// Merge node causes a loop so we remove NextIteration->Merge edge before
    // performing shape inference. But removing those edges also prevents us
    // from inferring output shape for Merge node (we need shapes for all its
    // inputs).
    // For loop invariant resource input's Merge node, we set output resource
    // shape as Enter node's resource shape.
    // TODO(b/129367850): clean this up.

# RUN: tf-mlir-translate -graphdef-to-mlir -tf-enable-shape-inference-on-import=false %s -o - | FileCheck %s

# Verify that the _input_shapes attribute of the FunctionDef is respected.
# This also checks that the output type is correctly inferred based on
# that.
#CHECK: func private @identity_function0(%arg0: tensor<i32>) -> tensor<i32>

node {
  name: "Placeholder"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_BOOL

Keith Randall <******@****.***> 1627678827 -0700

// Verifies that constants generated post shape inference are propagated.
// get_shape result in this test.
module attributes {tf.versions = {producer = 179 : i32}} {

  // CHECK-LABEL: func @main
  func.func @main(%arg0: tensor<10x19xf32>, %arg1: tensor<19x10xf32> {mhlo.is_same_data_across_replicas = true}) -> tensor<?xi64> {
    %0 = "tf.Shape"(%arg0) : (tensor<10x19xf32>) -> tensor<2xi64>

  xs.push_back(Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape)));
  auto y = AddN(scope_, xs);
  RunTest(xs, {shape, shape, shape}, {y}, {shape});
}

TEST_F(NaryGradTest, Add) {
  TensorShape x1_shape({3, 2, 5});
  TensorShape x2_shape({2, 5});
  auto x1 = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x1_shape));
  auto x2 = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x2_shape));

  ParallelTensor(const ParallelDevice& device,
                 std::vector<TensorHandlePtr> tensors,
                 absl::Span<const int64_t> shape, const TF_DataType dtype)
      : device_(device),
        tensors_(std::move(tensors)),
        shape_(std::vector<int64_t>(shape.begin(), shape.end())),
        dtype_(dtype) {}
  ParallelTensor(const ParallelDevice& device,

// RUN: tf-opt -split-input-file -verify-diagnostics --tf-shape-inference %s | FileCheck %s

module attributes {tf.versions = {producer = 888 : i32}} {
// CHECK-LABEL: testConv2dShapeValidPadding
func.func @testConv2dShapeValidPadding(%arg0: tensor<1x112x80x128xf32>, %arg1: tensor<128x3x3x128xf32>, %arg2: tensor<128xf32>) -> tensor<1x?x?x128xf32> {

    llvm::ArrayRef<int64_t> shapes) {
  return llvm::to_vector(llvm::map_range(shapes, [](int64_t shape) {
    return shape == kTFDynamicSize ? mlir::ShapedType::kDynamic : shape;
  }));
}

llvm::SmallVector<int64_t> ConvertMlirShapeToTF(
    llvm::ArrayRef<int64_t> shapes) {
  return llvm::to_vector(llvm::map_range(shapes, [](int64_t shape) {
    return mlir::ShapedType::isDynamic(shape) ? kTFDynamicSize : shape;
  }));
}

// Extracts shape from XlaArgument as TensorShape. If shape is a xla::Shape,
// that is converted to a TensorShape.
absl::StatusOr<TensorShape> GetTensorShapeFromXlaArgument(
    const XlaArgument& arg) {
  if (absl::holds_alternative<xla::Shape>(arg.shape)) {
    TensorShape arg_shape;
    TF_RETURN_IF_ERROR(
        XLAShapeToTensorShape(std::get<xla::Shape>(arg.shape), &arg_shape));
    return arg_shape;

^bb0(%arg0: tensor<4x4x4xf32>) :
  %shape0 = arith.constant dense<[16, 4]> : tensor<2xi32>
  %shape1 = arith.constant dense<[64]> : tensor<1xi32>
  %0 = "tfl.reshape"(%arg0, %shape0) : (tensor<4x4x4xf32>, tensor<2xi32>) -> tensor<16x4xf32>
  %1 = "tfl.reshape"(%0, %shape1) : (tensor<16x4xf32>, tensor<1xi32>) -> tensor<64xf32>
  %2 = "tfl.reshape"(%0, %shape1) : (tensor<16x4xf32>, tensor<1xi32>) -> tensor<64xf32>