swap(x,y)

[Blair P. Houghton]

In article <714 at philmtl.philips.ca> ray at philmtl.philips.ca (Raymond Dunn) writes:
>In article <4151 at buengc.BU.EDU> bph at buengc.bu.edu (Blair P. Houghton) writes:
>>Because if a machine can do it with a coupla gates and half a cycle,
>>I'd like to do it with an operator.
>
>This sounds dangerously like the arguments made by Herman Rubin that 'C'
>should provide facilities to access all the functionality of the machine
>architecture in some direct way.

What is this, the Nuremberg-defense rebuttal?  ( :-) )

The issue I'd like to address is that there are some constructs
that seem entirely obvious, but that are not a part of my favorite
language.  Some of them aren't there because there's no rational
way to implement them in a compiler.  A swap-operator, however,
is not at all that way.

Swapping a couple of variables is the utter basis of (almost) all sorting
algorithms.  Every one of them (except hashing) depends on the machine's
ability to put things A and B into places B and A, respectively, and
at least meta-simultaneously.  (I can see Gwyn salivating in the
background...maybe someone should throw him a pointer-sum...)  

I don't see a need for explicit register loading, PC munging, or
even specific selection of the method of addressing objects.  I think
the concepts of 'near' and 'far' are too chummy with the hardware, to
paraphrase some bigwig in the biz.

>Since when was that the goal of *any* language other than assemblers?

Swapping is a data operation, nor specifically a CPU-defined operation.
I bet it doesn't even require the CPU, just a very smart MMU.  There
are (as we've seen) plenty of ways, whether straightforward or
obfuscated-but-extra-efficient, to accomplish it.  Selection of the
optimal method should be up to the compiler, which shouldn't have to be
in -O mode to get it right.  C is almost silly to have omitted it.

>Which particular architecture is the source of these language features?

In this case, it's (allegedly) a Data General; a DG-6000, I think.
Someone used it in an early example in this thread of discussion.
He showed some C that the compiler had turned into a single-instruction
using the swap-opcode.

I thought, as soon as I saw that, "why does it take a special and
expensive pile of optimization algorithms to recognize a swap,
then replace it with this one instruction?"

It shouldn't.  You should just say ( a swap_op b ) and let fly.

We could even use the @-character.  It looks a little like it's twirling
around, anyway.

				--Blair
				  " 'a @ b' or fight!"