Yufeng,
> I've recently brought the "prestruct-c" compiler back to "life"
Great archeology! It seems you've unearthed a snapshot from the brief
period when Dennis was struggling to reconcile byte addressing with BCPL
pointers--the seminal innovation of C. In characteristic Unix fashion, he
was trying out his ideas as they developed.
I had forgotten that product types were under con-struct-ion at the same
time. That really was a big bang.
Doug
Hi again,
I've recently brought the "prestruct-c" compiler back to "life" (https://github.com/TheBrokenPipe/C-Compiler-Dec72) and thought it might be worth documenting here. One thing I have to say first - it's barely working and probably never worked to begin with.
There were some efforts in the distant past to revive this compiler; however, the compiled compiler never worked. The reasons are as follows:
- The compiled executable is too big (exceeds 32K, making pointers effectively negative). This triggers a bug in the liba I/O routines.
- The compiler assumes an origin of 0 and writes temp data at the NULL pointer. Without an MMU, this kills the interrupt vectors and possibly the kernel on the 11/20.
- The compiler is missing ALL code/tables written in assembly language. This is pretty fatal, and internal changes rendered files from the last1120c compiler incompatible.
- Calling convention changes make the s2/last1120c libc library incompatible.
I'm a big fan of the C programming language, and the reason I was so insistent on getting this compiler to work is that it has a funny struct syntax not seen in any other C compiler. Structs are defined like:
struct name (
type field;
...
);
... with round brackets (parentheses) instead of curly braces.
Another notable thing introduced in this compiler is that certain things are no longer lvalues. In the past (B and last1120c), functions, labels, and arrays were lvalues, meaning they could be assigned. For instance, this code:
func1() { return (1); }
func2() { return (2); }
main() {
printf("func1() = %d\n", func1());
printf("func2() = %d\n", func2());
printf("func1 = func2\n");
func1 = func2;
printf("func1() = %d\n", func1());
}
produces the output:
func1() = 1
func2() = 2
func1 = func2
func1() = 2
This code:
main() {
second = first;
goto second;
first:
printf("first\n");
second:
printf("second\n");
}
produces the output:
first
second
And this code:
main(argc, argv) {
int arr1[10];
int arr2[10];
arr1[0] = 5;
arr2[0] = 8;
printf("arr1[0] = %d\n", arr1[0]);
printf("arr2[0] = %d\n", arr2[0]);
printf("arr1 = arr2\n");
arr1 = arr2;
printf("arr1[0] = %d\n", arr1[0]);
}
produces the output:
arr1[0] = 5
arr2[0] = 8
arr1 = arr2
arr1[0] = 8
Now, the rules of the game have changed with the prestruct-c compiler, and these are no longer lvalues. I don't know why this change was made, but if I had to guess, speed was probably the biggest driving factor, with security also playing a role. Anyhow, they're no longer lvalues, so there's now one less level of indirection involving functions and labels. This means the codegen tables from last1120c have to be modified to suit this compiler change.
However, even with it generating the correct code, there is still one fatal problem - the libc. The libc from s2/last1120c was designed for the older compiler and therefore has one extra layer of indirection for functions. Luckily, the source code of the libc is available on the last1120c tape, and it wasn't too much work to remove the indirection manually.
Okay, what else? Well, this compiler also seems to be the first to introduce the "modern" pointer syntax. Before this compiler, pointers were declared using the same syntax as arrays/vectors, like "char name[];". This compiler introduced the "modern" syntax of "char *name;". No big deal, right? Well, the compiler itself was written using the old syntax, meaning it cannot compile itself. I think this indicates that this compiler (or the new syntax) was so unstable that the production compiler still used the old syntax.
With everything carefully put back into place, I managed to get this to work:
struct foo (
char x;
int y;
char *z;
);
main(argc, argv)
char **argv;
{
struct foo bruh;
bruh.x = 'C';
bruh.y = 123;
bruh.z = "test";
printf("x = '%c', y = %d, z = \"%s\"\n", bruh.x, bruh.y, bruh.z);
}
However, if I rename the variable "bruh" to something like "bar", it throws the error "Unimplemented pointer conversion". I have no clue why. I've also never gotten struct pointers to work - it always complains about "Illegal structure ref" when I try to use "->". It also seems to accept "." on structure pointers (and does not actually dereference the pointer when accessing the members), so something is probably very wrong with referencing and dereferencing.
Anyway, there are plenty of other issues with the compiler, like the code may not compile correctly with pointers and switch statements. I'm not sure if the issues are caused by my poor reconstruction of the assembly tables, or if the compiler itself never worked properly in the first place (or both). Either way, I've managed to get it to spit out a correct hello world, as well as the struct test above, so I think I've fulfilled my goal of seeing this compiler "work".
