Bad asymptotics on gencgc for pinned objects

[Paul Khuong]

Pinned large objects objects are considered completely dirty and remain so until they're not pinned anymore. For instance, the following shows a pretty clear linear growth in the size of the array, even though there is only a single write to the vector (linux/amd64/1.0.49.x).

The remembered set should be more precise, at least for very large vectors, which I expect are the only cases that could happen in the wild.

(defun test (length &optional (repeat 10))
  (let ((x (make-array length)))
    (setf (aref x 0) 0)
    (time (dotimes (i repeat)
            (gc)))))

* (gc :full t) (test (* 64 4096 4096))

NIL
*
Evaluation took:
  18.129 seconds of real time
  17.877118 seconds of total run time (17.313082 user, 0.564036 system)
  [ Run times consist of 17.877 seconds GC time, and 0.001 seconds non-GC time. ]
  98.61% CPU
  50,634,721,036 processor cycles
  4,512 bytes consed

NIL
* (gc :full t) (test (* 8 4096 4096))

NIL
*
Evaluation took:
  2.284 seconds of real time
  2.244140 seconds of total run time (2.172135 user, 0.072005 system)
  [ Run times consist of 2.245 seconds GC time, and 0.000 seconds non-GC time. ]
  98.25% CPU
  6,377,437,976 processor cycles
  4,512 bytes consed

NIL
* (gc :full t) (test (* 4096 4096))

NIL
*
Evaluation took:
  0.289 seconds of real time
  0.288018 seconds of total run time (0.276017 user, 0.012001 system)
  [ Run times consist of 0.284 seconds GC time, and 0.005 seconds non-GC time. ]
  99.65% CPU
  805,982,844 processor cycles
  3,920 bytes consed

NIL