circularity detection of FORMAT NIL behaves oddly in PRINT-OBJECT methods

[3b]

(defclass foo () ())
(defmethod print-object ((o foo) s)
  (let ((x "foo"))
    (flet ((twice (a s y)
             (format s "~a ~a ~a" a y y)))
      (format s "s ~a ~a~%" x x)
      (twice "twice " s x)
      (format s "~&nil ~s~%" (format nil "~a ~a" x x))
      (twice "twice " s x)
      (format s "~&nil ~s~%" (format nil "~a ~a" x x))
      (format s "nil/twice ~s~%" (twice "" nil x))
      (format s "nil ~s~%" (format nil "~a ~a ~a" "" x x))
      (format s "~&wots ~s~%"
              (with-output-to-string (ss)
                (format ss "~a ~a ~a" "" x x)))
      (format s "nil ~s~%" (format nil "~a ~a ~a" "" x x))
      (format s "wots/twice ~s~%"
              (with-output-to-string (ss)
                (twice " " ss x)))
      (format s "nil ~s~%" (format nil "~a ~a" x x))
      (format s "nil/twice ~s~%" (twice "" nil x))
      (format s "nil ~s~%" (format nil "~a ~a" x x)))))

(let ((*print-circle* t))
  (princ (make-instance 'foo))
  (values))

s #1=foo #1#
#2=twice #1# #1#
nil "foo foo"
#2# #1# #1#
nil "foo foo"
nil/twice "#3= #1# #1#"
nil " foo foo"
wots "#3# #1# #1#"
nil "#3# #1# #1#"
wots/twice " #1# #1#"
nil "#1# #1#"
nil/twice "#3# #1# #1#"
nil "#1# #1#"

not completely sure what the correct output is, but (format nil "~a ~a" x x) should probably return either "#1=foo #1#" or "#1# #1#" (or "#1=foo #1#" once then "#1# #1#") depending on whether you consider "creates and returns a string containing the output from control-string" from http://www.lispworks.com/documentation/HyperSpec/Body/22_c.htm to mean it is printing to a new stream or not, and if so if it is a new one per call. http://www.lispworks.com/documentation/HyperSpec/Body/v_pr_cir.htm specifies that circularity detection starts over when printing to a stream other than the one passed to the PRINT-OBJECT method.

Either way, returning "#1# #1#" when the NIL isn't a constant, then returning "foo foo" a few times when NIL is passed directly, then "#1# #1#" when NIL is passed directly after calling with-output-to-string seems wrong.

"#1# #1#" shoring up in with-output-to-string also looks wrong, probably https://bugs.launchpad.net/sbcl/+bug/309090

see also https://bugs.launchpad.net/sbcl/+bug/308948

[3b (00003b)]

reported by _death on #lisp in context of condition reports (which use print-object) in http://paste.lisp.org/display/330728