Where should the type information be: in tags and descriptors 443

Edward A. Feustel

memory

be

Windows, recoded, more

Porting Linux to a new architecture is something that has occurred many times.

Of course, in one of the threads on this topic, it had been noted that the C language allows people to do just the 'bad things' that a tagged architecture exists to prevent. Thus a major conversion effort might well be required for Linux, one unlikely to happen.

In the case of Windows, of course, it would not happen, or it would be irrelevant if it did; the advantage of using Windows is being able to run the many programs written for it, and distributed in binary form, without buying new copies when you upgrade.

But does all this mean tagged architecture is bad, or at least irrelevant? Not really. The forthcoming Windows "Longhorn" is supposed to include support for 'managed code', using a P-machine to nicely sandbox the ActiveX controls on the pages you visit. Thus, adding a feature to the x86 architecture to run programs in, say, 'guest mode', might not be a bad idea at all.

But will tagged architecture ever fly as the primary architecture of computers, if we could leave aside considerations of x86 compatibility?

As also noted in previous threads, languages can be type-safe even on conventional architectures. Given that programming in buttembler is uncommon these days, it would seem that a Pascal compiler, rather than a tagged architecture, suffices to save the programmer from himself.

And, as long as "anyone" is allowed to write a *compiler*, writing data to disk is sufficient of an end-run around type safety (as writing data to tape apparently was not on the Burroughs systems, hence, I suppose, nonprivileged users could not be allowed to bring in tapes written upon by IBM systems - or perhaps they were just labelled in some way, with a switch thrown on the tape drive, to prevent the computer from trusting them to contain programs) that a tagged architecture would not provide one of the most desirable potential features of such a design - making writing viruses impossible (or at least a lot more difficult).

The problem is that, today, despite hardware being so cheap compared to its cost in the past - today's desktop computer has surpbutted the 360-195, and is approaching the power of a Cray 2 - the appebreaste for more and better is insatiable. IEEE 754 for floating-point only got adopted because Intel revealed some of the good ideas used in the 8087 chip that allowed its features to be implemented at very limited cost.

Having cast aside the tyranny of x86 compatibility, we thus come face-to-face with the tyrrany of the benchmark.

However, as also noted in one of the threads, the Burroughs machines performed their type-checking in parallel with everything else. The cost, therefore, was extra transistors, not an irreducible cost in time.

At the present time, new chips contain ever more transistors in order to increase their speed. However, there is an upper limit to this, when one is considering problems that are not fully parallelizable. (When something is parallelizable, of course, one can just throw any number Mersenne Primes Search.)

Hence, it may well happen that once we start making 64-core chips and the like, it will become possible to give the idea of a tagged architecture a serious look once again. Again, as noted, even today the idea of allowing a chip to provide a tagged 'guest mode' for some programs has a clear value that can be appreciated, but that is very different from having an architecture fundamentally oriented towards tagged data.

John Savard