if err != nil {
					c.Expires = time.Time{}
					break
				}
			}
			c.Expires = exptime.UTC()
			continue
		case "path":
			c.Path = val
			continue
		case "partitioned":
			c.Partitioned = true
			continue
		}
		c.Unparsed = append(c.Unparsed, parts[i])
	}
	return c, nil
}

// readSetCookies parses all "Set-Cookie" values from

using PartitionedClusterOutputMap = absl::flat_hash_map<
    uint64_t, llvm::SmallVector<mlir::TF::TPUPartitionedOutputV2Op, 8>>;

// Returns the partitioned output ops from the cluster if there are any,
// along with any single user identity ops between them. Not all outputs
// of a cluster must be partitioned, so the output is a map from cluster
// output ids to ops.
std::tuple<PartitionedClusterOutputMap,

		"cookie-15=samesite-none; SameSite=None",
	},
	{
		&Cookie{Name: "cookie-16", Value: "partitioned", SameSite: SameSiteNoneMode, Secure: true, Path: "/", Partitioned: true},
		"cookie-16=partitioned; Path=/; Secure; SameSite=None; Partitioned",
	},
	// The "special" cookies have values containing commas or spaces which
	// are disallowed by RFC 6265 but are common in the wild.
	{

  added_func_op.setName(builder.getStringAttr(function_name));
  added_call_op.setCallee(builder.getStringAttr(function_name));
}

// Raises partitioned sequential `Operations` from a block to a new function
// definition. `Operations` are partitioned into classes from the cartesian
// product of possible devices and inference datatypes. For example, we might

  // Add TF->TF lowering patterns.
  TF::PopulateTFLoweringBeforeHLOPatterns(context, &legalize_lower_patterns);

  if (tf2xla_fallback_device_type && prefer_tf2xla) {
    VLOG(1) << "TF to XLA legalization patterns are partitioned by op into "
               "either native MLIR legalization, or TF2XLA fallback "
               "legalzation, with a preference toward TF2XLA.";
  } else if (tf2xla_fallback_device_type) {

      FunctionType::get(func.getContext(), new_input_types,
                        func.front().getTerminator()->getOperandTypes()));
}

// Contains cached information for decomposed callee functions for (stateful)
// partitioned call ops.
struct PartitionedCallStackOpsInfo {
  bool signature_change;
  func::FuncOp decomposed_callee;
  llvm::SmallDenseMap<int64_t, int64_t> stack_var_arg_to_size_arg;
};

std::unique_ptr<OperationPass<ModuleOp>> CreateTPUResourceReadForWritePass();

// Creates a pass that reorders partitiioned resource reads and replicated
// inputs.
std::unique_ptr<OperationPass<func::FuncOp>>
CreateTPUReorderReplicateAndPartitionedInputsPass();

// Creates a pass that partitions unpartitioned resource read/write to
// partitioned resource variables.
std::unique_ptr<OperationPass<func::FuncOp>>

// in this function should be on the same host.
struct FunctionMetadata {
  // The original function name before partition.
  llvm::StringRef original_name;
  // The insertion point of partition functions.
  Block::iterator insertion_point;
  // The partitioned function name.
  llvm::StringRef partition_name;
  // The input values of the function.
  llvm::SmallVector<Value, 4> inputs;

func.func @cluster_func(%arg0: tensor<*xi32>) -> tensor<*xi32> {
  func.return %arg0 : tensor<*xi32>
}

// -----

// Tests output sharding of unpartitioned resource write takes on same sharding
// as unpartitioned resource.

// CHECK-LABEL: func @partitioned_input_output
func.func @partitioned_input_output(%arg0: tensor<!tf_type.resource<tensor<f32>>>) {

// which returns a dict of string key -> EITHER:
// 1. python callable (taking a checkpoint key) returning SaveableObject OR
// 2. variable (partitioned/resource/reference or otherwise)
// https://github.com/tensorflow/tensorflow/blob/ddc1bbad3dfd4a089eb96014f26cc16664b1b2f8/tensorflow/python/training/saving/saveable_object.py#L58.