Some of the posts in this thread confuse pipes in UNIX shells (the | symbol) with the pipe system call.

In the shell, how two processes can be connected with a pipe is very constrained (only one unidirectional pipe). But the pipe system call can be used to build much more elaborate connections. Back in 1980, when I was still at Bell Labs, I wrote a shell called 2dsh ("two dimensional shell) that had a more complex syntax. The memo I wrote, "2DSH—An experimental shell for connecting processes with multiple data streams", wasn't published externally, but exists as a Bell Labs memo. I found a reference here: https://scholar.google.com/scholar_lookup?title=2DSH%E2%80%94An+experimental+shell+for+connecting+processes+with+multiple+data+streams&author=M.+J.+Rochkind&publication_year=1980.

Here are two examples from that memo:

I stumbled across another paper from 2017 titled "Extending Unix Pipelines to DAGs," which references my un-published Bell Labs memo. I haven't read it since I don't subscribe to IEEE Transactions on Computers.

A while ago Doug McIlroy was kind enough to send me a scan of my memo, but I don't think I'm allowed to publish it here. In that memo, I credit Doug for coming with a very similar idea around the same time ("A Notation for Arboreal Plumbing").

Marc Rochkind

On Thu, Dec 5, 2024 at 2:13 PM Dan Cross <crossd@gmail.com> wrote:

On Thu, Dec 5, 2024 at 3:56 PM <arnold@skeeve.com> wrote:
> Chet Ramey via TUHS <tuhs@tuhs.org> wrote:
> > On 12/5/24 10:19 AM, Dan Cross wrote:
> >
> > > Unix pipelines, on the other hand, tend to be used in a manner that is
> > > strictly linear, without the fan-out and fan-in capabilities described
> > > by Morrison. Of course, nothing prevents one from building a
> > > Morrison-style "network" from Unix processes and pipes, though it's
> > > hard to see how that would work without something like `select`, which
> > > didn't yet exist in 1978. Regardless, Unix still doesn't expose a
> > > particularly convenient syntax for expressing these sorts of
> > > constructions at the shell.
> >
> > Process substitution is about as close as we can get, but most programs
> > still process their filename arguments one at a time, beginning to end.
> >
> > The canonical process substitution example is
> >
> > diff <(old-program-version) <(new-program-version)
> >
> > to do simple regression testing.
>
> And fanout is simply
>
> ... | tee >(pipeline1) >(pipeline2)

And indeed these things are pretty nifty, but don't they generate
trees, and not arbitrary dags? They don't quite capture the full
generality of Morrison-style networks since it doesn't seem like
there's a way to connect process substitution fan-out with fan-in; at
least, not conveniently.

It is, perhaps, notable that Go allows me to do this sort of thing
with channels and goroutines, but it has `select` built into the
language.

Funny, despite using Unix almost daily for over 30 years now, I don't
think I've ever felt limited by the power of pipelines. On the
contrary, I've lost count of the times I've felt limited on systems
that do Not support pipes.

- Dan C.