Re: architecture, The Design of Design

Tim Rentsch <tr.17687@z991.linuxsc.com> writes:

John Levine <johnl@taugh.com> writes:

Well, yes, but another 360 innovation was the whole idea of computer
architecture, as well as the term.  It was the first time that the
programmer's view of the computer was described independently of any
implementation.

>
I don't buy it.  An architecture is just a description of system
behavior, and surely there were descriptions of system behavior
before System/360.  Even in the 1950s companies must have changed
implementations of a given model while still conforming to its
earlier description.

Sure, the 7094 was a compatible successor of the 704, but the idea of
implementation independence turns out to be much more profound than
most people (probably including its inventors at the time) realized.
The hardware behind the z16 is vastly different from that of any of
the initial members of the 360 family, its microarchitecture is also
vastly different (caches, virtual memory, out-of-order execution,
speculative execution, superscalar execution), and yet it can run
software written for the S/360 60 years ago.

And the important part is not just what to put into the architecture,
but also what to leave out.  There is no end to clever ideas that
would improve performance of a particular implementation, that are bad
ideas when it comes to architecture.  A widely accepted example is
branch delay slots.

An interesting example is IA-64: It was designed as a long-lived
architecture (while the VLIW machines that provided some of the ideas
for IA-64 seem to be mainly designed as implementations), but it
turned out that the special architectural features it had could be
provided through microarchitecture (in particular, OoO execution) to
earlier architectures, and that these features were pointless for OoO
microarchitectures.

Another interesting example is Alpha.  It was claimed to be designed
for a 25-year lifetime (actually there were 9 years between the
introduction of the 21064 in 1992 and the cancellation of the 21464 in
2001).  They left out all features that (they claimed) hindered
performance, such as a flags register and byte/word-access (BWX)
instructions; the BWX case was supposedly because they required ECC
for write-back caches.  But the first implementations (EV4, EV45, EV5,
EV56) all have write-through L1 caches, so BWX instructions would have
been no problem (and they were added in EV56 in 1996, i.e. after 4
years).  EV6, which has a write-back L1 cache, has a write buffer, so
generating ECC for the BWX instructions was not particularly
expensive.  The downside of leaving an architectural feature like BWX
out in the first implementation is that much software would forego
using BWX for a very long time (if Alpha had lived that long).

As for the word architecture, it seems like
an obvious and natural word choice, given the hundreds (or more)
of years of experience with blueprints and buildings.

I don't think that their achievement is in choosing the word
"architecture".  It reflects on the division of labor between the
planner of a building and the people who implement the plan, but it
does not transport the fact that computer ISA "architects" design the
interface between software and hardware for a vast amount of software
and a vast difference in potential hardware across the decades, much
of it unforeseeable for the computer architect.  By contrast, building
architects are more like computer microarchitects, designing for
building materials of the day, and the uses of the buildings tend to
be less varied than the software that runs on a general-purpose ISA;
ok, my office is in a building that was built as a residence building
in 1913, but that's not because the architect designed it for
general-purpose usage.

- anton
--
'Anyone trying for "industrial quality" ISA should avoid undefined behavior.'
  Mitch Alsup, <c17fcd89-f024-40e7-a594-88a85ac10d20o@googlegroups.com>