dynamic-extent propagation failure make-array/make-struct/list

[James Y Knight]

For some reason, dynamic extent fails to be propgated through the make-array/make-struct to the list call in the following example program. I don't know why, though. Rather irritating, because it doesn't even print a couldn't d-x allocate compiler note when this happens...

compile-file/load/running testit on the following code prints something like:

array: 2AAAAD8C17FF struct1: 2AAAAD8C1831 struct2: 2AAAAD8C1811
struct1-a: 2AAAAD8C1857 struct2-a: 100334B657

That is: everything got DX-allocated except "(list 6)".
The make-array was dx-allocated, both make-foo were dx-allocated.
The (list 5)" allocated as part of "s" is dx-allocated.
But *not* the (list 6) through make-foo through make-array.

All the cruft in that example isn't necessary, btw, the failure happens just the same way with a "trivial-example", as below. I just added the cruft to attempt to figure out where the problem was occurring.

(declaim (optimize (safety 0) (speed 3) (debug 0)))

(declaim (inline make-foo))
(defstruct foo
  a)

(defun testit ()
  (let* ((s (make-foo :a (list 5)))
         (result (make-array 2 :initial-contents (list s (make-foo :a (list 6))))))
    (declare (dynamic-extent s result))
    (locally (declare (optimize (speed 1)))
      (format t "array: ~X struct1: ~X struct2: ~X struct1-a: ~X struct2-a: ~X"
              (sb-kernel:get-lisp-obj-address result)
              (sb-kernel:get-lisp-obj-address (elt result 0))
              (sb-kernel:get-lisp-obj-address (elt result 1))
              (sb-kernel:get-lisp-obj-address (foo-a (elt result 0)))
              (sb-kernel:get-lisp-obj-address (foo-a (elt result 1)))))))

(defun trivial-example ()
  (let ((result (make-array 1 :initial-contents (list (make-foo :a (list 6))))))
    (declare (dynamic-extent result))
    (car (foo-a (elt result 0)))))

Tested with SBCL 1.0.38.12 Linux x86-64.

[Nikodemus Siivola (nikodemus)]

The problem lies in the array optimizations making LIST NOTINLINE in order to pick it apart -- which foils DX for it later on.

[Nikodemus Siivola (nikodemus)]

Apropos: when trying to figure out how something is decided to be DX or not, a good strategy is to put a break after HANDLE-NESTED-DYNAMIC-EXTENT-LVARS in RECHECK-DYNAMIC-EXTENT-LVARS, and trace H-N-D-E-L and its callees.

[Nikodemus Siivola (nikodemus)]

In SBCL 1.0.42.12.