The code, build instructions and pre-built binaries are here:
https://github.com/TheBrokenPipe/C-Compiler-Dec72
Ideally, it runs under a PDP-11/45 environment with 0 as the origin and generates code for the PDP-11/45. However, I made it target the 11/20 since I couldn't get the 11/45 toolchain to work, and I haven't implemented 11/45 instructions in my simulator yet. If anyone wants to pick up the baton and get it working for the 11/45 or fix my bugs, be my guest!
Sincerely,
Yufeng
> From: Yufeng Gao
> The s1 kernel is, to date, the earliest machine-readable UNIX kernel,
> sitting between V1 and V2.
It will be interesting to see what it reveals, as it's in the UNIX 'dark age'
between V1 and V4. Working from hints and clues in the extant 'UNIX
Programmer's Manual: Second Edition', I had tried to figure out how V2
differed from V1:
https://gunkies.org/wiki/UNIX_Second_Edition
but I was mostly interested in 'big picture' issues (like how a process'es
address space was laid out), not details like 'the foo() call was added', or
'how exec() differs'. (If someone _does_ create lists of the calls in V1 and
V2, and their details, and compares them, that _will_ be of value, don't get
me wrong; I was mostly just trying to work out how the mysterious KS11
worked.)
> It's somewhat picky about the environment. So far, aap's PDP-11/20
> emulator .. is the only one capable of booting the kernel. SIMH and
> Ersatz-11 both hang before reaching the login prompt.
It would be very interesting to know what fails. By 'hang', do you mean
'ceases making progress', or 'halts'?
If the former, since I've almost always had good experiences with Ersatz-11,
my _guess_ would be a problem with the RF11 emulation. (The RF11 was a very
early, and smalll, disk, so I wouldn't be surprised if there hasn't been a
lot of software run on those emulators that uses it, to flush out bugs. It's
also kind of an odd duck; it's word-oriented, not block-orientd.) So, for
instance, a 'lost' disk interrupt would produce this symptom. Are there any
RF11 diagnostics online? That would be the thing I would start with.
And I guess this system doesn't include the KS11; a pity, code that uses it
would allow re-creation of the programming manual (the way the:
https://gunkies.org/wiki/ANTS/ISI_IMP_Interface
programming instructions were re-created).
> From: Angelo Papenhoff
> So the next step would be to restore the assembly source? :)
Having only the binary to work from (to start with) is not optimal; those
early versions of UNIX ran on a number of very different hardware
configurations (e.g. with or without the KS11), with conditional assembly to
handle different configurations. Having only the dis-assembled code for _this_
configuration would obviously leave the code for the others missing.
Still, having _this_ source _would_ be useful; e.g. the 'hang-up' problem
above; the easiest way to debug that would to put 'print' statements in the
code, where a disk operation was started, and completes. If it's 'losing' a
disk interrupt completion, that will show right up. (Been there, done that, on
the RK11 hardware emulator Bridgham and I built, when UNIX wouldn't boot, just
hung.) Although I suppose one could put break-points there. Trying to debug it
any other way would be painfu beyond belief.
Noel
Hi everyone,
I've cobbed together a crude Teletype Model 37 emulator that generates PDF files (https://github.com/TheBrokenPipe/Teletype-37-PDF) It produces sane-looking PDFs for most (all?) of the early UNIX ROFF/NROFF documents.
The biggest advantage of this over something like "roff $1 | enscript -c -f Courier12 -l -M Letter --margins=67:-9:0:-9 -p $1.ps -s -0.05" is it supports half (forward/reverse) line feeds which enscript does not. Early ROFF stuff like the UNIX manuals and memos made extensive use of subscripts (and superscripts), making them rather painful to typeset.
As an experiment, I re-set Ken Thompson's "Users' Reference to B" memo from early 1972 (https://github.com/TheBrokenPipe/kbman-reset) I picked this one because it contains a BNF-alike description of the grammar as well as fractions in code comments, both of which make extensive use of sub/superscripts. I went to the extent of overlaying the re-set pages on top of the originals to make sure everything lined up.
I'd really appreciate it if someone could review my work on the B manual. If everything looks good, I may tackle other documents, starting with low hanging fruits like the "V0" manual and potentially moving on to re-setting the V1 and V2 manuals in the future (building on aap's work).
Sincerely,
Yufeng
Hi everyone,
First-time poster here. Near the end of last year, I did some forensic analysis on the DMR tapes (https://www.tuhs.org/Archive/Applications/Dennis_Tapes) and had some fun playing around with them. Warren forwarded a few of my emails to this list at the end of last year and the beginning of this year, but it was never my intention for him to be my messenger, so I'm posting here myself now.
