Bit Addressable Architectures

[Henry Spencer]

> >I once had the opportunity to ask Bill Wulf what he thought of bit-oriented
> >machines; his answer was "I wish they weren't so damned slow".  I'm afraid
> >I haven't seen anything since that invalidates that assessment.  There is
> >something to be said for providing bit addressability, but one must realize
> >that actually exploiting it will be slow and that there will still be a
> >large payoff for trying to work on byte or word boundaries whenever possible.
> 
> It seems to me that aligned access to all items larger than a bit would allow
> a bit addressable machine to be every bit as fast as a byte or word addressable
> machine.  Am I missing something?

No and yes.

No, in that this is exactly what I said in the last sentence of my comments,
although somewhat obscurely.  (Note that "bit-oriented" and "bit-addressable"
aren't the same thing in the terminology I was using.)  As an extreme case,
one can envision a bit-addressable machine -- that is, one whose pointers
use the low-order three bits to indicate a bit within a byte -- that traps
whenever those bits aren't zero, leaving the actual use of bit pointers
entirely up to the software.  When all accesses were in fact aligned, this
would incur essentially no overhead except the reduction in address space.

Yes, in that almost any attempt to make bit-aligned objects easier to handle
is going to mean extra hardware, quite possibly in a critical path where
every added gate slows the whole machine down.  Even if it's not in a
critical path, it will steal chip area from other things that could boost
performance.  The tradeoffs depend on the design details.
-- 
Those who do not understand Unix are |  Henry Spencer @ U of Toronto Zoology
condemned to reinvent it, poorly.    | {allegra,ihnp4,decvax,utai}!utzoo!henry