}

  SmallVector<int64_t> shape;
  bool refined_shape = false;
  // Build the shape of the refined type, if lhs is unranked it
  // will be directly the shape of the refined type, otherwise we merged by
  // taking the most specialized. This combines `10x?x?` and `?x?x8` into
  // `10x?x8`.
  if (!lhs_shape_type.hasRank()) {
    if (rhs_shape_type.hasRank()) {
      shape.append(rhs_shape_type.getShape().begin(),

    merged[indices[m][i], ...] = data[m][i, ...]
```

Each `data[i].shape` must start with the corresponding `indices[i].shape`,
and the rest of `data[i].shape` must be constant w.r.t. `i`.  That is, we
must have `data[i].shape = indices[i].shape + constant`.  In terms of this
`constant`, the output shape is

    merged.shape = [max(indices) + 1] + constant

      }
      SmallVector<int64_t, 4> result_shape = {leading_dim_v};
      ArrayRef<int64_t> shape = element_type.getShape();
      result_shape.append(shape.begin(), shape.end());
      result_type =
          tensorflow::GetTypeFromTFTensorShape(result_shape, element_dtype);
    }

    // Create a 1-D RankedTensorType for result's shape. Number of elements in
    // it is equal to the rank of the result, if known. Otherwise, the number of

#include "mlir/Dialect/Func/IR/FuncOps.h"  // from @llvm-project
#include "mlir/Dialect/Quant/QuantOps.h"  // from @llvm-project
#include "mlir/Dialect/SCF/IR/SCF.h"  // from @llvm-project
#include "mlir/Dialect/Shape/IR/Shape.h"  // from @llvm-project
#include "mlir/Dialect/Tensor/IR/Tensor.h"  // from @llvm-project
#include "mlir/InitAllPasses.h"  // from @llvm-project
#include "mlir/Support/LogicalResult.h"  // from @llvm-project

  let summary =
    "Replicate TensorList init ops for correct shape assignments in shape inference";

  let description = [{
    If we pass same TensorList to a while op as multiple arguments or just use
    the same TensorList at multiple places and assign different
    TensorListSetItem to elements of TensorList, the shape inference is then
    unable to identify the Shape of these args and thus the input TensorList

    if (ShapedType::isDynamic(num_index_dims)) return failure();

    auto updates = op.getUpdates();

    // Broadcast scalar `updates` in into expected shape as following shape:
    // updates.shape == indices.shape[:-1] + tensor.shape[indices.shape[-1]:]
    if (updates_ty.getRank() == 0 &&
        (std::is_same<OpTy, TF::TensorScatterUpdateOp>::value ||
         std::is_same<OpTy, TF::TensorScatterAddOp>::value)) {

  %cst = "tf.Const"() <{value = dense<1> : tensor<i32>}> {device = ""} : () -> tensor<i32>
  %0 = "tf.MultiDeviceIteratorFromStringHandle"(%arg0) <{output_shapes = [#tf_type.shape<>], output_types = [!tf_type.string]}> {device = ""} : (tensor<!tf_type.string>) -> tensor<!tf_type.resource>

  zero->setAttr(kICIWeightDistributionMlirBridgeMarker,
                builder.getBoolAttr(true));
  ConstOp shape = builder.create<ConstOp>(loc, shape_attr);
  shape->setAttr(kICIWeightDistributionMlirBridgeMarker,
                 builder.getBoolAttr(true));
  FillOp fill = builder.create<FillOp>(loc, shape, zero);
  fill->setAttr(kICIWeightDistributionMlirBridgeMarker,
                builder.getBoolAttr(true));
  return fill;

  import_config.upgrade_legacy = true;
  // Disable shape inference during import as some TensorFlow op fails during
  // shape inference with dynamic shaped operands. This in turn causes the
  // import to fail. Shape inference during import is going to be removed and
  // the shape inference pass is run early in the pass pipeline, shape inference
  // during import is not necessary.

#include "tensorflow/compiler/mlir/tensorflow/utils/convert_type.h"
#include "tensorflow/compiler/mlir/tensorflow/utils/serialize_mlir_module_utils.h"
#include "xla/mlir_hlo/mhlo/IR/register.h"
#include "xla/shape.h"
#include "xla/translate/mhlo_to_hlo/type_to_shape.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h"
#include "tsl/platform/errors.h"