Here's an update on my work with the s1/s2 tapes - I've managed to get a working system out of them. The s1 tape is a UNIX INIT DECtape containing the kernel, while s2 includes most of the distribution files.
The s1 kernel is, to date, the earliest machine-readable UNIX kernel, sitting between V1 and V2. It differs from the unix-jun72 kernel in the following ways:
- It supports both V1 and V2 a.outs out of the box, whereas the unmodified unix-jun72 kernel supports only V1.
- The core size has been increased to 16 KiB (8K words), while the unmodified unix-jun72 kernel has an 8 KiB (4K word) user core.
On the other hand, its syscall table matches that of V1 and the unix-jun72 kernel, lacking all V2 syscalls. Since it aligns with V1 in terms of syscalls, has the V2 core size and can run V2 binaries, I consider it a "V2 beta".
login: root
root
# ls -la
total 42
41 sdrwrw 7 root 80 Jan 1 00:02:02 .
41 sdrwrw 7 root 80 Jan 1 00:02:02 ..
43 sdrwrw 2 root 620 Jan 1 00:01:30 bin
147 l-rwrw 1 root 16448 Jan 1 00:33:51 core
42 sdrwrw 2 root 250 Jan 1 00:01:51 dev
49 sdrwrw 2 root 110 Jan 1 00:01:55 etc
54 sdrwrw 2 root 50 Jan 1 00:00:52 tmp
55 sdrwrw 7 root 80 Jan 1 00:00:52 usr
# ls -la usr
total 8
55 sdrwrw 7 root 80 Jan 1 00:00:52 .
41 sdrwrw 7 root 80 Jan 1 00:02:02 ..
56 sdrwrw 2 28 60 Jan 1 00:02:22 fort
57 sdrwrw 2 jack 50 Jan 1 00:02:39 jack
58 sdrwrw 2 6 30 Jan 1 00:02:36 ken
59 sdrwrw 2 root 120 Jan 1 00:00:52 lib
60 sdrwrw 2 sys 50 Jan 1 00:02:45 sys
142 s-rwrw 1 jack 54 Jan 1 00:52:29 x
# ed
a
main() printf("hello world!\n");
.
w hello.c
33
q
# cc hello.c
I
II
# ls -l a.out
total 3
153 sxrwrw 1 root 1328 Jan 1 00:02:12 a.out
# a.out
hello world!
#
It's somewhat picky about the environment. So far, aap's PDP-11/20 emulator (https://github.com/aap/pdp11) is the only one capable of booting the kernel. SIMH and Ersatz-11 both hang before reaching the login prompt. This makes installation from the s1/s2 tapes difficult, as aap's emulator does not support the TC11. The intended installation process involves booting from s1 and restoring files from s2.
What I did was I extracted the files from the s1 tape and placed them on an empty RF disk, then installed the unix-jun72 kernel. After booting from the RF under SIMH, I extracted the remaining files from s2. Finally, I replaced the unix-jun72 kernel with the s1 kernel using a hex editor, resulting in an RF disk image containing only files from s1/s2. This RF image is bootable under aap's emulator but not SIMH.
The RF disk image can be downloaded from here (https://github.com/TheBrokenPipe/Research-UNIX-V2-Beta)
Direct link - https://github.com/TheBrokenPipe/Research-UNIX-V2-Beta/raw/refs/heads/main/…
Interestingly, its init(7) program does not mount the RK to /usr, suggesting that /usr was stored on the RF.
Sincerely,
Yufeng
Tom Van Vleck just posted to the multicians mailing list that he is
doing an update to the Unix page at multicians.org and is soliciting
feedback. I figure some folks here may have useful suggestions.
His draft is here: https://multicians.org/unix2.html
Comments directly to Tom, I suppose, but if interested parties would
rather discuss here I'd be happy to summarize and send to him as well.
- Dan C.
For the non-TUHS folks who don't know me, I worked in
Center 1127 (the Bell Labs Computing Science Research
Center) 1984-1990, and had some hand in 9th and 10th
Edition Manuals and what passed for the V8-V10
`distributions.'
To answer Branden's points:
A. I do know what version of troff was used to typeset
the 8th through 10th Edition manuals. It was the version
we were using in 1127 at the time, which was indeed
Kernighan's. The macro packages probably matter more
than the particular troff edition.
For the 10th Edition (which files I have at hand), there
was an individual mkfile (mk(1)) for each paper, so
in principle there was no fixed formatting package,
but in practice everything appears to have used troff -mpm,
with various preprocessors according the paper: prefer,
tbl, pic, ideal, and in some cases additional macros and even
odds and ends of sed and awk.
If you wanted to re-render things from scratch you'd
want all the tools. But if you have the real troff
sources you'll have all the mkfiles--things were stored
one paper per directory.
