The More We Share, The More We Have: Improving GPU performance through Register Sharing