GNU Emacs 19.28 and 21.2 for VAX OpenVMS 7.3

GNU Emacs 19.28 and 21.2 for VAX OpenVMS 7.3

I have been running GNU Emacs on my VAXstations under OpenVMS VAX 7.3, and today there is a kit other people can install: Emacs 19.28 and Emacs 21.2, terminal-only, side by side, with an installer and a regression suite. Both were built on a VAXstation 4000/96, installed from the kit on a VAXstation 4000/60 and on a simulated MicroVAX 3900, and tested on all three.

Claude Code assisted, which is to say, did much of the hard work.

Downloads, with SHA-256 sums in SHA256SUMS.TXT:

  • Binary kit (32 MB): two BACKUP savesets, the installer, the test suite, and a README.TXT that says everything below in more detail.
  • Source kit (67 MB): the complete source of both versions as built, plus the two unmodified upstream archives they came from.
  • README.TXT on its own, and SHA256SUMS.TXT.

Emacs is GPL; the source kit is the corresponding source.

Where the two versions come from

  • 19.28 is Richard Levitte’s 1996 VMS kit. It is the version that has been on VMS the longest and it still runs well on a small VAX.
  • 21.2 is Thien-Thi Nguyen’s VMS port of Emacs 21, in the form Eric Brown brought up on OpenVMS x86-64 in 2025-2026. Eric’s tree is the base; the VAX work below is on top of it. 21.2 has font-lock, a text-mode menu bar, and twenty-five years of Lisp the older version lacks.

Nothing in either tree was rewritten. The changes are patches, each one a separate commit with the symptom, the cause and the fix in its message.

Installing

You need OpenVMS VAX 7.3 with TCP/IP Services (Emacs links against it), about 200,000 free blocks on the system disk, and a VT-series terminal or an emulator of one. No compiler is needed for the binary kit.

  1. Get the kit’s files into one directory on the VAX. Transfer modes matter: the two .BCK savesets in binary, every .COM, .EL and .TXT in ASCII. (UnZip for VMS gets both right by itself if you unzip on the VAX.) A procedure sent in binary arrives as 512-byte fixed records and DCL refuses it with %RMS-W-RTB; the kit is not damaged, send the text files again in ASCII.

  2. As SYSTEM:

     $ SET DEFAULT dev:[the.kit.directory]
 $ @EMACS_KIT_INSTALL CHECK        ! says what it would do
 $ @EMACS_KIT_INSTALL

    The installer repairs the savesets’ record attributes (they always arrive wrong), restores both versions under SYS$SYSDEVICE:[GNU...], copies the shared procedures, offers to add the two lines your SYLOGIN.COM needs (it asks first; it never edits SYSTARTUP_VMS.COM, it prints the one line to add there yourself), and then runs the regression suite. Its exit status is the suite’s.

  3. Log in again. EMACS19 and EMACS21 start a kept editor: a subprocess that survives C-z, so EMACS21 a second time re-attaches with your buffers intact. RUNEMACS19 and RUNEMACS21 run Emacs directly. Plain EMACS is 19.28 unless the logical name GNU_EMACS_DEFAULT is 21.

Installing the kit twice, or over an older copy of it, is safe: every file ends up as exactly one version. (The first version of the installer got that wrong; see below.)

Testing it

@SYS$SYSDEVICE:[GNU]EMACSTEST ALL from your own login runs the batch suite against both versions: 254 checks each, in twenty groups – VMS file versions and names, logical names, record formats, completion, dired, access checks that depend on your privileges, synchronous and asynchronous subprocesses, timers, GC, byte compilation, the editing primitives – and reports PASS/FAIL per check and a summary line. Every check was shown to fail at least once before it was trusted: the suite has a switch that negates any named check, and the checks that guard a specific fix were run against the build that lacked it.

There is also a second tier of scripted full-screen tests (a kept editor, M-!, M-x shell, completion, function keys, two kept editors at once) that drive a terminal from a host over telnet and assert on the rendered screen. Those are not in the kit; they are in the project repository.

What was changed, and why

Emacs 21.2 – ten changes. Eight are VAX-only and leave Alpha and x86-64 alone:

  1. The compile switch /OPTIMIZE=(LEVEL=5,TUNE=HOST) is Alpha’s; the VAX compiler rejects it.
  2. <libicb.h> does not exist on VAX and nothing used it.
  3. The VAX compiler has no 64-bit integer type. The port used unsigned __int64 for a timer request id and unsigned long long for the privilege mask; both became longwords (the mask as two of them, as 19.28 did it).
  4. The same in vmsproc.c (long long delta time, done with LIB$EMUL), plus XPROCESS() applied to something that was already a struct pointer – a type error every compiler should have caught.
  5. The dump step aborted with “unhandled overlap”: the VAX linker puts the BSS markers beside DATA, so the BSS range duplicated DATA.
  6. A function name from VSI C on x86-64 had been typed into the source; on VAX it is plain sys_access.
  7. make-docfile commands were too long for DCL when the build directory was deep, and one batch of docstrings was silently left out of DOC.
  8. All platforms: the last record a subprocess wrote before exiting was dropped. M-! SHOW TIME printed nothing.
  9. All platforms: under -batch, file access was checked as the account that built Emacs, because the UIC is cached in a static that the dump file preserves.
  10. The one that took a day: Compaq C for VAX inlines a function that calls the alloca builtin, then keeps using SP-relative slots after SP has moved. Every file-name completion – C-x C-f TAB, RET in dired – jumped into data and killed Emacs. That buffer now comes from the heap. 19.28 is immune because it uses Emacs’s own alloca.c.

Plus VMS terminal names in termcap, and a vt500.el so that on an LK-style keyboard without an Escape key, F11 acts as ESC.

A traditional C preprocessor (tradcpp, shipped in the 21.2 kit) is used instead of Compaq C’s, which puts a space between every token and wrecks the generated MMS files.

Emacs 19.28 needed build-system repairs, the same dropped-record fix as number 8, and one more: close() was being called on an index into Emacs’s own mailbox table, so emacs -batch -l file.el closed the file it was loading and died at the next read.

Since the first kit (which I never published), the subprocess path got a proper going-over, with a trace compiled into a test build to see what it was doing rather than guessing:

  • M-x shell in 21.2 was slow – a 600-file DIRECTORY took eleven seconds where M-! of the same command took one. The trace showed no waiting at all: each read returned one DCL record and Emacs spent about fifty milliseconds of its own CPU per record on filters and redisplay. A read now drains the records that follow. 1.8 seconds, faster than 19.28.
  • A one-shot asynchronous subprocess never exited: DCL, given both a command and an input device, kept reading for another command, so no sentinel ever ran and the process quota filled up. Both versions.
  • Behind that, three more in 21.2: output lost at exit, -batch pty records truncated to ten characters (the batch “frame” was ten columns wide), and delete-process hanging because the 21.2 port had regressed SIGKILL from $DELPRC to $FORCEX, which cannot stop a DCL sitting at its prompt.
  • In 19.28, a shell’s exit sentinel ran only at the next keystroke: after the exit AST, a drain read waited on a dead pty until any other event.
  • 21.2 reported a failing command’s VMS completion code as a Unix signal (“user defined signal 1”); it now shows the code and its message text. In 19.28, RET in dired said “Buffer is read-only” (the keymap never bound it). start-process with no argument crashed the image.

Rebuilding it

The source kit’s emacs-21.2-vax-source-*.zip is the whole tree as built. Unzip it under your own account on a VAX with Compaq C and MMS, then

$ SUBMIT/NOPRINT/LOG_FILE=SYS$LOGIN:E21.LOG -
      /PARAMETERS=(CONFIGURE) dev:[dir]BUILD_EMACS21.COM

configures and builds it (about 18 minutes on a 4000/96, 5 on a simulator), and /PARAMETERS=(INSTALL) as SYSTEM installs it. The install prefix is compiled in, so a kit built with one prefix runs only from that prefix. Building 19.28 is documented in the repository; its generated makefiles need a repair script before MMS will accept them.

The builds are reproducible in the sense that matters here: the same commit built on the simulator and on real hardware gives the same warnings in the same numbers and an image and dump of the same size, and this was checked every time anything changed.

Known limitations

  • No X or DECwindows: terminal only, by design.
  • Neither version notices a terminal window resize; M-x set-screen-width or C-l after resizing.
  • 21.2, unprivileged accounts: file-writable-p on your own home directory says nil although writing there works.
  • 19.28: run-at-time never fires in batch mode – it needs a helper program the VMS build never included.
  • Output captured from a subprocess ends with one extra blank line.
  • The link reports %LINK-W-MULDEF for the allocator and directory routines: Emacs’s own versions replacing the C library’s, by design.

How this was done

Two VAXstations, one simulator, and a small team of Claude sessions: one that plans and reviews, and one fresh session per run with a written brief. The rule that made it work was that nothing counts until a script fails without the fix and passes with it, on the simulator and on real hardware, with the transcripts kept. Several of the fixes above were found by tests that were written to fail first; the last one before release – the kit’s own self-test failing for every account except the one that installed it – was caught only by logging in as an ordinary user and running exactly what the README says, with nothing else staged. The repository with the tooling, the test suites and every review will be public in due course.

Credits: Richard Levitte for the 19.28 VMS port, Thien-Thi Nguyen for the 21.2 VMS port, and Eric Brown for the OpenVMS x86-64 work the 21.2 base comes from.