// 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> {

// this will shade it.
//
// 3. Once a goroutine's stack is black, the shade(ptr) becomes
// unnecessary. shade(ptr) prevents hiding an object by moving it from
// the stack to the heap, but this requires first having a pointer
// hidden on the stack. Immediately after a stack is scanned, it only
// points to shaded objects, so it's not hiding anything, and the

		}
	} else {
		// Note: can't use rawmem (which avoids zeroing of memory), because then GC can scan uninitialized memory.
		to = mallocgc(tomem, et, true)
		if copymem > 0 && writeBarrier.enabled {
			// Only shade the pointers in old.array since we know the destination slice to
			// only contains nil pointers because it has been cleared during alloc.
			//
			// It's safe to pass a type to this function as an optimization because

		runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
		if tail == atomic.Load(&pp.runqtail) {
			return head == tail && runnext == 0
		}
	}
}

// To shake out latent assumptions about scheduling order,
// we introduce some randomness into scheduling decisions
// when running with the race detector.
// The need for this was made obvious by changing the

  if (pack_axis < 0) {
    pack_axis += rank;
  }

  // Concat out shape.
  for (int i = 0; i < rank; ++i) {
    int64_t dim_size = input_type.getDimSize(i);
    if (i == pack_axis) {
      dim_size *= count;
    }
    concat_out_shape.push_back(dim_size);
  }

  // Pack out shape.
  int j = 0;
  for (int i = 0; i < rank + 1; ++i) {
    if (i == pack_axis) {

  return transposed_type;
}

// Returns shape of a ranked tensor.
// Precondition: output_val's is ranked tensor.
// Returns a truncated shape when `truncate` is set to true.
inline DenseElementsAttr GetShape(Value output_val, bool truncate = false) {
  auto output_shape = output_val.getType().dyn_cast<ShapedType>().getShape();

  SmallVector<int32_t> shape;
  shape.reserve(output_shape.size());

// 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;

    {
      "tensors": [
        {
          "shape": [1, 5, 2],
          "name": "input0"
        },
        {
          "shape": [2, 5],
          "buffer": 1,
          "name": "input2input_weights1"
        },
        {
          "shape": [2, 5],
          "buffer": 2,
          "name": "input2forget_weights2"
        },
        {
          "shape": [2, 5],
          "buffer": 3,

    // 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.

//  se::DeviceMemoryBase gpu_dst{device_tensor.data(), 4 * sizeof(float)};
//  xla::Shape shape(xla::F32, {2, 2}, {}, {})
//  tsl::AsyncValueRef<std::unique_ptr<se::Event>> done_event =
//      tsl::MakeConstructedAsyncValueRef<std::unique_ptr<se::Event>>(stream.parent());
//  done_event->Init();
//
//  XlaHostSendDeviceContext device_context(&stream, &gpu_dst,
//    shape, done_event);
//  device_context.CopyCPUTensorToDeviceSync(