Ramblings on languages and architectures (was Re: any benefit to throwing off lua51 constraints?)

Szabó Antal wrote:
> This topic interest me too, and I was wondering what a completely
> newly designed language + implementation (so it doesn't have any
> limitations or restrictions of existing languages) created by Mike
> would be like.

Well, I've been thinking about this for quite some time. But I'm
approaching it from the other side: what's the most efficient CPU
architecture and what would a language have to look like that
makes the best use of it?

I'll start with a simple observation from the perspective of a
compiler writer:

All modern and advanced compilers convert source code through
various stages and representation into an internal data-flow
representation, usually a variant of SSA. The compiler backend
converts that back to an imperative representation, i.e. machine
code. That entails many complicated transforms e.g. register
allocation, instruction selection, instruction scheduling and so
on. Lots of heuristics are used to tame their NP-complete nature.
That implies missing some optimization opportunities, of course.

OTOH a modern CPU uses super-scalar and out-of-order execution. So
the first thing it has to do, is to perform data-flow analysis on
the machine code to turn that back into an (implicit) data-flow
representation! Otherwise the CPU cannot analyze the dependencies
between instructions.

Sounds wasteful? Oh, yes, it is. Mainly due to the impedance loss
between the various stages, representations and abstractions.

That's just one example of a general problem: it's hard to
preserve programmer intent across all stages down to the silicon.
So when designing a language, one should strive to offer more
intent-based programming concepts.

I strongly believe in the Sapir-Whorf hypothesis when it comes to
computer language design. A likely conjecture is that if you offer
the right abstractions, people will use them. So better design
abstractions that map cleanly onto the target architecture(s).
Alas, if you're only looking at current CPU architectures, you'd
primarily offer plain imperative concepts. Thankfully, the arrival
of GPGPUs in the mainstream has opened up the perspective somewhat.

But we're still very much entrenched in the thinking that jointly
created the CPU architectures plus the matching languages in the
last decades. This is neither surprising nor a bad thing per se --
we've witnessed enormous progress in scalar execution performance
and (much less spectacular) progress in programming language
implementation efficiency. But the real advances in e.g. parallel
execution will only materialize by out-of-the-box thinking.

So, dreaming a little bit, a CPU based on a data-flow architecture
is probably the way to go. Obviously, data-flow programming
languages are a natural match. But most of the existing ones are
visual languages. That's nice for some tasks, but IMHO these are
not suitable for general programming. And before the functional
programming crowd cheers: nope, sorry, the usual blend of concepts
offered by FP languages is decidedly not what I'd personally
consider worthwhile (some concepts in isolation certainly are).

Ok, so I have plenty of ideas for all levels of the stack, but not
enough time to pursue them with sufficient rigor. And I can offer
little tangible results at this time. Duh. But then, you've seen
the subject of this posting, so that's what you got. :-) It's
tough to find the right mix of concepts for such a language and
yet make it attractive for programmers.

Anyway, please consider this my personal ramblings on what needs
to be done to advance the state of the art. Your mileage may vary.
Have fun thinking about these ideas and discussing them (but
probably not here on this list). And please use them freely. Maybe
start some research and revolutionize the world? Who knows ...

--Mike

Other related posts: