Re: 16 bit integers in streams

[stt@houdini.camb.inmet.com (Tucker Taft)]

Stephen Leake (Stephen.Leake@gsfc.nasa.gov) wrote:

: I assume size clauses control both the external and internal
: representation; see below.

You would be wrong if you made that assumption.

: > I suspect the "right" solution is to consider the base range to be
: > a 16 bit range for types like this, even though 32-bit arithmetic is
: > used for them. We have already considered doing this, but haven't
: > done it yet.  Of course, the longer we wait, the more people will 
: > begin to rely on the current behavior...

: Or complain about it :). Surely the intent of 'Read is to put the
: _memory_ image of a type into the stream, not the _register_ image?

Because 'Write/'Read are designed to handle the *base* range of the 
type, rather than the range of the first subtype, the stream 
representation is in some sense more related to the "register" image 
than the memory image.

: > Perhaps we could use an explicit size clause as a way to control
: > the base range.  We have frequently debated the meaning of
: > a "confirming" size clause.  Perhaps this could be one "real" effect
: > of such a clause.  It's an intriguing thought...

: This is definitely what I expect a size clause to mean; I'm telling the
: compiler I want exactly 16 bits, anyplace it matters (including
: streams). I assume that's what GNAT does now, although maybe GNAT is
: also doing 16 bit arithmetic.

A size clause currently has *no* effect on the base range of the type.
I am thinking that perhaps it ought to have such an effect, though
that "thought" is pretty radical.  And there are some problems with
the idea.  For example, given:

   type I16 is range 0..2**16-1;
   for I16'size use 16;

Clearly the base range of this type is required by the RM to be
at least -2**16+1 .. 2**16-1, since it is a signed integer type 
(even though the first subtype's range includes no negative values).
So what should be the stream representation for this type?  
Presuming each stream element is 8 bits, probably at least
3 stream elements, and more likely 4 stream elements, would be the
expected stream representation.

In general, the existence of the base subtype is more of an interesting
frill than a core feature, but with streams, the characteristics of
the base subtype suddenly become more important.  It's not clear
how to resolve this dilemma...

: -- 
: - Stephe

--
-Tucker Taft   stt@inmet.com   http://www.inmet.com/~stt/
Intermetrics, Inc.  Burlington, MA  USA