-mpm (mpm(6) in 10/e vol 1) was a largely ms-compatible
package with special expertise in page layout.
B. There was no such thing as a `release' after V7.
In fall 1984 we made a single V8 snapshot. Making
that involved a lot of fiddly work, because we didn't
normally try to build systems from scratch; when we
brought in a new computer we cloned it from an existing
one. So there was lots of fiddly work to make sure
every program in /bin and /usr/bin on the tape compiled
correctly from the source code that would be on the tape
when the cc and as and ld and libraries on the tape were
used.
We sent V8 tapes to about a dozen external places, few
of which did anything with it (many probably never even
installed it). Which makes sense, by then we really
weren't a central source for Unix even within AT&T, let
alone to the world. Neither did we want the support
burden that would have carried--the group's charter was
research, after all, not software support. So the 9th
and 10th editions existed as manuals, but not as releases.
We did occasionally make one-off snapshots for other parts
of AT&T, and maybe for a university or two. (I definitely
remember taking a snapshot to help the official AT&T System N
Unix people set up a Research system at one point, and have
a vague memory that I may have carried a tape to a university
under a special one-off license letter.)
On the other hand, troff wasn't a rapid moving target, and
unlike the stars of the modern software world, we tried not
to break things unless there was a real reason to do so.
So I suspect the troff from any system of that era would
render the Volume 2 papers properly, and am all but certain
the 10th-edition-era troff would do so even for older manuals.
C. Just to be clear, the official 10th Edition manuals
published by Saunders College Publishing were made from
camera-ready copy prepared by us in 1127 (Doug McIlroy
did all the final work, I think) and printed on our
phototypesetter. We didn't ship them troff source, nor
even Postscript. We did everything including the tables
of contents and indexes and page numbering.
D. troff is indeed not TeX, and some of us think of that
as a feature, not a bug.
I think the odds are fairly good (but not 100%) that
groff would do a reasonable job of rendering the papers;
as I said, the hard part is the macro packages. I'm
not sure -mpm ever made it out of Research.
And there are probably copyright issues not just with
the software but with the papers themselves. The published
manuals bear a copyright notice, after all.
Norman Wilson
Toronto ON
(A much nicer place than suburban NJ, which is why
I left the Labs when I did)
Although I edited the v7 through v10 manuals, I have no recollection of
why "system" crept into the title between v7 and v8. Resistance to
trademark edicts did grow. In v10, the cover and the man pages proclaimed
"Unix". However, the fossilized spelling, "UNIX", still appeared in the
introduction to Volume 1 and scattered throughout Volume 2.
Doug
So in most technical circles and indeed in the research communities surrounding
UNIX, the name of the system was just that, UNIX, prefixed often with some
descriptor of which stream, be it Research, USG, BSD/Berkeley, but in any case
the name UNIX itself was descriptive of the operating system for many of its
acolytes and disciples.
However, in AT&T literature and media, addition of "System" to the end of the
formal name seemed to become de facto if not de jure. This can be seen for
instance in manual edits in the early 80s with references to just "UNIX" being
replaced with variations on "The UNIX System", sometimes haphazardly as if done
via a search and replace with little review. This too is evident in some
informative films published by AT&T, available on YouTube today as
"The UNIX Operating System" and "UNIX: Making Computers Easier to Use"[1][2].
Discrepancies in the titles of the videos notwithstanding, throughout it seems
there are several instances where audio of an interviewee saying
"The UNIX System" were edited over what I presume were instances of them simply
saying UNIX.
I'm curious if anyone has the scoop on whether this was an attempt to echo the
"One Bell System" and related terminology, marketing tag lines like
"The System is the Solution", and/or the naming of the revisions themselves as
"System <xyz>". On the other hand, could it have simply been for clarity, with
the uninitiated not being able to glean from the product name anything about it,
making the case for adding "System" in formal descriptions to give them a little
bit of a hint.
Bell Labs folks especially, was there ever some grand thou shalt call it
"The UNIX System" in all PR directive or was it just something that organically
happened over time as bureaucratic powers at be got their hands on a part of the
steering wheel?
- Matt G.
[1] - https://www.youtube.com/watch?v=tc4ROCJYbm0
[2] - https://www.youtube.com/watch?v=XvDZLjaCJuw
> My understanding is that Unix V8-V10 were not full distributions but
patches.
"Patch" connotes individually distributed small fixes, not complete
working systems. I don't believe Brendan meant that v8 was only a patch on
v7, but that's the natural interpretation of the statement.
V8-v10 were snapshots, yes, possibly not perfectly in sync with the
printed editions. But this was typical of Research editions, and especially
of Volujme 2,
which was originally called something like "Documents for Use with Unix".
Doug
