The Big Switcheroo

:: Racket, Emacs

In my previous post I talked about how Racket Mode now will often want the back end command server, before actually needing a live REPL — but it has to start a REPL anyway. This was bothering me. So I went on to address that. Doing so entailed reversing the I/O model for the back end. As a bonus, that set up fairly short strokes to supporting multiple REPLs.

In the beginning, Racket Mode was a thin wrapper around xrepl. As a result, it used a comint-mode buffer to start Racket, and the REPL stdin/stdout was connected to the comint-mode buffer. The back end also started a little TCP server; connecting to that established the I/O for commands.

This was mostly fine, especially when racket-run tended to be the first command you would use.

However it might take a second or two for the TCP server to be ready to accept connections. As a result, the front end either needed to block (boo!) or use an Emacs timer to retry at, say, one second intervals until success. That meant some extra complexity. More basically, it meant more delay until a command request and response could occur.

Flip

So I decided to flip that around. Now, the stdin/stdout of the back end process is the command request/response channel. The same format of sexpr command requests and responses are sent — they just happen to go over ports that are stdin/stdout instead of TCP ports. The TCP server is for creating a REPL session, if/when that is desired. Make a TCP connection, read a session ID, et voila you have a new REPL session. The back end parameterizes the REPL’s current-{input output error}-mode to the TCP ports. And fortunately, comint-mode makes it as easy to connect the buffer to I/O from a TCP connection, as to stdin/stdout of a process.

The other thing is, a process can have only one stdin/stdout. But of course it can accept multiple TCP connections. Now that we use the latter for REPL I/O, this removes a basic barrier to an old feature request: Supporting multiple REPLs would now be possible.

Of course “possible” is different than “implemented and working correctly”. So at first I focused mainly on the back end: Make sure that it wasn’t using global variables when it could use per-thread parameters, for things related to the REPL. Add a lookup table from unique REPL session IDs to information about each REPL session. Have the front end obtain the session ID and supply it later with commands — but otherwise, for now, the front end still just created one REPL session at a time.

Multiple REPLs

After some time to get that working, I wanted to tackle supporting multiple REPLs in the front end. For a day or two, I procrastinated. I really did not want to write a bunch of “ceremonial” code related to REPL sessions. I did not want some whole UI for that, complete with more code and bugs.

I’m glad I hesitated, because I realized something simple. Emacs already has the concept that variable values can be global to all buffers — but also can be set so that a buffer has its own, local value.

So, assume a variable racket-repl-buffer-name. It only has meaning for a racket-mode buffer. It means, “what is the name of the racket-repl-mode buffer that commands like racket-run should use”. Initially such a variable could have a global value, *Racket REPL*. That default is the status quo behavior: All edit buffers take turns sharing the same, one and only REPL buffer.

Now, what if a racket-mode buffer for foo.rkt did a (setq-local racket-repl-buffer-name "*Racket REPL foo.rkt*")? Now commands issued with that buffer selected would use (creating if necessary) a REPL buffer of that name. If every edit buffer did this, you’d have one REPL per edit buffer (much like DrRacket).

You could also imagine wanting edit buffers for files that are in the same projectile project, to share a REPL buffer: “One REPL per project”. OK: setq-local their racket-repl-buffer-name to some common name that includes the project name or root directory.

And so that is the mechanism I went with. A customization variable racket-repl-buffer-name-function points to a function that will set the racket-repl-buffer-name variable value for a racket-mode edit buffer when it is created. I defined three such functions, for the three strategies described above. And best of all, as a user you can supply your own such function.

I also added a kill-buffer-hook: When you kill a racket-mode edit buffer, and it is the last one using a racket-repl-mode buffer, we y-or-n-prompt offering to kill the REPL buffer, too. I figure that will help keep things cleaner, especially for the case where people want 1:1 edit:REPL buffers.

As a result, there is no Big UI about “sessions”. There are just edit buffers, and names of the REPL buffer(s) each should use. Because it rides directly on Emacs semantics for variables and buffers, there is non-trivial code that I did not need to write — and “out of sync” bugs that I could not create.

That cleaving feeling

Of course there are some bugs lurking, about which I don’t yet know. Someday I’ll probably feel discouraged about this code, for reasons I don’t yet know. A lot of programming “innovation” amounts to trading an old set of problems for a new set. Sometimes the new problems aren’t better, they just feel better, at first, because novelty.

All stipulated and admitted. Even so, I also feel this was one of those situations where, finally, I saw a way to define a problem so that it aligns with the material I had to work with: Just a few gentle taps, and the material cleaves along the lines I need and want. That is a pretty awesome feeling. While it lasts.

Availability

The stuff I’m describing in this blog post initially lived on a new-comm branch. I did just merge it back to the check-syntax branch. It’s not yet merged to master. I’m still reviewing documentation, dogfooding, and such. I’m coming up on two full months working on this, so, I’d like to merge to master in the near future.