Follow-up: C and tagged pointers

[jdb at s1-c]

A while back I posted a message to this group concerning the use
of C on a machine with tagged pointers.  (In my particular case,
the machine is the S-1 Mark IIa.)  I received a number of thoughtful
replies, but for various reasons only a few of my direct-mail responses
made it to their destinations.

Several respondents pointed out that C guarantees that no object
resides at location zero ("...it is guaranteed that the assignment of
the constant 0 to a pointer will produce a null pointer distinguishable
from a pointer to any object" -- C Programming Language, p. 192).  I
was aware of this; it motivated my suggestion that (char *)0 produce a
pointer with a "nil" tag.  Unfortunately, there are situations where
the use of a "true" pointer to zero is desirable; e.g.

	&((struct xyzzy *)0)->abcde

to obtain the offset of "abcde" in a structure of type "xyzzy".  It
is aesthetically displeasing to treat 0 specially so that

	((int *)XYZZY)[i]

is an invalid reference through a null pointer if VWXYZ is zero, but
is OK if VWXYZ is anything else.  (Consider code written in C which
patches interrupt vectors in low core.)

Other problems which show up are the initialization of the BSS to zero
(which MANY programs rely upon to initialize BSS pointers to NULL)
and routines such as "execl", for which the last argument in the
argument list is defined to be the integer 0.

I can't simply #define NULL to the appropriate nil pointer unless I
want to type-cast each use or suffer through pointer type conflict
messages from "(p)cc" and "lint".

We have tenatively decided to change the architecture of the S-1 so
that a tag of 0 (formerly an invalid pointer tag) will be treated as
a "user-defined" tag.  This will allow the use of pointer manipulation
instructions (move pointer, validate pointer, etc.) without producing
an invalid pointer fault.

	John Bruner
	S-1 Project/LLL