Re: RISC

[Tucker Taft <stt@averstar.com>]

Marin David Condic wrote:
> 
> Oh sure. Most garden variety representation clauses get listened to and
> obeyed. At least in my experience. I just like to complain about the
> inability to coerce behavior out of compilers sometimes. I know I've come up
> with perfectly reasonable and legal representation specifications from time
> to time and had them rejected by compilers because a) they didn't understand
> what I wanted or b) they *did* understand and refused to do it because
> obviously they knew better than I did what it is I *really* wanted, if only
> I was as smart and omnipotent as the compiler.

This is my cue to pipe up with the Compiler Writer's Lament.
Rep clauses are difficult to support because every special case
makes a compiler more complicated and more buggy, and makes
optimization harder.  So compilers really "want" to do everything
one way if possible, and then optimize the heck out of that.

With rep-clauses, the programmer is sometimes asking the compiler
to violate some of its basic assumptions (e.g., the "tag" is always
the first word of the record, the non-variant part precedes the
entire variant part, every dynamic-sized component is dumped at the
end of the record, with an offset word identifying where it starts,
every object is aligned on its "natural" alignment boundary, bit fields
never cross word boundaries, packed array components never cross
byte boundaries, ...).  

Of course any one of these assumptions can be examined 
and declared stupid, and the compiler writer can
be scolded for making any such assumptions.  On the other hand,
the programmer expects that if by chance there is an efficient
way of implementing the specified representation, the compiler jolly-well
better recognize that fact and do the right thing.  E.g., allow me
to specify fields that cross word boundaries, but if they don't, then
do the efficient in-line field extract.

I can assure you that we devote huge quantities of code in our compilers
to dealing with programmer-specified representations, and still
we find that there are some representations we can't handle.
Sometimes we know it, so we reject the rep-clause.  Sometimes we
don't know it, and generate goofy or incorrect code.

So feel a little sympathy for us, and/or get out your check books
and pony up some big bucks for the ultimate in rep-clause support ;-).

> That's when you have to get creative in the stories you tell to the
> compiler! :-)
> 
> MDC
> --
> Marin David Condic
> Senior Software Engineer
> Pace Micro Technology Americas    www.pacemicro.com
> Enabling the digital revolution
> e-Mail:    marin.condic@pacemicro.com
> Web:      http://www.mcondic.com/

-- 
-Tucker Taft   stt@avercom.net   http://www.averstar.com/~stt/
Chief Technology Officer, AverCom Corporation (A Titan Company) 
Burlington, MA  USA (AverCom was formerly the Commercial Division of AverStar:
http://www.averstar.com/services/ebusiness_applications.html)