Extremely CISC instructions

Paul Koning paulkoning at comcast.net
Tue Aug 24 10:38:42 CDT 2021



> On Aug 24, 2021, at 6:34 AM, Diane Bruce <db at db.net> wrote:
> 
> On Mon, Aug 23, 2021 at 09:09:55PM -0400, Paul Koning via cctalk wrote:
>> 
>> 
>>> On Aug 23, 2021, at 8:38 PM, Tom Stepleton via cctalk <cctalk at classiccmp.org> wrote:
>>> 
>>> Hello,
>>> 
>>> For the sake of illustration to folks who are not necessarily used to
>>> thinking about what computers do at the machine code level, I'm interested
>>> in collecting examples of single instructions for any CPU architecture that
>>> are unusually prolific in one way or another.
>>> ...
>>> Although I don't know it well, I suspect VAX will place well in one way or
>>> another.
>> 
>> Indeed. Just its addressing modes, with indirection and two separate register indexing operations, are hairy enough.  Then consider the decimal arithmetic instructions that might have up to 6 operands.
> 
> And who can ever forget the built in CRC instruction?
> 
> Amusingly it was demonstrated that many of the CISC instructions were
> faster if one just did them with the 'normal' instruction set.

Not too surprising given that the instruction worked 4 bits at a time while a reasonable software implementation goes 8 bits at a time.

The more general point is one I learned on the Moto 68040.  The task was to write a fast packet forwarding module (for an FDDI -> Ethernet switch, wihch ended up being the DECswitch 900).  When I studied the instruction timings I realized the "RISC subset" of the 68040 instruction set and addressing modes runs in one or two cycles per instruction, while all the other cases take a great deal more.  So I wrote it in the RISC subset, and ended up processing 60k packets per second, in a 25 MHz processor.

	paul




More information about the cctalk mailing list