensor<1x8x8x1024xf32>) -> (tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>)

  %4, %5 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x2xf32>)
  %6:2 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x?xf32>)
  %7:2 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<?x2xf32>, tensor<16x?xf32>)

# CHECK-DAG:       %[[VAL_3:.*]] = "tfl.no_value"() <{value}> : () -> none
# CHECK-DAG:       %[[VAL_6:.*]] = arith.constant dense<0> : tensor<i32>
# CHECK:           %[[VAL_7:.*]]:2 = "tfl.split"(%[[VAL_6]], %[[VAL_0]]) <{num_splits = 2 : i32}> : (tensor<i32>, tensor<2x5x3xf32>) -> (tensor<1x5x3xf32>, tensor<1x5x3xf32>)
# CHECK:           %[[VAL_9:.*]] = "tfl.transpose"(%[[VAL_1]], %[[VAL_2]]) : (tensor<3x7xf32>, tensor<2xi32>) -> tensor<7x3xf32>

                                   RankedTensorType input_type, int split_dim,
                                   int num_splits, PatternRewriter* rewriter) {
  SmallVector<Value, 4> slice_outputs;
  int begin = 0;
  for (int i = 0; i < num_splits; ++i) {
    // Create slice op.
    // Populate begin & size.
    SmallVector<int32_t, 4> slice_begin;
    SmallVector<int32_t, 4> slice_size;

  if (dim_size % num_splits != 0)
    return op.emitOpError("'num_splits' should evenly divide 'split_dim' axis");

  // Verifies output tensor types.
  RankedTensorType expected_output_type = SubstituteRankedTensorTypeDimSize(
      input_type, split_dim, dim_size / num_splits);
  return VerifySplitOpOutputTypes(
      op.getOperation(), num_splits,

  %0 = "tfl.dequantize"(%arg) : (tensor<4x!quant.uniform<u8:f32, 1.0>>) -> tensor<4xf32>
  %1:2 = "tfl.split"(%cst, %0) {num_splits = 2 : i32} : (tensor<i32>, tensor<4xf32>) -> (tensor<2xf32>, tensor<2xf32>)
  %2 = "tfl.quantize"(%1#0) {qtype = tensor<2x!quant.uniform<u8:f32, 1.0>>} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.0>>