#include "xla/translate/mhlo_to_hlo/type_to_shape.h"

namespace mlir {

static FailureOr<std::vector<int64_t>> GetTPUInfeedLayoutFromAPI(
    RankedTensorType t) {
  // Call the TPU API to determine the right infeed layout. Note that
  // this can fail if we're not running on a TPU-enabled node.
  // TODO(kramm): Move this into a separate pass. See b/184944903
  xla::Shape old_shape = xla::TypeToShape(t);
  XLA_Shape old_shape_c = {};

  }
};

// Converts InfeedDequeueTuple to XLA HLO create_token, infeed and
// get_tuple_element ops.
//
// All HLO infeed ops expect a HLO token type operand and produce a tuple
// containing a token. This HLO token type is used to order multiple infeed
// operations within a computation. The token type can come from other
// infeed/outfeed/send/recv ops or can be generated using create_token op with

    TF_Tensor:$output
  );

  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
}

def TF_InfeedDequeueTupleOp : TF_Op<"InfeedDequeueTuple", []> {
  let summary = "Fetches multiple values from infeed as an XLA tuple.";

  let arguments = (ins
    OptionalAttr<StrAttr>:$_XlaSharding,
    OptionalAttr<ArrayAttr>:$layouts
  );

  let results = (outs
    Variadic<TF_Tensor>:$outputs
  );

// CHECK: [[TOKEN:%.*]] = mhlo.create_token  : !mhlo.token
// CHECK: [[INFEED:%.*]]:3 = "mhlo.infeed"([[TOKEN]]) <{infeed_config = ""{{.*}}}> : (!mhlo.token) -> (tensor<1x8x4x4xi32>, tensor<1x100x1xf32>, !mhlo.token)
// CHECK: return [[INFEED]]#0, [[INFEED]]#1
  %0:2 = "tf.InfeedDequeueTuple"() : () -> (tensor<1x8x4x4xi32>, tensor<1x100x1xf32>)

g,p,v,d=t(i),m=u.invert(r[0],r[1]),y={point:o,lineStart:c,lineEnd:s,polygonStart:function(){y.point=l,y.lineStart=f,y.lineEnd=h,g=[],p=[],i.polygonStart()},polygonEnd:function(){y.point=o,y.lineStart=c,y.lineEnd=s,g=Xo.merge(g);var n=Le(m,p);g.length?we(g,Ne,n,e,i):n&&(i.lineStart(),e(null,null,1,i),i.lineEnd()),i.polygonEnd(),g=p=null},sphere:function(){i.polygonStart(),i.lineStart(),e(null,null,1,i),i.lineEnd(),i.polygonEnd()}},x=Ce(),M=t(x);return y}}function Ae(n){return n.length>1}function Ce(){var n,t=[];...

  }];

  let arguments = (ins
    TF_ShapeAttr:$shape
  );

  let results = (outs
    Res<TF_Tensor, [{A tensor that will be provided using the infeed mechanism.}]>:$output
  );

  TF_DerivedResultTypeAttr dtype = TF_DerivedResultTypeAttr<0>;
}

def TF_InitializeTableOp : TF_Op<"InitializeTable", []> {
  let summary = [{