Issue: ARRAY-TYPE-ELEMENT-TYPE-SEMANTICS (Version 8)

[Jim McDonald <jlm@lucid.com>]

     ... It would be valid for an implementation to ... remember the
     original array element-type or its canonical or expanded form ...
     and satisfy all of the constraints of this proposal.

This observation makes the proposal much more palatable, so I won't
press the issue beyond the following two observations:

(1)  The extra storage per array for remembering the original type can
     be quite minimal, depending on the implementation.  If arrays
     already allocate a full pointer for the type, that can just as
     well point at the pair (<upgraded-type> . <given-type>) shared
     among all such arrays.  If arrays use a special encoding of a few
     bits to identify the updgraded type, then a small number of extra
     bits suffices to index into a table of given types, overflowing
     to the use of a full pointer.  This does not necessarily
     complicate the runtime array access code, but could require enough
     trickiness in laying out arrays to be considered non-trivial. 

(2)  I find it esthetically disturbing that the existing proposal
     seems to force the following to be legal in many implementations:

       (setf (aref (the (array (unsigned-byte 5)) foo) 0)
	     127)

     Any proposal to make such code illegal would seem to be
     incompatible with ARRAY-TYPE-ELEMENT-TYPE-SEMANTICS:UNIFY-UPGRADING 
     since then, e.g., for any implementation for which 
     (array (unsigned-byte 5)) upgrades to (array (unsigned-byte 8))
     the following code would be illlegal or legal depending whether 
     baz is declared inline or not:

       (defun baz (x) (setf (aref x 0) 127)) 

       (defun peculiar ()
         (let ((x (make-array 0 :element-type '(unsigned-byte 4))))
	   (declare (type (array (unsigned-byte 4)) x))
	   (baz x)
           (aref x 0)))

     In fact, I can't even portably assume that PECULIAR returns a
     number, since in implementions which upgrade to (array t),
     BAZ could be consistently redefined to install a string.

     I keep wondering what the potential convert from FORTRAN or
     PASCAL would think of all this. 

Thanks for bearing with me.

 jlm